Paper deep dive
Training-inference input alignment outweighs framework choice in longitudinal retinal image prediction
Liyin Chen, Nazlee Zebardast, Mengyu Wang, Tobias Elze, Jason I. Comander
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 100%
Last extracted: 4/27/2026, 3:08:11 AM
Summary
The paper investigates the impact of training-inference distributional alignment on longitudinal retinal image prediction. Through a controlled comparison of five conditioning configurations, the authors demonstrate that aligning the training and inference input distributions is more critical for performance than the choice of generative framework (e.g., diffusion vs. deterministic regression). This is attributed to the low task entropy in slowly progressing retinal diseases, where inter-visit change is dominated by acquisition variability. Based on these findings, the authors introduce TRU (Temporal Retinal U-Net), a deterministic direct-regression model using continuous time-delta conditioning and multi-scale history aggregation. TRU was validated across 28,902 eyes using various imaging platforms (Optos FAF, Heidelberg Spectralis, and Cirrus en-face fundus) and demonstrated superior performance compared to state-of-the-art benchmarks.
Entities (13)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Quantitative prediction of future retinal appearance from longitudinal imaging would support clinical decisions in progressive macular disease that currently rely on qualitative comparison or scalar progression scores. Recent methods have moved toward increasing generative complexity, but whether this complexity is necessary for slowly progressing retinal disease is unclear. We tested this through a controlled comparison of five conditioning configurations sharing one architecture and training dataset, spanning standard conditional diffusion, inference-aligned stochastic training, and deterministic regression. In our evaluation, aligning the training and inference input distributions produced large gains (delta-SSIM +0.082, SSIM +0.086, both p < 0.001), while the choice among aligned frameworks did not significantly affect any primary metric. Task-entropy and posterior-concentration analyses, replicated on two fundus autofluorescence (FAF) platforms, provided a mechanistic account: the predictable component of inter-visit change is small relative to time-invariant acquisition variability, leaving stochastic sampling with little width to exploit. Guided by these findings, we developed TRU (Temporal Retinal U-Net), a deterministic direct-regression model with continuous time-delta conditioning and multi-scale history aggregation. We evaluated TRU on 28,902 eyes across three imaging platforms: a mixed-disease Optos FAF cohort (9,942 eyes), zero-shot transfer to Stargardt macular dystrophy on Optos (288 eyes) and Heidelberg Spectralis (125 eyes), and a boundary evaluation on Cirrus en-face fundus images from a glaucoma cohort (18,547 eyes). TRU matched or exceeded delta-SSIM, SSIM, and PSNR in every FAF cohort against three state-of-the-art benchmarks, and its advantage grew monotonically with available history length.
Tags
Links
- Source: https://arxiv.org/abs/2604.16955v1
- Canonical: https://arxiv.org/abs/2604.16955v1
Trouble viewing inline? Open PDF directly →
Full Text
122,661 characters extracted from source content.
Expand or collapse full text
Distributional Alignment Supersedes Generative Complexity in Longitudinal Retinal Image Prediction Liyin Chen 1,4 , Nazlee Zebardast 1,2 , Mengyu Wang 2 , Tobias Elze 2 , Jason I. Comander* 1,3 1 Department of Ophthalmology, Massachusetts Eye and Ear, Harvard Medical School, Boston, MA, USA 2 Harvard Ophthalmology AI Lab, Schepens Eye Research Institute, Harvard Medical School, Boston, MA, USA 3 Ocular Genomics Institute, Massachusetts Eye and Ear, Harvard Medical School, Boston, MA, USA 4 Broad Institute of MIT and Harvard, Cambridge, MA, USA *Corresponding author Abstract Quantitative prediction of future retinal appearance from longitudinal imaging would support clinical decisions in progressive macular disease that currently rely on qualitative comparison or scalar progression scores. Recent methods have moved toward increasing generative complexity, but whether this complexity is necessary for slowly progressing retinal disease is unclear. We tested this through a controlled comparison of five conditioning configurations sharing one architecture and training dataset, spanning standard conditional diffusion, inference-aligned stochastic training, and deterministic regression. At the single-institution training scale evaluated, aligning the training and inference input distributions produced large gains (ΔSSIM +0.082, SSIM +0.086, both p < 0.001), while the choice among aligned frameworks did not significantly affect any primary metric. Task-entropy and posterior-concentration analyses, replicated on two fundus autofluorescence (FAF) platforms, provided a mechanistic account: the predictable component of inter-visit change is small relative to time-invariant acquisition variability, leaving stochastic sampling with little width to exploit. Guided by these findings, we developed TRU (Temporal Retinal U-Net), a deterministic direct-regression model with continuous time-delta conditioning and multi-scale history aggregation. We evaluated TRU on 28,902 eyes across four cohorts: a mixed-disease Optos FAF cohort (9,942 eyes), zero-shot transfer to Stargardt macular dystrophy on Optos (288 eyes) and Heidelberg Spectralis (125 eyes), and a boundary evaluation on Cirrus en-face fundus images from a glaucoma cohort (18,547 eyes). TRU matched or exceeded ΔSSIM, SSIM, and PSNR in every FAF cohort against three state-of-the-art benchmarks, and its advantage grew monotonically with available history length. Keywords: fundus autofluorescence, spatial-temporal model, disease progression prediction, exposure bias, retinal imaging, deep learning, retinal diseases 1. Introduction Progressive retinal diseases such as geographic atrophy (GA) and Stargardt macular dystrophy cause irreversible vision loss through expansion of hypoautofluorescent lesions in the retinal pigment epithelium (RPE), visible on fundus autofluorescence (FAF) imaging [1,2]. Quantitative prediction of future retinal appearance from longitudinal imaging history would support clinical decisions that currently rely on qualitative serial comparison, including treatment timing, monitoring interval selection, and clinical trial stratification by expected progression rate. Most existing computational tools address this need through binary atrophy segmentation or scalar progression scores [3–8], both of which discard clinically relevant information. Specifically, binary lesion segmentation methods concentrate the continuous-intensity image into a binary mask, discarding continuous intensity patterns that carry independent prognostic value. Scalar progression scores provide a single number per patient that lacks clinically important spatial specificity. Dense image prediction, by contrast, preserves the full spatial and intensity information. Predicted images enable post-hoc extraction of any scalar or binary metric, and support direct visual comparison with the patient's actual follow-up image. Recent efforts to produce dense longitudinal predictions in medical imaging have converged on a methodological trajectory of increasing generative complexity, from neural ODEs [9,10] through temporal conditional diffusion [11] and multi-stage latent diffusion pipelines [12] to flow matching formulations [13,14]. Each step along this trajectory adds capacity, parameters, and inference complexity in pursuit of prediction quality. However, to our knowledge, none have examined whether this complexity is necessary for slowly progressing retinal diseases. In this work, we constructed a controlled comparison of five conditioning configurations that share one architecture, one training dataset, and one set of optimizer hyperparameters, differing only in how the training input is constructed and how inference is initialized. The configurations span the methodological space from standard DDPM-style conditional diffusion with iterative sampling [15], through inference-aligned stochastic, to fully deterministic regression. This design isolates two questions. First, how much performance improvement can we achieved by replacing the training input of standard conditional diffusion (i.e noised versions of the the targaet image) with the same intput as in inference (i.e the patient's most recent observation). Second, once this alignment is in place, whether the remaining choice among training frameworks affects prediction quality. Our first finding was that, at our training scale, distributional alignment between training and inference inputs dominated the choice among aligned generative frameworks by more than an order of magnitude. Correcting the training–inference mismatch yielded ΔSSIM +0.082 and SSIM +0.086 (both p < 0.001), while switching among aligned frameworks produced no statistically significant difference on any primary clinical outcome. The alignment principle itself has precedent in the image-to-image translation literature [16–19]. What is new here is its controlled isolation against current longitudinal baselines in a clinical imaging task, and the demonstration that the mismatch penalty is destructive rather than merely suboptimal. Our second finding provided a mechanistic explanation of the observations above. Two complementary analyses showed that the conditional posterior over future FAF images in slowly progressing retinal disease had collapsed to an effective point mass. A task entropy analysis across 587.9 million pixels revealed that inter-visit change was dominated by time-invariant acquisition variability rather than disease progression, with changed-pixel fraction barely correlated with elapsed time (r = 0.116). A direct posterior concentration measurement across ten independent noise realizations showed that 99.99% of remaining prediction error was systematic bias, with stochastic variance contributing less than 0.02%. The disease-relevant signal occupied a low-dimensional subspace of total inter-visit variability, and within that subspace the posterior was sharply peaked, leaving stochastic sampling with no meaningful width to exploit. These results were consistently observed between datasets from two different types of imaging devices. These two findings together motivate a task-adaptive design principle that we propose as a hypothesis for empirical testing in other longitudinal medical imaging domains. The appropriate level of generative complexity for a longitudinal prediction task should match the entropy of the predictable component of its conditional posterior, with distributional alignment as a prerequisite in all regimes. The principle is diagnostic rather than prescriptive, specifying two model-light measurements that practitioners can perform on any candidate dataset before committing to a framework. We believe this pre-commitment diagnostic is the most transferable contribution of this work, and present its operational form, domain scope, and boundary cases in Section 7. To instantiate the principle and validate it against current longitudinal baselines, we developed TRU (Temporal Retinal U-Net), a deterministic direct-regression model with continuous time- delta conditioning and multi-scale history aggregation. The role of TRU is not to introduce architectural novelty, but to show that the simple architecture consistent with the principle can match or exceed substantially more complex published methods. We validated TRU across a four-cohort hierarchy spanning 28,902 eyes, three imaging manufacturers, and two imaging modalities. The primary FAF benchmark comprised 9,942 mixed-disease eyes with 1,001 held out for testing. Zero-shot transfer to Stargardt macular dystrophy (288 eyes) tested generalization to an unseen rare inherited disease within the same modality. Zero-shot transfer to glaucoma en- face scanning laser ophthalmoscopy (18,547 eyes) tests simultaneous cross-disease and cross- modality robustness. TRU overall outperformed ImageFlowNet [9], TADM [11], and a 2D adaptation of BrLP [12] on primary metrics across all three cohorts, and its advantage grew with the number of available history frames, a property with direct relevance to patients under active longitudinal monitoring. Contributions. The contributions of this work are as follows. 1. A controlled isolation of distributional alignment between training and inference inputs as the dominant performance driver for longitudinal retinal image prediction in a clinical setting. A five-configuration comparison sharing one architecture and training dataset shows that correcting the alignment mismatch contributes significant gains in ΔSSIM and SSIM, while switching among aligned frameworks produces no statistically significant difference on any primary clinical metric. 2. A mechanistic explanation of the observed framework equivalence, grounded in task entropy analysis and posterior concentration measurement on FAF. The predictable component of inter-visit change is small relative to time-invariant acquisition variability, producing a conditional posterior that collapses to an effective point mass and leaves stochastic sampling with no meaningful width to exploit. A design principle based on these two measurements was proposed to help practitioners assess whether stochastic generative machinery is warranted on a candidate longitudinal imaging dataset before committing to a framework. 3. TRU, a deterministic instantiation of the design principle, validated across a four- cohort hierarchy comprising 28,902 eyes. To our knowledge, this is the largest-scale evaluation of dense longitudinal image prediction reported on retinal imaging, and the first to span vendor and modality shift on this task. 2. Related Work We organize prior work into three areas that together establish the gap this paper addresses. The first is deep learning for retinal disease progression prediction, which motivates dense image prediction as the appropriate output modality. The second is longitudinal medical image prediction with deep generative models, which documents the methodological trajectory toward increasing generative complexity. The third is the training and inference distributional mismatch in conditional diffusion, which introduces the alignment problem that our controlled analysis isolates as the operative bottleneck. 2.1 Deep learning for retinal disease progression prediction Most existing deep learning approaches to progression prediction in retinal disease produce either binary segmentation outputs or scalar risk scores, and both output types discard information that the underlying image signal contains. Binary lesion segmentation methods predict whether each retinal location will become atrophic at a future timepoint. Salvi et al. [4] forecast geographic atrophy growth from baseline FAF as per-pixel atrophy probability maps, and Cluceru et al. [7] extended this approach with a topographic analysis of where progression occurs most rapidly across the macula. Liefers et al. [6] developed a multi-feature segmentation system that quantifies AMD-related lesion types from individual images, and Mishra et al. [8] proposed a recurrent architecture that jointly predicts Stargardt and geographic atrophy progression by modeling temporal dependencies between lesion masks at successive visits. Across this line of work, the output lies in a lesion-class label space and therefore discards the continuous FAF intensity signal, including the hyperautofluorescent rim that often precedes lesion expansion and the graded transition zones between intact and atrophic RPE. A second family of methods produces scalar progression scores. Yim et al. [5] trained a system to predict conversion to neovascular AMD as a binary patient-level event, and Dai et al. [3] addressed time to progression for diabetic retinopathy as a scalar regression problem. Methods of this class support population-level risk stratification but collapse the spatial heterogeneity of disease progression into a single number and therefore cannot indicate where on the retina change is expected, which is the question that drives treatment timing decisions in geographic atrophy and trial enrichment in inherited macular dystrophies. Dense image prediction preserves the information that both discards. A predicted future FAF image can be post-processed to extract any binary mask or scalar score, while retaining the spatial pattern and graded intensity transitions that those derived metrics summarize. Despite this advantage, dense image prediction has been largely absent from the retinal imaging literature, which is the gap our model addresses. 2.2 Longitudinal medical image prediction with deep generative models In medical imaging more broadly, dense longitudinal prediction has been pursued primarily in brain MRI for neurodegenerative disease progression, and the methodological trajectory in this area moves consistently toward increasing generative complexity. ImageFlowNet is the most direct prior work on longitudinal retinal image prediction and the only published method targeting dense image-level forecasting in this modality. It parameterizes temporal evolution as a neural ordinary differential equation in a learned latent space, trained with a multiscale contrastive regularizer against a population-level temporal prior rather than an individual-level supervised target. TADM applies conditional diffusion to longitudinal brain MRI by modeling the temporal residual between consecutive timepoints in an epsilon-prediction framework, conditioning on a single source frame through a dedicated encoder branch, and performing inference by iterative DDPM sampling initialized from Gaussian noise. It represents the most direct adaptation of standard conditional diffusion to the longitudinal prediction problem. BrLP, published recently in Medical Image Analysis, advances individual-level brain MRI forecasting through a three-stage latent diffusion pipeline comprising a variational autoencoder for compression, an unconditional diffusion model in the learned latent space, and a ControlNet for subject-specific conditioning. BrLP introduces a Latent Average Stabilization procedure that averages multiple latent samples at inference for temporal consistency, and it incorporates auxiliary regional volume prediction to inject prior knowledge about disease-specific anatomical change. BrLP achieves strong results on Alzheimer’s disease cohorts and represents the current state of the art for individualized 3D brain MRI prediction. Two recent works extend the landscape to flow matching. Disch et al. introduce Temporal Flow Matching for 4D longitudinal medical imaging, learning velocity fields that interpolate between consecutive image timepoints, and Chen et al. propose a latent flow matching variant that conditions trajectories on patient-specific covariates. 2.3 Training and inference distributional mismatch in conditional generation Standard conditional diffusion models exhibit a structural mismatch between training and inference inputs. During training, the model learns to denoise corrupted versions of the target image, which is the unknown future observation the model is being asked to predict. At inference, the target is unavailable and the sampling procedure must be initialized from something else, typically pure Gaussian noise in iterative sampling or a noised version of the source image in single-pass approaches. The distribution of inputs the model encounters at deployment is therefore not the distribution it was trained on, and the resulting performance penalty has been recognized in the broader generative modeling literature as a form of exposure bias [16]. Several methods address this mismatch through alternative training formulations. Image-to-Image Schrödinger Bridge [17] reformulates image translation as a learned stochastic bridge between paired source and target distributions, eliminating the noise-source asymmetry by training on paths anchored at both endpoints. Cold Diffusion [18] generalizes the diffusion framework to deterministic non-noise corruption operators, and Inversion by Direct Iteration [19] proposes a training procedure in which the model learns to invert a deterministic degradation operator applied to the source image. These methods share a common conceptual move. They align the training input distribution to the inference input distribution by training on paths that begin from a source image rather than from noise. A related class of methods reframes the problem as flow matching, which learns continuous-time velocity fields rather than discrete denoising steps. Standard flow matching trains the velocity field along a path interpolating between a Gaussian source distribution and the target data distribution, and therefore inherits the same source-distribution mismatch as DDPM-style diffusion when applied to image-to-image tasks. Bridge variants of flow matching, including the bridge formulation underlying Image-to-Image Schrödinger Bridge, train directly on source-to-target paths and fall within the same class of inference-aligned methods. The common structural property of all methods in this class is that training and inference operate on paths anchored at the source image rather than at a distribution-free noise endpoint. Despite the methodological richness of this literature, no prior work has examined the specific impact of the training and inference distributional mismatch on longitudinal retinal image prediction, nor isolated the contribution of distributional alignment from the choice of generative framework. 3. Methods 3.1 Problem Formulation Given a sequence of N registered retinal images 퐼 1 ,...,퐼 푁 from a single eye, acquired at timepoints 푡 1 ,...,푡 푁 , the goal is to predict the image 퐼 ∗ at a future timepoint 푡 ∗ . For each history frame 퐼 푘 , we define the time delta 훥푡 푘 =푡 ∗ −푡 푘 (in years) and denote the target gap as 훥푡 ∗ = 푡 ∗ −푡 푁 . The longitudinal history is ℋ= ( 퐼 푘 ,훥푡 푘 ) 푘=1 푁 . TRU learns a mapping 푓 휃 ( 퐼 푁 ,ℋ,훥푡 ∗ ) → 퐼 ̂ ∗ that predicts the future retinal image from the most recent observation and the full available history. The architecture and hyperparameters are identical across imaging modalities (FAF and SLO) and across all evaluation cohorts. 3.2 Temporal Conditioning Architecture TRU is built on a multiscale temporal U-Net backbone [20] with approximately 22.8M parameters. The architecture integrates two core design elements: continuous time-delta conditioning encoding the temporal distance from each history frame to the prediction target, and multi-scale history feature extraction with delta-weighted aggregation. The backbone is a four- level encoder–decoder with channel widths (64, 128, 256, 512) and skip connections at all scales (Figure 1). Full architectural details (layer counts, ResBlock distribution, GroupNorm configuration) are provided in Supplementary Methods. Continuous Time-Delta Conditioning The temporal distance from each history frame to the prediction target is encoded as a continuous 256-dimensional embedding [15,21]. For a time delta 훥푡 (in years), the encoding proceeds as: 푒 ( 훥푡 ) ∗2푖=sin ( log ( 1+훥푡 ) ⋅푓 푖 ) , 푒 ( 훥푡 ) ∗2푖+1=cos ( log ( 1+훥푡 ) ⋅푓 푖 ) where 푓 푖 =exp ( −푖⋅log ( 100 ) /128 ) for 푖=0,...,127, spacing frequencies log-uniformly from 1 to 1/100 . The log transform log ( 1+훥푡 ) compresses the dynamic range of inter-visit intervals and ensures that the encoding is well-conditioned for both short and long prediction horizons. The raw sinusoidal encoding is passed through a two-layer MLP to produce the final delta embedding 퐞 훥 [22]. Per-frame delta embeddings are computed for all 푁 history frames at each forward pass. The embedding of the most recent valid frame, 퐞 훥,푁 , serves as the global delta conditioning signal passed to every ResBlock. Multi-Scale History Feature Extraction and Aggregation A weight-sharing convolutional encoder extracts features from each history frame at three spatial scales (256², 128², 64²) through successive convolution, normalization, and pooling stages. All 퐵×푁 frames are processed in a single batched forward pass by reshaping, producing three feature banks 퐅∗푘 ( 푠 ) 푘=1 푁 for 푠∈1,2,3. At each spatial scale 푠, the 푁 per-frame feature maps are aggregated via a delta-weighted attention mechanism that scores frames by their temporal proximity to the prediction target: 훼 푘 =softmax∗푘!(퐰 푇 퐞 훥,푘 ), 퐆 ( 푠 ) =∑훼 푘 푁 푘=1 ,퐅 푘 ( 푠 ) where 퐰∈ℝ 256 is a learnable projection applied to the delta embedding of each frame. Three independent delta-weighted attention modules operate at scales 1, 2, and 3, each with its own projection 퐰 ( 푠 ) . History Injection and Residual Prediction The aggregated history features 퐆 ( 푠 ) are injected into the U-Net encoder at the corresponding scales via channel-wise concatenation followed by a 1×1 convolution acting as a pixel-wise temporal mixer, followed by GroupNorm with SiLU activation [23], providing normalization and non-linear refinement of the fused representation. The model predicts the residual change from the most recent history frame rather than the absolute target [24]: 퐼 ̂ ∗ =퐼 푁 +푓 휃 ( 퐼 푁 ,ℋ,훥푡 ∗ ) where 푓 휃 denotes the U-Net output (prior to the residual addition). This residual formulation concentrates capacity on the disease-relevant change signal. The output layer is initialized near zero, so the initial prediction approximates the copy-last baseline 퐼 푁 . 3.3 Training Configuration Our primary model, TRU, is trained to predict the ground-truth target 퐼 ∗ via a per-pixel masked mean squared error loss: ℒ= 1 퐵 ∑ 1 푁 푏 퐵 푏=1 ∑푀 푏 푖,푗 ( 푖,푗 ) ⋅(퐼 ̂ 푏 ∗ ( 푖,푗 ) −퐼 푏 ∗ ( 푖,푗 ) ) 2 where 푀 푏 is the validity mask (excluding non-retinal border regions introduced in registration, Supplemental Methods), 푁 푏 = ∑ 푀 푏푖,푗 ( 푖,푗 ) is the number of valid pixels in sample 푏, and the per-sample normalization by 푁 푏 ensures equal contribution from each eye regardless of fundus coverage area. Inference is a single deterministic forward pass. The model variants below share the same architecture, optimizer, and hyperparameter choices as TRU and only differ in the conditioning configurations on how the training input is constructed and how inference is performed. They are used in the controlled analysis in Section 6. 1) Std-DDIM-50step: standard v-prediction diffusion training [21,25,26] that noises the ground- truth target I*; 50-step DDIM sampling initialized from pure Gaussian noise. Both the training input distribution (q(x t | I*)) and the inference initialization (pure noise, bearing no relationship to the patient’s retinal anatomy) are maximally mismatched relative to the task requirement. 2) Std-DDIM-1step: identical training to Std-DDIM-50step (noising I*), but inference is initialized from noised I N at t = 200 with single-pass x ̂ 0 prediction. This replaces the pure-noise starting point with the patient’s last observation, but was still trained on inputs derived from I*, not I N . 3) IA-Nonlinear: diffusion training is done with the same noising schedule as the above two variants, but it noises I N rather than I* during training. Same single-pass x ̂ 0 inference at t = 200 as Std-DDIM-1step. The training input distribution now matches the inference input distribution: both are q(x t | I N ). 4) IA-Linear: Stochastic training with a linear interpolation noise schedule. The training input is constructed as (1 − t) · I N + t · ε, with t ~ U[0.05, 0.95], and inference uses single-pass x ̂ ₀ prediction at t = 0.20. This configuration tests whether the shape of the noise schedule, linear versus the nonlinear cumulative-product schedule of IA-Nonlinear, affects prediction quality under alignment. All training is done with geometric augmentation (random horizontal flipping and random affine transformation) applied jointly to all frames of each longitudinal sequence to preserve spatial correspondence. Training uses the AdamW [27] optimizer (learning rate 1 × 10⁻⁴, weight decay 1 × 10⁻⁴), linear warmup (1,000 steps), cosine decay [28], gradient clipping at 1.0, EMA (decay 0.999), batch size 12, for 150 epochs. Figure 1. Architecture of Temporal Retinal U-Net (TRU). (A) Registered longitudinal FAF images from a single eye serve as input. (B) A weight-sharing convolutional encoder extracts spatial features from each frame at three scales (256², 128², 64²). At each scale, per-frame features are aggregated via δ-weighted attention scored by the time-delta embedding, which encodes the temporal distance from each history frame to the prediction target through a log-sinusoidal encoding followed by a two-layer MLP. The aggregation produces a single spatial feature tensor per scale (G⁽¹⁾, G⁽²⁾, G⁽³⁾), weighting frames by temporal proximity. (C) A four-level U-Net backbone receives the most recent frame I N as input. Aggregated history features are injected at three encoder scales (⊕) via concatenation and 1×1 convolution. The time-delta embedding conditions every ResBlock through learned affine modulation. (D) The network output f θ is added to the most recent frame I N as a residual, producing the predicted future FAF image Î*. 4. Experimental Setup 4.1 Datasets Three datasets constitute our validation hierarchy, summarized in Table 1. Optos FAF mixed-disease dataset (primary training and evaluation). The primary dataset comprises 8,718 patients (55.6 ± 19.3 years; 52.0% female) contributing 9,942 eyes and 24,419 FAF images that were registered post-hoc within each time series. Registration details of FAF images are in Supplementary Methods. The cohort includes eyes with various retinal pathologies and eyes without macular disease. All images were acquired on Optos ultra-widefield FAF platforms (models P200DTx, 200Tx, and P200TxE). The number of frames per eye varies from 2 to 10, with irregular clinical intervals (median 0.51 years, IQR [0.16, 1.40]). An 90/10 patient- level split yields 1,001 held-out test eyes. Optos Stargardt disease dataset (zero-shot rare-disease evaluation). Stargardt macular dystrophy is the most common inherited macular dystrophy, producing progressive expansion of hypoautofluorescent RPE lesions on FAF that share some, but not all, morphological similarities to age-related macular degeneration geographic atrophy lesions [29]. This independent cohort comprises 195 patients (41.6 ± 18.6 years; 55.4% female) with clinically confirmed Stargardt disease, contributing 288 eyes and 699 FAF images with inter-visit intervals of median 2.01 years (IQR [1.02, 3.00]) acquired on Optos ultra-widefield FAF platforms (models 200Tx, P200DTx, and P200TxE). Spectralis Stargardt cross-device dataset (zero-shot cross-device evaluation). This independent cohort comprises 71 patients with Stargardt disease, contributing 125 eyes and 360 FAF images acquired on Heidelberg Spectralis platforms (30° and 55° field-of-view blue autofluorescence). Inter-visit intervals have a median of 1.01 years (IQR [0.36, 1.96]) with a median follow-up duration of 1.73 years (IQR [0.76, 3.87]). Glaucoma en-face SLO dataset (zero-shot cross-modality evaluation). To test generalization beyond FAF imaging entirely, we evaluate the FAF-trained model on en-face scanning laser ophthalmoscopy (SLO) images from a primary open-angle glaucoma (POAG) cohort of 12,363 patients (18,547 eyes with ≥2 visits; drawn from a total dataset of 26,166 patients, 51,690 eyes, and 85,866 images; 63.1 ± 13.8 years, 57.3% female; inter-visit intervals of median 1.17 years, IQR [1.00, 2.07]). All images were acquired on Zeiss Cirrus HD-OCT systems as en-face SLO from optic disc cube scans. Registration details of SLO images are in Supplementary Methods. This dataset presents two simultaneous domain shifts: different imaging physics (RNFL reflectance vs RPE autofluorescence) and different disease progression (gradual, patchy RNFL thinning vs focal atrophy expansion). All deep learning methods (TRU, ImageFlowNet, TADM, and 2D-BrLP) were trained exclusively on FAF data and comparisons among them reflect relative cross-domain robustness rather than SLO-optimized performance. Table 1: Dataset characteristics. Dataset Purpose Modality Disease Patients (N) Eyes (N) Images (N) Split Strategy Zero- Shot? Optos Mixed- Disease Primary Training & Eval FAF RPE Pathologies, Healthy 8,718 9,942 24,419 90/10 Patient- level No Optos Stargardt Rare- Disease Transfer FAF Stargardt Dystrophy 195 288 699 No split Yes Spectralis Stargardt Cross- Device Transfer FAF Stargardt Dystrophy 71 125 360 No split Yes Glaucoma SLO Cross- Modality Transfer SLO Primary Open- Angle Glaucoma 12,363 18,547 52,723 No split Yes 4.2 Evaluation Metrics We prioritize our evaluation metrics to promote eventual clinical utility, emphasizing measures of disease dynamics through lesion-level spatial accuracy over standard pixel-level reconstruction fidelity. This ordering reflects the clinical priority of the prediction task: a model that correctly captures the rate and spatial pattern of disease progression is more useful than one that merely reconstructs a plausible-looking retinal image. Tier 1: Progression Dynamics We calculate the change-map SSIM [30] (훥SSIM) as the primary progression metric. A structural metric that quantifies whether a model captures the spatial pattern of temporal change. For each eye, we compute two difference images: 훿 gt =퐼 ∗ −퐼 푁 (ground-truth change) and 훿 pred =퐼 ̂ −퐼 푁 (predicted change), where 퐼 ∗ is the ground-truth target, 퐼 ̂ is the prediction, and 퐼 푁 is the most recent history frame, all in a normalized [ 0,1 ] intensity space. 훥SSIM is the structural similarity index between these two change maps, computed within the bounding box of the image validity mask with a data range of 2.0 (reflecting the [ −1,1 ] range of difference images). A 훥SSIM of 1.0 indicates that the predicted change is structurally identical to the true change; low values indicate that the model fails to capture the spatial pattern of disease progression even if the static prediction appears reasonable. Unlike a simple difference of two SSIM values, this metric operates directly on the change signal and is therefore sensitive to the spatial distribution of progression, not merely whether the predicted image globally resembles the target. Tier 2: Lesion-Level Accuracy To assess whether models preserve clinically relevant hypoautofluorescent structure, we report binary-mask-derived metrics, Dice [31] and 95th percentile Hausdorff distance (HD95) [32]. Although our model does not predict segmentation masks, we compute these metrics by generating atrophy masks post-hoc, because lesion-level measurements remain the clinical currency for longitudinal monitoring of RPE diseases. Instead of using a learned segmentation model to generate the atrophy masks, we use an algorithmic segmentation method where adaptive intensity thresholding is applied to the predicted and ground-truth images. Critically, the same algorithm with identical parameters is applied to both the predicted and ground-truth images, ensuring that any systematic bias in the segmentation (over- or under-detection of atrophy) affects both sides of the comparison equally. We use this transparent, parameter-explicit algorithmic segmentation to avoid the “black-box” confound of using a second learned model to grade the first. The thresholding algorithm and parameters are detailed in Supplementary Methods. Lesion-level metrics are used only in the FAF dataset, because glaucomatous RNFL loss on SLO does not have discrete dark foveal lesions. Tier 3: Pixel-Level Reconstruction Fidelity Standard pixel-level metrics are reported as secondary measures of overall image quality. These include mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and Structural similarity index (SSIM). These choices reflect more clinically-relevant progression metrics rather than learned perceptual metrics, such as LPIPS [33], because they would penalize the correct suppression of clinically uninformative acquisition texture that our conditional-mean predictions produce by design [34]. 4.3 Baseline Methods Copy-last. The most recent history frame 퐼 푁 , representing the clinical status quo without any predictive model. This is a particularly informative baseline for slowly progressing retinal diseases, where short inter-visit intervals produce targets that differ only subtly from the most recent observation. A model’s improvement over copy-last demonstrates that the model is adding predictive information in the time domain. Linear spline. Per-pixel linear regression through the two most recent history frames, extrapolated to the target timepoint. This is the simplest model capable of temporal extrapolation and serves to test whether learned temporal representations provide benefit over purely linear pixel-level dynamics. ImageFlowNet. Neural ODE-based continuous-time longitudinal prediction with multiscale contrastive regularization [9]. We trained ImageFlowNet on our FAF dataset using the published codebase with the recommended hyperparameters (smoothness weight 0.10, latent weight 0.001, contrastive weight 0.01). ImageFlowNet represents the current benchmark for continuous-time longitudinal retinal image prediction. TADM. A diffusion-based approach to longitudinal medical image prediction designed for brain MRI [11]. TADM models the temporal residual between consecutive timepoints in a standard epsilon-prediction diffusion framework, conditioning on a single input frame via a dedicated condition encoder. It represents the most direct application of conditional diffusion to the longitudinal prediction task. 2D-BrLP (2D Brain Latent Progression). A latent diffusion approach to longitudinal medical image prediction adapted from BrLP [12]. The adapted model follows the three-stage pipeline in the original method. Specifically, a variational autoencoder (VAE) [35], a diffusion U-Net with epsilon, and a ControlNet [36]. The total model comprises 247.91M trainable parameters, an order of magnitude larger than TRU (22.8M). Inference follows BrLP’s published protocol: 50- step DDIM sampling with Latent Average Stabilization (averaging 푚=5 independent latent samples before decoding). Adaptation details and per-stage parameter counts are in Supplementary Methods. Additional methods considered. We evaluated several additional longitudinal prediction methods for inclusion in the benchmark. SADM was designed for longitudinal brain MRI generation with a sequence-aware attention mechanism. Δ-LFM applies flow matching in a learned latent space for patient-specific trajectories. Despite substantial adaptation efforts, neither codebase could be successfully trained on our FAF data within the ultra-widefield retinal imaging domain: SADM's sequence conditioning mechanism assumes a fixed temporal grid that is incompatible with the irregular clinical visit intervals in our dataset, and Δ-LFM's latent space training pipeline did not converge when adapted from brain MRI to retinal FAF resolution and intensity characteristics. Rather than report results from inadequately trained models, we restrict the benchmark to methods we could train to convergence with validated hyperparameters. 5. Results: Benchmarking and Generalization 5.1 Primary Benchmark: Mixed-Disease FAF Hold-out We evaluated all methods on the held-out FAF test set (1,001 eyes), which includes various retinal pathologies, and eyes without macular disease. TRU achieved the highest score on every metric in this benchmark (Table 2). On the two clinically most informative measures, TRU’s ΔSSIM (0.327) exceeded the next-best method (2D-BrLP, 0.303; p = 1.83 × 10 −6 ), and its atrophy Dice (0.671) exceeded copy-last (0.657; p = 0.061) and 2D-BrLP (0.649; p = 4.29 × 10 −4 ). On pixel-level metrics, TRU improved SSIM by 0.067 over the copy-last clinical baseline (p = 2.13 × 10 −156 ) and by 0.026 over the strongest deep learning baseline, 2D-BrLP (p = 3.23 × 10 −141 ). The improvement over copy-last was most pronounced on ΔSSIM (+0.070), confirming that TRU captured spatial progression patterns beyond what is available in the most recent observation alone. Among the deep learning baselines, 2D-BrLP and ImageFlowNet achieved comparable pixel- level performance to each other (SSIM 0.577 vs 0.575), with 2D-BrLP producing a higher ΔSSIM (0.303 vs 0.286), suggesting better temporal change prediction despite similar overall reconstruction fidelity. TADM, the conditional diffusion baseline, performed near copy-last on ΔSSIM (0.260 vs 0.257) and below all other deep learning methods on every metric. To confirm that model rankings on the atrophy segmentation metrics are not artifacts of threshold parameters in the adaptive segmentation method, we conducted a sensitivity analysis and showed that the relative ordering of model performance was preserved across the majority of settings (Supplementary Results, Table S1) Table 2. Cross-cohort benchmark of TRU against state-of-the-art longitudinal medical image prediction methods and classical references. Four cohorts are organized by increasing distribution shift relative to the FAF mixed- disease training set: (A) in-distribution Optos FAF hold-out; (B) zero-shot Optos Stargardt (disease shift); (C) zero- shot Heidelberg Spectralis Stargardt (vendor + disease shift); (D) zero-shot glaucoma en-face SLO (modality + disease shift). All deep learning methods were trained exclusively on the Optos FAF mixed-disease training set with no Stargardt or SLO cases. Values are mean ± SD across eyes. Best value per metric per cohort is shown in bold. Arrows indicate metric direction (↑ higher is better, ↓ lower is better). Dice and HD95 are not applicable (n/a) to glaucoma SLO because glaucomatous RNFL loss does not produce discrete foveal lesions. The SLO cohort is included as a scope-boundary demonstration; absolute performance on SLO is well below the FAF benchmark and does not indicate clinical deployability in the SLO domain (see Section 5.4). Method ΔSSIM ↑ SSIM ↑ PSNR ↑ MAE ↓ Dice ↑ HD95 ↓ A. Optos FAF mixed-disease hold-out (N = 1,001 eyes) copy-last 0.257 ± 0.11 0.534 ± 0.11 22.2 ± 2.7 0.0624 ± 0.021 0.657 ± 0.26 11.8 ± 16 linear spline 0.196 ± 0.14 0.462 ± 0.18 20.4 ± 4.5 0.0880 ± 0.071 0.605 ± 0.28 14.8 ± 18 ImageFlowNet 0.286 ± 0.095 0.575 ± 0.10 22.7 ± 2.6 0.0587 ± 0.019 0.633 ± 0.26 12.0 ± 15 2D-BrLP 0.303 ± 0.10 0.577 ± 0.092 22.7 ± 2.7 0.0587 ± 0.020 0.649 ± 0.26 12.0 ± 16 TADM 0.260 ± 0.13 0.543 ± 0.096 22.3 ± 2.3 0.0615 ± 0.019 0.621 ± 0.25 12.4 ± 16 TRU (ours) 0.327 ± 0.14 0.601 ± 0.093 23.2 ± 2.4 0.0559 ± 0.017 0.671 ± 0.24 11.1 ± 15 B. Optos Stargardt — zero-shot rare-disease transfer (N = 288 eyes) copy-last 0.294 ± 0.13 0.568 ± 0.11 22.5 ± 2.9 0.0581 ± 0.021 0.702 ± 0.25 13.7 ± 18 linear spline 0.233 ± 0.16 0.514 ± 0.15 21.3 ± 3.7 0.0712 ± 0.043 0.675 ± 0.27 15.2 ± 19 ImageFlowNet 0.311 ± 0.086 0.593 ± 0.11 23.1 ± 2.5 0.0549 ± 0.016 0.680 ± 0.26 17.2 ± 23 5.2 Zero-Shot Within-Vendor Rare-Disease Transfer: Stargardt Macular Dystrophy We evaluated zero-shot transfer to Stargardt macular dystrophy on the same Optos ultra- widefield platform used for training (288 eyes). All deep learning models were trained with the same mixed-disease FAF training dataset that do not contain any identified Stargardt patients based on diagnosis codes. Despite having encountered no Stargardt cases during training, TRU achieved strong prediction performance that exceeded in-distribution FAF metrics on several measures (e.g., SSIM 0.615 vs 0.601; ΔSSIM 0.351 vs 0.327), suggesting that the morphological similarity between Stargardt atrophy and other RPE pathologies on FAF enables effective anatomical transfer (Table 2). 2D-BrLP 0.331 ± 0.11 0.590 ± 0.108 23.0 ± 2.9 0.0552 ± 0.020 0.703 ± 0.25 13.5 ± 18 TADM 0.272 ± 0.13 0.568 ± 0.112 22.6 ± 2.4 0.0585 ± 0.018 0.683 ± 0.27 15.4 ± 20 TRU (ours) 0.351 ± 0.15 0.615 ± 0.11 23.7 ± 2.4 0.0520 ± 0.016 0.704 ± 0.25 14.4 ± 20 C. Heidelberg Spectralis Stargardt — zero-shot cross-vendor transfer (N = 125 eye-pairs) copy-last 0.245 ± 0.11 0.671 ± 0.11 20.2 ± 3.7 0.0761 ± 0.032 0.659 ± 0.24 18.2 ± 15 linear spline 0.148 ± 0.16 0.558 ± 0.20 17.9 ± 4.8 0.113 ± 0.079 0.582 ± 0.26 22.3 ± 17 ImageFlowNet 0.265 ± 0.087 0.689 ± 0.11 20.6 ± 3.6 0.0729 ± 0.029 0.671 ± 0.22 17.5 ± 14 2D-BrLP 0.257 ± 0.082 0.659 ± 0.10 20.4 ± 3.5 0.0745 ± 0.030 0.645 ± 0.25 18.8 ± 16 TADM 0.191 ± 0.12 0.598 ± 0.10 20.4 ± 2.1 0.0771 ± 0.020 0.633 ± 0.22 21.0 ± 16 TRU (ours) 0.272 ± 0.15 0.695 ± 0.10 21.8 ± 2.4 0.0670 ± 0.020 0.669 ± 0.20 18.7 ± 14 D. Glaucoma en-face SLO — zero-shot cross-modality transfer (N = 18,547 eyes) copy-last 0.110 ± 0.038 0.342 ± 0.10 16.5 ± 2.0 0.109 ± 0.023 n/a n/a linear spline 0.0616 ± 0.060 0.267 ± 0.13 14.7 ± 3.0 0.149 ± 0.068 n/a n/a ImageFlowNet 0.129 ± 0.062 0.380 ± 0.10 17.0 ± 2.0 0.103 ± 0.021 n/a n/a 2D-BrLP −0.003 ± 0.16 0.373 ± 0.09 13.8 ± 1.4 0.173 ± 0.031 n/a n/a TADM 0.102 ± 0.089 0.350 ± 0.09 17.2 ± 1.4 0.109 ± 0.020 n/a n/a TRU (ours) 0.134 ± 0.089 0.392 ± 0.10 17.4 ± 1.6 0.104 ± 0.020 n/a n/a TRU achieved the highest score on all pixel-level metrics and the highest Dice (0.704) among all methods. However, 2D-BrLP achieved better HD95 (13.50 pixels) than TRU (14.39). The dissociation between Dice (where TRU led by 0.001, essentially tied) and HD95 (where 2D- BrLP led by 0.89 pixels) reflects a known and expected tradeoff of conditional-mean prediction. Specifically, TRU averages over unpredictable acquisition variability, which suppresses clinically uninformative noise but also produces slightly softer lesion boundaries than a stochastic latent-space model that samples from a higher-entropy output distribution. This tradeoff is discussed further in Limitations. More importantly, similar to what was observed from the FAF benchmark, model differentiation was stronger on pixel-level and change-map metrics than on lesion-level metrics. This is consistent with the task structure analyzed in 6.3, which shows that the disease-relevant change signal is small, and that the primary value of dense image prediction lies in capturing continuous intensity patterns that binary atrophy metrics do not fully reflect. Figure 2 shows representative Stargardt predictions and change maps across all methods. Figure 2. Qualitative prediction examples on the Stargardt hold-out set to demonstrate the zero-shot model performance. Each panel (blue box) contains one representative eye. The top row in each panel shows the longitudinal history frames, ground-truth target, and predictions from all evaluated methods. The bottom row in each panel shows the change maps display the pixel-wise difference between the predicted image and the most recent history frame, with blue indicating darkening (new atrophy) and red indicating brightening relative to the last observation; the ground-truth change map (green border) serves as the reference. Copy-last produces a blank change map by construction, confirming that it captures no progression signal. Among deep learning methods, TRU's change maps most closely reproduce the spatial pattern and magnitude of the ground-truth change signal, particularly the localized darkening at lesion expansion boundaries. Pixel-wise difference maps between each model's prediction and the ground-truth image are provided in Supplemental Figure 1. ‘IFN’ indicates ImageFlowNet. 5.3 Zero-Shot Cross-Vendor Transfer: Heidelberg Spectralis Stargardt The Optos Stargardt evaluation in Section 5.2 tests generalization across disease labels but within a single FAF imaging platform. We therefore evaluated zero-shot transfer to an independent Stargardt cohort from Heidelberg Spectralis, with no retraining, no fine-tuning, and no architectural modification. This cohort tests simultaneous vendor shift and disease shift relative to the Optos training distribution. One cohort property should be noted before reading the benchmark. The Spectralis FAF field of view (~30°, macula-centered) is substantially narrower than the Optos ultra-widefield field of view (~200°), so after macula-centered cropping to 256×256 the Spectralis frames contain proportionally more temporally stable macular anatomy. This elevates the copy-last reference on Spectralis (SSIM 0.671) relative to Optos Stargardt (SSIM 0.568, Table 2), and method comparisons should be interpreted against the cohort-specific copy-last baseline rather than across cohorts. TRU achieved the best performance on pixel-level fidelity metrics (i.e MAE and PSNR) with statistical significance (all comparisons p < 10⁻³) (Table 2). TRU also achieved the highest SSIM (0.695), ΔSSIM (0.272) in the cohort, although the improvements over some models did not reach statistical significance (P < 0.05). On the segmentation metrics, the performance of all learned models was similar and none of the differences reached statistical significance (P<0.05). 5.4 Zero-Shot Cross-Modality Transfer: Glaucoma en-face SLO As a final stress test for the generality of TRU's learned temporal representations, extending beyond the vendor shift in Section 5.3 to a simultaneous disease and modality shift, we evaluated the FAF-trained model on en-face SLO from a glaucoma cohort (18,547 eyes) with no SLO training data or architectural modification. All deep learning methods were trained exclusively on FAF; comparisons reflect relative cross-domain robustness on equal footing. Because glaucomatous RNFL loss does not produce discrete foveal lesions, atrophy metrics are not applicable. TRU achieved the highest SSIM (0.392), PSNR (17.37), and ΔSSIM (0.134), with all improvements highly significant (p < 10⁻¹⁰) (Table 2). The absolute performance gap between SLO and FAF (SSIM 0.39 vs 0.60) delineates the boundary beyond which modality-specific training is required. The most revealing comparison is 2D-BrLP's failure: negative ΔSSIM (−0.003) and PSNR below linear spline (13.80 vs 14.71), demonstrating that latent representations tightly coupled to FAF statistics are fragile under domain shift. TRU's pixel- space residual formulation, by contrast, partially generalizes across modality-specific statistics. We interpret this result as demonstrating boundary robustness, not clinical deployability in the SLO domain. Figure 3 summarizes the cross-cohort benchmark on the three metrics most directly tied to the prediction task. The pattern is consistent across the domain-shift hierarchy: TRU achieves the highest mean ΔSSIM and SSIM in every cohort, and the highest mean Dice on the in-distribution FAF and zero-shot Optos Stargardt cohorts. On Spectralis Stargardt, where vendor and disease shift are combined, the deep-learning methods converge on Dice (Section 5.3); the change-map and pixel-fidelity advantages persist. On glaucoma SLO, where modality and disease both shift, all methods degrade in absolute terms, but the relative ordering on ΔSSIM and SSIM holds. The advantage of the proposed method is therefore not a single-cohort artifact but a property that survives each successive layer of distribution shift. Figure 3. Cross-cohort benchmark of TRU against state-of-the-art longitudinal medical-image-prediction methods. Bars show mean values; error bars are ±1 SD across eyes within each cohort. SD bars reflect across-eye variability and do not reflect the per-eye paired comparison structure used in the statistical tests; pairwise Wilcoxon comparisons in Sections 5.1–5.4 establish statistical significance on most cells where TRU leads. Cohorts are ordered left-to-right by increasing distribution shift relative to the FAF mixed-disease training set. Within each cohort, four learned methods are compared: ImageFlowNet, TADM, 2D-BrLP, and TRU (proposed, blue). The dashed line within each cohort group marks the copy-last reference (i.e., predicting that the future image equals the most recent observation), which is the relevant clinical no-prediction baseline. Triangles mark the best-performing method with statistical significance per cohort per metric. A. Change-map ΔSSIM. B. Pixel SSIM. C. Atrophy Dice; the SLO cohort is omitted from this panel because glaucomatous RNFL loss does not produce discrete foveal lesions. The trivial linear-spline baseline (consistently below copy-last on all metrics in all cohorts) is omitted from this figure for clarity but reported in Tables 2, including PSNR, MAE, and HD95 results. 5.5 Effect of History Length on Prediction Quality We assessed the effect of history lengths on benchmarking performance by evaluating all methods on subsets of the FAF hold-out stratified by the number of visits per eye: ≥3 visits (N = 273) and ≥4 visits (N = 100). These subsets represent eyes with richer temporal follow-up and longer prediction horizons with greater disease progression. TRU’s advantage over all competing methods grew monotonically with the number of available history frames (Table S8). On ΔSSIM, the metric most directly measuring temporal change prediction, TRU’s advantage over the next-best method more than tripled from +0.024 (all eyes) to +0.078 (≥4 visits). On PSNR, the advantage nearly doubled from +0.52 dB to +1.02 dB. Interestingly, on atrophy Dice, TRU’s advantage over copy-last was non-significant for the full cohort (p = 0.061) but became significant for eyes with ≥3 visits (p = 2.77 × 10 −4 ). These results overall support that TRU’s benefit emerged specifically in the subset with richer history and more active disease, precisely the patients for whom longitudinal prediction is most clinically consequential. 6. The Central Analysis: Why Alignment, Not Framework, Is the Bottleneck Section 5 establishes that TRU outperforms published baselines across three evaluation domains. This section addresses why, through a controlled analysis that isolates the contribution of distributional alignment between training and inference inputs. What we mean by 'distributional alignment.' A model trained to denoise corrupted versions of image X will, at deployment, perform best when the image it receives is also a corrupted version of X. If training noises the target image, I*, which is unavailable at deployment, and inference instead starts from the most recent image, I N , the model sees a distribution at deployment that it never saw during training. We call this gap a ‘training–inference input mismatch’, and we call the absence of it ‘distributional alignment’. Table 3 summarizes the model configuration variants we compare in this section. Table 3. Configuration taxonomy. The five configurations share one architecture, one training dataset, and one set of optimizer hyperparameters. Config Training: which image is noised Inference: initialized from Training & inference inputs match? Inference steps Framework Std-DDIM- 50step I* (target) Pure Gaussian noise No (maximum mismatch) 50 Stochastic Std-DDIM- 1step I* (target) Noised I N Partial (inference side only) 1 Stochastic IA- Nonlinear I N (last observation) Noised I N Yes (full alignment) 1 Stochastic IA-Linear I N (last observation) Noised I N (linear schedule) Yes (full alignment) 1 Stochastic TRU I N (last observation) I N (no noise) Yes (full alignment) 1 Deterministic 6.1 The Mismatch Diagnosis: Inference Alignment as the Primary Performance Driver We isolated the contribution of this mismatch through a controlled comparison of the model variants and decomposed the alignment effect through two targeted comparisons. (i) Comparing between Std-DDIM-50step and Std-DDIM-1step. Std-DDIM-50step is the standard diffusion setup, where the model was trained to denoise corrupted versions of the target image I*, and at inference it was initialized from pure Gaussian noise and iterated for 50 steps. Std-DDIM-1step shares the identical trained weights but instead of starting from pure noise at inference, it started from the patient's last observation I N corrupted to the same noise level. This comparison therefore isolates the contribution of aligning the inference input to the patient's anatomy, with training held fixed. From Table 4, we see that switching from Std-DDIM-50step to Std-DDIM-1step produced large improvements across all metrics. Specifically, SSIM improved from 0.516 to 0.595 (+0.079, p = 2.48 × 10−165), ΔSSIM from 0.246 to 0.292 (+0.046, p = 6.11 × 10−22), PSNR from 21.7 to 22.8 dB (+1.03, p = 2.61 × 10−62), and atrophy Dice from 0.609 to 0.645 (+0.036, p = 5.06 × 10−11). Critically, Std-DDIM-50step fell below copy_last on both SSIM (0.516 vs 0.534) and ΔSSIM (0.246 vs 0.257), demonstrating that a model initialized from pure noise actively degraded temporal prediction relative to making no prediction at all. Initializing inference from I N partially addressed the bottleneck by anchoring the starting point to the patient’s actual retinal anatomy. (i) Comparing Std-DDIM-1step and IA-Nonlinear. Std-DDIM-1step and IA-Nonlinear shared the same inference procedure but differ in which image was corrupted during training. Std- DDIM-1step is trained on noised I* and therefore sees a distribution at deployment (noised I N ) that it never sees during training; IA-Nonlinear was trained on noised I N and so the training distribution matches what it receives at deployment. This training-side alignment yielded further significant improvements: ΔSSIM from 0.292 to 0.328 (+0.036, p = 5.04 × 10−12), SSIM from 0.595 to 0.602 (+0.007, p = 4.20 × 10−26), PSNR from 22.77 to 23.26 dB (+0.49, p = 2.67 × 10−27), and atrophy Dice from 0.645 to 0.671 (+0.027, p = 1.20 × 10−6). Together, these two comparisons decomposed the path from maximum mismatch (Std-DDIM- 50step, SSIM 0.516) to full alignment (IA-Nonlinear, SSIM 0.602) into two additive contributions. On ΔSSIM, the primary clinical metric, the inference-side correction contributed +0.046 and the training-side alignment contributes a further +0.036, with both reaching high significance. The combined evidence established that, at the training scale tested here, the distributional mismatch between training and inference inputs is the operative bottleneck for longitudinal prediction quality. The mismatch penalty was so significant that Std-DDIM-50step produced predictions worse than the trivial copy-last baseline. The practical consequence persists under distribution shift on both the Optos Stargardt cohort (Supplementary Table S2) and, more stringently, under simultaneous vendor and disease shift on the Heidelberg Spectralis Stargardt cohort (Supplementary Table S4). All alignment-progression comparisons remain highly significant on Spectralis (all p < 0.005). The finding that training– inference mismatch is the dominant bottleneck is therefore reproduced across disease shift within vendor, and across the simultaneous vendor and disease shift in Section 5.3. Table 4: Alignment progression on the FAF hold-out (N = 1,001 eyes). Three conditioning configurations ordered by degree of distributional alignment, from maximum mismatch to full alignment. copy_last included as the no-prediction reference. Method ΔSSIM ↑ SSIM ↑ PSNR ↑ MAE ↓ Dice ↑ HD95 ↓ copy_last 0.257 ± 0.11 0.534 ± 0.11 22.2 ± 2.7 0.0624 ± 0.021 0.657 ± 0.26 11.8 ± 16 Std-DDIM-50step 0.246 ± 0.12 0.516 ± 0.091 21.7 ± 2.2 0.0658 ± 0.018 0.609 ± 0.25 12.3 ± 14 Std-DDIM-1step 0.292 ± 0.12 0.595 ± 0.099 22.8 ± 2.8 0.0587 ± 0.021 0.645 ± 0.26 11.8 ± 15 IA-Nonlinear 0.328 ± 0.14 0.602 ± 0.093 23.3 ± 2.4 0.0557 ± 0.017 0.671 ± 0.24 11.0 ± 15 6.2 The Equivalence Principle: Noise Schedule and Stochastic Corruption Are Secondary to Alignment. Section 6.1 establishes that distributional alignment is the primary performance driver. A natural question follows: given alignment, does the specific form of stochastic corruption or its presence at all provide additional benefit? We test this by comparing three inference-aligned configurations that span a spectrum from stochastic corruption with the nonlinear DDPM- derived [15] noise schedule (IA-Nonlinear), through stochastic corruption with a linear interpolation schedule (IA-Linear), to fully deterministic regression with no noise injection (TRU). We found that the three aligned configurations produced nearly identical results (Table S9). SSIM spanned a range of 0.0005 (0.6012–0.6017), ΔSSIM spanned 0.007 (0.321–0.328), and no single method dominated across all metrics: IA-Nonlinear achieved the highest PSNR and lowest HD95, IA-Linear achieved the highest SSIM and atrophy Dice, and TRU achieved the highest ΔSSIM and lowest MAE. This lack of consistent ordering already suggests that the differences are noise rather than signal. We confirmed this through formal statistical testing (Table S10). No comparison between TRU and either stochastic aligned configuration reached significance on ΔSSIM, SSIM, or atrophy Dice (all p > 0.05). On the primary clinical metrics, direct regression and the stochastic frameworks were statistically indistinguishable. TRU vs IA-Nonlinear showed no significant difference on any metric (all p > 0.30). TRU vs IA-Linear showed significance only on PSNR (p = 0.003), but the absolute effect size was 0.10 dB, clinically imperceptible and an order of magnitude smaller than the alignment effects in Section 6.1 (+0.49 to +1.03 dB). This isolated significance on a secondary metric, achieved with N = 1,001, reflects the statistical power of the sample size rather than a meaningful performance distinction. The pattern was consistent across all evaluated metrics. Once distributional alignment was achieved, stochastic noise injection, whether following a nonlinear DDPM-derived schedule or a linear interpolation schedule, did not outperform deterministic regression on any clinically relevant metric (SSIM 0.601–0.602, ΔSSIM 0.321–0.328, atrophy Dice 0.671–0.676). The equivalence pattern replicates on both out-of-distribution cohorts. On Optos Stargardt (Supplementary Table S3), no primary clinical metric separated TRU from either stochastic aligned configuration. On Heidelberg Spectralis Stargardt, where vendor and disease shift are combined, TRU versus IA-Nonlinear and TRU versus IA-Linear again show no significant difference on ΔSSIM, atrophy Dice, or HD95 (all p > 0.17; Supplementary Table S5); isolated significance appears only on pixel-level secondary metrics with clinically imperceptible effect sizes (e.g., TRU vs IA-Linear PSNR advantage 0.33 dB, p = 0.007), matching the pattern observed on Optos. Once distributional alignment is achieved, the choice among aligned training frameworks does not affect prediction quality on primary clinical metrics under any of the three evaluation cohorts. Given this statistical equivalence, we adopted the simplest model, TRU (direct regression), as the primary model on parsimony grounds, allowing efficient training and fully reproducible inference. 6.3 Mechanism of Equivalence: Why Stochastic Sampling Provides No Benefit Sections 6.1 and 6.2 establish that inference alignment produces large performance gains, and once aligned, stochastic configurations and deterministic regression converge to statistically indistinguishable prediction quality. This equivalence invites explanation. Diffusion models offer two properties that direct regression does not: trajectory diversity (the ability to sample multiple plausible futures from a stochastic posterior) and iterative refinement (the ability to traverse a learned score field across multiple denoising steps). Both advantages require that the conditional posterior 푝 ( 퐼 ∗ ∣퐼 푁 ,ℋ ) has meaningful variance. To examine the conditional posterior distribution of our longitudinal FAF datasets, we performed present two complementary analyses, where one characterizes the task itself, independent of any model, and the other one directly measures the posterior width learned by the stochastic models. 6.3.1 Task Characterization: Inter-Visit Change Is Substantial but Unpredictable To characterize the intrinsic entropy of the prediction task without reference to any model, we computed per-pixel absolute intensity change | 훿 | = | 퐼 ∗ −퐼 푁 | between the most recent history frame and the prediction target across all 9,009 ( 퐼 푁 ,퐼 ∗ ) pairs in the FAF training set, encompassing 587.9 million valid pixels within the fundus mask (Figure 4a). Inter-visit pixel change is not negligible. Specifically, approximately half of all pixels exceed 5% intensity change between consecutive visits, and the median absolute change is 4.7% of the full intensity range. However, the critical observation is that this change is largely time-invariant. Table 5a stratifies the same statistics by inter-visit interval. If inter-visit change were dominated by disease progression, it would scale with time: longer intervals would produce proportionally more change, as atrophy has more time to expand. Instead, the changed pixel fraction is nearly constant across strata (0.486–0.507) and the correlation between inter-visit interval and changed fraction is weak (푟=0.116) (Figure 4b). This indicates that the dominant source of inter-visit pixel change is acquisition variability, such as differences in illumination, instrument alignment, or intensity normalization between images, rather than disease progression. The temporally structured disease signal that prediction models can learn to capture is a relatively small component of total inter-visit difference, embedded within a larger background of time-invariant imaging noise. This decomposition explains the structure of the prediction task. The conditional distribution of future images 푝 ( 퐼 ∗ ∣퐼 푁 ,ℋ ) has substantial total variance (driven by unpredictable acquisition variability), but the predictable component of that variance, meaning the disease progression signal, is highly constrained. Stochastic sampling from the full posterior would primarily produce diversity in the acquisition-noise dimensions, which are clinically uninformative. The disease-relevant signal, which determines clinical utility, occupies a low-dimensional subspace of the overall change, and within that subspace the posterior is sharply peaked. The task-entropy signature is not specific to Optos ultra-widefield imaging. Replicating this analysis on the Heidelberg Spectralis Stargardt cohort (Supplementary Table S6) reproduces the Optos pattern to within three significant figures on most statistics: the median absolute pixel change is identical (0.0471), the 99th percentile identical (0.2706), the fraction of pixels below 5% change is 50.2% versus 51.2% on Optos, and the correlation between inter-visit interval and changed-pixel fraction is r = 0.138 versus r = 0.116. The higher copy-last SSIM on Spectralis (0.684 vs 0.525 on Optos consecutive-visit pairs) is attributable to the narrower Spectralis field of view, which places proportionally more temporally stable macular anatomy in the frame (Section 5.3), and does not reflect a different entropy regime. That the task-entropy signature reproduces across confocal narrow-field and non-confocal ultra-widefield FAF acquisition with different intensity dynamics and different registration-pipeline behavior indicates that the low- entropy characterization is a property of slowly progressing FAF as a prediction task, not an artifact of the specific Optos acquisition pipeline. Figure 4. Task entropy characterization of inter-visit change. (a) Distribution of per-pixel absolute intensity change |δ| = |I* − I N | across 587.9 million pixels from consecutive-visit pairs in the FAF training set (log-scale y- axis). Dashed lines indicate cumulative thresholds: 11.9% of pixels change by less than 1%, 51.2% by less than 5%, and 79.6% by less than 10% of the full intensity range. The broadly distributed profile confirms that inter-visit pixel change is substantial, not concentrated near zero. (b) Fraction of changed pixels (|δ| > 0.05) per eye as a function of inter-visit interval Δt, shown as hexbin density (log-scale color) with marginal histogram of Δt. The near-flat regression line (r = 0.116) indicates that the magnitude of inter-visit change is largely independent of elapsed time. If disease progression were the dominant source of change, longer intervals would produce proportionally more changed pixels. The weak temporal dependence indicates that the dominant source is time-invariant acquisition variability rather than temporally structured disease progression. Together, these panels establish that the prediction task occupies a low-entropy regime in which the predictable disease signal is a small component embedded within a larger background of unpredictable imaging noise (Section 6.3). Table 5. Task entropy and posterior concentration analyses of the longitudinal FAF prediction task. Panel A characterizes the intrinsic predictability of inter-visit change independently of any model, by stratifying per-pixel absolute intensity change |δ| = |I* − I N | by elapsed time Δt. Panel B measures the width of the learned conditional posterior p(I* | I N , history) by generating K = 10 predictions per eye from each stochastic aligned model while varying only the noise realization at inference, with all conditioning inputs held fixed. Together these two model- light measurements form the diagnostic framework proposed in Section 7.1 for assessing whether stochastic generative machinery is warranted on a candidate longitudinal imaging dataset. Replication of both analyses on the Heidelberg Spectralis Stargardt cohort is reported in Supplementary Tables S6 and S7. A. Task entropy: inter-visit change stratified by prediction interval Δt (FAF training pairs; N = 9,009; 587.9 million valid pixels). Stratum N eyes Frac. pixels |δ| < 5% Median changed fraction per eye (|δ| > 0.05) Mean |δ| Copy-last SSIM Short (Δt < 0.25 y) 3,493 0.514 0.486 0.063 ± 0.021 0.528 ± 0.107 Medium (0.25 ≤ Δt < 1 y) 2,915 0.521 0.478 0.062 ± 0.021 0.532 ± 0.108 Long (Δt ≥ 1 y) 2,601 0.500 0.507 0.066 ± 0.022 0.513 ± 0.107 All pairs (median Δt = 0.51y) 9,009 0.512 0.490 0.064 ± 0.021 0.525 ± 0.108 B. Posterior concentration: bias–variance decomposition of prediction error across K = 10 independent noise realizations per eye (FAF hold-out; N = 1,001 eyes). Metric IA-Nonlinear IA-Linear Inter-sample SSIM (mean ± SD across eyes) 0.99984 ± 0.00006 0.99998 ± 0.00001 Prediction MSE 5.5 × 10⁻³ 5.7 × 10⁻³ Inter-sample variance / Prediction MSE 0.015% 0.002% Bias² fraction of total error 99.99% 100.00% Variance fraction of total error 0.01% < 0.01% 6.3.2 Posterior Concentration: Stochastic Models Converge to a Single Output The task characterization above establishes that the predictable component of inter-visit change is small. We next verify that the stochastic models have learned this structure, but their posteriors have concentrated to effective point masses, producing the same prediction regardless of the noise realization used at inference. For each of the 1,001 eyes in the FAF hold-out set, we generated 퐾=10 independent predictions from IA-Nonlinear and IA-Linear, varying only the random seed that determines the noise realization at each model’s operating point. All other inputs were held constant. Table 5b reports inter-sample agreement and the bias–variance decomposition [37] of prediction error. The posteriors have concentrated entirely. Ten independent noise realizations, each producing a different corrupted version of the patient’s retina as input, yielded outputs that agreed to four decimal places of SSIM (0.9998 for IA-Nonlinear, 0.99998 for IA-Linear). The bias–variance decomposition was definitive: 99.99% of IA-Nonlinear’s prediction error and 100.00% of IA- Linear’s prediction error was systematic bias, with stochastic variance contributing less than 0.02% in both cases. The models have learned to project through the input noise to a single deterministic output, regardless of the noise realization. The inter-sample variance is substantially smaller relative to the prediction error; in fact, it is five orders of magnitude smaller (6.8×10 −7 vs 5.5×10 −3 ). No amount of additional sampling can reduce the systematic bias that constitutes effectively all of the remaining error. The posterior-collapse pattern reproduces on the Heidelberg Spectralis Stargardt cohort (Supplementary Table S7). IA-Nonlinear inter-sample SSIM 0.9997 and IA-Linear 1.0000, with stochastic variance contributing ≤ 0.01% of total prediction error in both cases. Because the Spectralis models were trained exclusively on Optos data and received no Spectralis inputs during training, the collapse is a property of the learned posterior rather than of the training distribution. Together, sections 6.3.1 and 6.3.2 establish that, across both evaluation cohorts, the stochastic models' posteriors had collapsed to effective point masses, and that this collapse reflected a property of the prediction task itself, where the predictable component of inter-visit change is small relative to time-invariant acquisition variability, leaving little for stochastic sampling to capture. This mechanistic account explains the equivalence observed in Section 6.2. 7. Discussion 7.1 Toward a task-adaptive design principle for longitudinal medical image prediction The controlled comparison in Section 6 shows that distributional alignment dominates framework choice at the training scale we evaluated, and the posterior concentration analysis in Section 6.3 explains why. These findings challenge an implicit assumption underlying the field's current methodological trajectory toward ever-increasing generative complexity, namely that the conditional distribution over plausible future images are broad enough to require stochastic sampling machinery. For slowly progressing retinal disease at the imaging resolutions and follow-up intervals tested here, this assumption fails. The predictable component of inter-visit change is small relative to time-invariant acquisition variability, the conditional posterior collapses to an effective point mass, and the additional model capacity, training cost, and inference complexity that diffusion frameworks introduce buy no measurable accuracy. We therefore propose, as a hypothesis for further empirical testing in other domains, the following design principle. The appropriate level of generative complexity for a longitudinal medical image prediction task should match the entropy of the predictable component of the conditional posterior, with distributional alignment between training and inference inputs as a prerequisite in all regimes. Operationally this principle is diagnostic rather than prescriptive. It does not prescribe deterministic regression universally. It prescribes a measurement to perform on a candidate dataset before committing to a framework. The two measurements that constitute the diagnostic are exactly those reported in Section 6.3. The first is the ratio of temporally structured to time-invariant inter-visit variance, computed from raw image pairs without any model, by stratifying inter-visit pixel change by elapsed time and asking whether the magnitude of change scales with the interval (Table 5a, Figure 4b). The second is the posterior concentration of a stochastic model trained on the same data, measured as inter-sample agreement across multiple noise realizations at fixed conditioning (Table 5b). A task in which the first measurement reveals weak temporal scaling and the second reveals collapse to a near point mass is one for which inference-aligned deterministic regression should be sufficient, and one in which the additional cost of stochastic machinery is unlikely to be repaid in accuracy. A task in which either measurement points the other way is one in which the stochastic machinery may be essential, and where deterministic regression would underperform. Whether other slowly progressing conditions occupy the same low entropy regime as FAF in age-related macular degeneration and Stargardt disease is an empirical question our data do not answer. Plausible candidates for the same regime include neurodegeneration on structural MRI over short follow-up, chronic glaucomatous optic neuropathy on SLO or OCT, and indolent tumor monitoring on serial imaging. Plausible candidates for the opposite regime, where stochastic machinery may be necessary, include long prediction horizons over which disease trajectories diverge, branching responses to therapeutic intervention, rapid or stochastic disease processes such as exudative AMD activity, and high variability imaging where the predictable signal is genuinely multimodal. These remain hypotheses, and the value of the diagnostic framework above is precisely that it provides a low-cost way to test them on any candidate dataset before committing to a modeling framework. A second qualification is internal to our characterization. The low entropy regime we identify reflects the predictable signal available through current FAF and SLO imaging, not the intrinsic predictability of the underlying disease biology. Future imaging modalities or multi-modal inputs that resolve earlier indicators of disease trajectory, for example OCT-derived photoreceptor or RPE-band biomarkers paired with FAF, could reveal predictable variance that current single- modality imaging does not capture, and could shift the same diseases into a regime in which stochastic sampling would provide genuine benefit. The principle is a statement about the prediction task as currently posed, not about the disease. 7.2 Clinical utility of deterministic, history-conditioned prediction Two properties of TRU together determine its suitability for clinical longitudinal monitoring, and each follows from a specific design choice rather than from incidental implementation details. The first property is deterministic inference. Because TRU is a single forward pass with no stochastic component at inference, the same patient history always produces the same predicted image. This is a precondition for clinical interpretation that stochastic models do not satisfy. A clinician comparing today's prediction with one generated last month cannot distinguish a meaningful change in the model's expectation from a sampling artifact unless they know the model is deterministic. Reproducibility under identical inputs is also a baseline expectation in regulatory evaluation rather than an optional feature. The second property is the compounding benefit of multi-frame temporal conditioning with history length. As reported in Section 5.5, TRU's predictive advantage grows monotonically with the number of available history frames. This scaling property is clinically consequential because the patients who accumulate the longest imaging histories are precisely those in active monitoring programs for whom progression prediction would most directly influence management decisions. A method whose advantage grows with the length of follow-up is, in expectation, a method whose marginal value to the clinic also grows over time. 7.3 Generalization across cohorts The three zero-shot evaluations in Section 5 provide converging evidence that TRU's learned temporal representations transfer beyond the training distribution in ways that are clinically meaningful but bounded in scope. The three regimes (i.e within-vendor disease shift, cross- vendor disease shift, and cross-modality cross-disease shift) each delineate a different boundary of that transfer. The Stargardt result is the most encouraging case. Prediction quality on an unseen rare inherited retinal disease (SSIM 0.615, atrophy Dice 0.704) approaches and on some metrics exceeds the in-distribution FAF benchmark, despite the absence of any Stargardt cases in the training set. The most straightforward interpretation is that the morphological similarity between Stargardt- related RPE atrophy and the broader spectrum of RPE pathologies represented in the training cohort enables effective anatomical transfer without disease-specific training. We note an alternative contributing factor that we cannot fully separate from the present data. The Stargardt cohort has a substantially longer median inter-visit interval than the FAF training set (2.0 years versus 0.5 years), which means the disease-relevant component of inter-visit change is, in relative terms, larger compared with the time-invariant acquisition variability that dominates short-interval pairs. In other words, the Stargardt evaluation cohort sits in a slightly higher signal-to-noise regime than the in-distribution FAF cohort, which would mechanically improve any model that captures the disease signal correctly. Both explanations may contribute and untangling them would require an evaluation cohort matched on inter-visit interval to the training set. The Heidelberg Spectralis Stargardt evaluation extends this finding across the most clinically important vendor boundary in FAF imaging. TRU retained a clear and statistically significant advantage over TADM and 2D-BrLP on all metrics, indicating that the core finding of the paper that complex generative pipelines are not necessary on this task holds under vendor shift and not only within vendor. Beyond the vendor shift evaluated in Section 5.3, the cross-modality SLO transfer is a more cautious finding that delineates where modality-specific training becomes necessary. TRU is overall the best performing model among all evaluated ones on the glaucoma SLO cohort. However, the absolute SSIM achieved on SLO (0.39) is well below the FAF benchmark (0.60), and we read this gap as the boundary beyond which modality-specific training becomes necessary. The SLO evaluation is best interpreted as a demonstration that TRU's temporal representations can generalize across modalities, not as a demonstration that FAF-trained models are clinically deployable on SLO without further adaptation. 7.4 Limitations Several boundary conditions qualify the generalizability of these findings. Training-regime scope. We establish the alignment principle at ~9,000 training eyes, a realistic single-institution cohort scale. Whether diffusion-based frameworks gain ground at substantially larger pretraining scales is an open empirical question; what we establish here is that at clinically realistic dataset scales for single-disease retinal prediction, alignment dominates framework choice. Spatial resolution. All training and evaluation used images cropped to the macular region and resampled to 256×256, which fits within the memory budget of a single 40 GB GPU but may obscure fine lesion boundary details and sub-resolution RPE changes visible at clinical viewing resolution. Whether the temporal representations learned at this resolution transfer to or benefit from higher-resolution inputs is an open question with direct implications for clinical utility. Image-based evaluation only. The current evaluation is purely image-based. We assess prediction quality through pixel-level, change-map, and lesion-level metrics, but we do not correlate predictions with functional outcomes such as visual acuity change, microperimetry sensitivity loss, or patient-reported outcomes. The proxy atrophy metrics reported here are derived from algorithmic intensity thresholding rather than expert annotation. The sensitivity analysis across 27 parameter combinations confirms that model rankings are robust to threshold choice (Supplementary Table S1), but validation against expert segmentation on a representative subset would provide stronger lesion-level evidence. Conditional-mean smoothing and edge sharpness. Predictions from conditional-mean models such as TRU are inherently smoother than clinical images because they average over unpredictable acquisition variability. This smoothing suppresses clinically uninformative noise while preserving the temporally consistent disease signal, a property that is advantageous for the change-detection task TRU is designed for, and that is reflected in TRU's strong performance on area-based metrics such as atrophy Dice. The same property has a measurable cost. The HD95 result on the Stargardt cohort, where 2D-BrLP achieves a tighter boundary estimate than TRU (13.50 vs 14.39 pixels), reflects this tradeoff directly. For atrophy area estimation, which is the quantity that drives treatment-timing decisions in geographic atrophy, TRU and 2D-BrLP are essentially equivalent on Stargardt. For boundary localization, the conditional-mean smoothing imposes a modest penalty. 7.5 Future work The clearest near-term extensions are clinical. Validation against expert-annotated lesion progression would strengthen the lesion-level evidence beyond algorithmic thresholding, and correlation with functional endpoints such as visual acuity and microperimetry would establish whether the temporally consistent disease signal TRU captures align with patient outcomes that matter clinically. A second direction tests the design principle itself. Incorporating multi-modal inputs such as OCT-derived photoreceptor and RPE-band biomarkers paired with FAF may reveal predictable variance that single-modality imaging misses, shifting the task into a higher-entropy regime where stochastic machinery would prescribe a different framework than the one we adopt here. Applying the diagnostic of Section 6.3 to other longitudinal medical imaging tasks, particularly those with longer horizons or more active disease dynamics, would test whether the principle generalizes or is specific to the low-entropy regime characterized here. 8. Conclusion This study establishes that distributional alignment between training and inference inputs, rather than the choice of generative framework, is the dominant determinant of prediction quality for longitudinal retinal imaging at the training scale evaluated. A controlled comparison of five conditioning configurations sharing one architecture and one training dataset isolates this effect, and the resulting deterministic direct-regression model, TRU, achieves the highest prediction quality across a three-cohort evaluation hierarchy spanning in-distribution mixed-disease FAF, zero-shot transfer to Stargardt macular dystrophy, and zero-shot cross-modality transfer to glaucoma en-face SLO. We propose, as a hypothesis for further empirical testing in other domains, that the appropriate level of generative complexity for a longitudinal medical image prediction task should match the entropy of the predictable component of the conditional posterior, with distributional alignment as a prerequisite in all regimes. Adopted as a diagnostic, this principle reframes the engineering question for longitudinal medical image prediction from which generative framework to use to whether the chosen framework's machinery has anything informative to model on the task at hand. Funding Research to Prevent Blindness Physician Scientist Award NIH R01 EY036222 (M.W.) Code and Data Availability Statement The source code for the Temporal Retinal U-Net (TRU) will be made publicly available on GitHub (link to be provided) upon acceptance of this manuscript. The longitudinal retinal imaging datasets used in this study (FAF Mixed-Disease, Stargardt External, and Glaucoma SLO) consist of de-identified clinical data from multiple institutions. Due to the presence of sensitive clinical information and institutional data-sharing agreements, the raw imaging data are not publicly available. Bibliography [1] R.W. Strauss, X. Kong, A. Ho, A. Jha, S. West, M. Ip, P.S. Bernstein, D.G. Birch, A.V. Cideciyan, M. Michaelides, J.-A. Sahel, J.S. Sunness, E.I. Traboulsi, E. Zrenner, S. Pitetta, D. Jenkins, A.H. Hariri, S. Sadda, H.P.N. Scholl, ProgStar Study Group, Progression of Stargardt Disease as Determined by Fundus Autofluorescence Over a 12-Month Period: ProgStar Report No. 11, JAMA Ophthalmol 137 (2019) 1134–1145. https://doi.org/10.1001/jamaophthalmol.2019.2885. [2] F.G. Holz, A. Bindewald-Wittich, M. Fleckenstein, J. Dreyhaupt, H.P.N. Scholl, S. Schmitz-Valckenberg, FAM-Study Group, Progression of geographic atrophy and impact of fundus autofluorescence patterns in age-related macular degeneration, Am J Ophthalmol 143 (2007) 463–472. https://doi.org/10.1016/j.ajo.2006.11.041. [3] L. Dai, B. Sheng, T. Chen, Q. Wu, R. Liu, C. Cai, L. Wu, D. Yang, H. Hamzah, Y. Liu, X. Wang, Z. Guan, S. Yu, T. Li, Z. Tang, A. Ran, H. Che, H. Chen, Y. Zheng, J. Shu, S. Huang, C. Wu, S. Lin, D. Liu, J. Li, Z. Wang, Z. Meng, J. Shen, X. Hou, C. Deng, L. Ruan, F. Lu, M. Chee, T.C. Quek, R. Srinivasan, R. Raman, X. Sun, Y.X. Wang, J. Wu, H. Jin, R. Dai, D. Shen, X. Yang, M. Guo, C. Zhang, C.Y. Cheung, G.S.W. Tan, Y.-C. Tham, C.-Y. Cheng, H. Li, T.Y. Wong, W. Jia, A deep learning system for predicting time to progression of diabetic retinopathy, Nat Med 30 (2024) 584–594. https://doi.org/10.1038/s41591-023-02702-z. [4] A. Salvi, J. Cluceru, S.S. Gao, C. Rabe, C. Schiffman, Q. Yang, A.Y. Lee, P.A. Keane, S.R. Sadda, F.G. Holz, D. Ferrara, N. Anegondi, Deep Learning to Predict the Future Growth of Geographic Atrophy from Fundus Autofluorescence, Ophthalmol Sci 5 (2025) 100635. https://doi.org/10.1016/j.xops.2024.100635. [5] J. Yim, R. Chopra, T. Spitz, J. Winkens, A. Obika, C. Kelly, H. Askham, M. Lukic, J. Huemer, K. Fasler, G. Moraes, C. Meyer, M. Wilson, J. Dixon, C. Hughes, G. Rees, P.T. Khaw, A. Karthikesalingam, D. King, D. Hassabis, M. Suleyman, T. Back, J.R. Ledsam, P.A. Keane, J. De Fauw, Predicting conversion to wet age-related macular degeneration using deep learning, Nat Med 26 (2020) 892–899. https://doi.org/10.1038/s41591-020- 0867-7. [6] B. Liefers, P. Taylor, A. Alsaedi, C. Bailey, K. Balaskas, N. Dhingra, C.A. Egan, F.G. Rodrigues, C.G. Gonzalo, T.F.C. Heeren, A. Lotery, P.L. Müller, A. Olvera-Barrios, B. Paul, R. Schwartz, D.S. Thomas, A.N. Warwick, A. Tufail, C.I. Sánchez, Quantification of Key Retinal Features in Early and Late Age-Related Macular Degeneration Using Deep Learning, Am J Ophthalmol 226 (2021) 1–12. https://doi.org/10.1016/j.ajo.2020.12.034. [7] J. Cluceru, N. Anegondi, S.S. Gao, A.Y. Lee, E.M. Lad, U. Chakravarthy, Q. Yang, V. Steffen, M. Friesenhahn, C. Rabe, D. Ferrara, Topographic Clinical Insights From Deep Learning-Based Geographic Atrophy Progression Prediction, Transl Vis Sci Technol 13 (2024) 6. https://doi.org/10.1167/tvst.13.8.6. [8] Z. Mishra, Z. Wang, E. Xu, S. Xu, I. Majid, S.R. Sadda, Z.J. Hu, Recurrent and Concurrent Prediction of Longitudinal Progression of Stargardt Atrophy and Geographic Atrophy, (2024). https://doi.org/10.1101/2024.02.11.24302670. [9] C. Liu, K. Xu, L.L. Shen, G. Huguet, Z. Wang, A. Tong, D. Bzdok, J. Stewart, J.C. Wang, L.V. Del Priore, S. Krishnaswamy, ImageFlowNet: Forecasting Multiscale Image-Level Trajectories of Disease Progression with Irregularly-Sampled Longitudinal Medical Images, in: ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, Hyderabad, India, 2025: p. 1–5. https://doi.org/10.1109/ICASSP49660.2025.10890535. [10] D. Lachinov, A. Chakravarty, C. Grechenig, U. Schmidt-Erfurth, H. Bogunovic, Learning Spatio-Temporal Model of Disease Progression With NeuralODEs From Longitudinal Volumetric Data, IEEE Trans Med Imaging 43 (2024) 1165–1179. https://doi.org/10.1109/TMI.2023.3330576. [11] M. Litrico, F. Guarnera, M.V. Giuffrida, D. Ravì, S. Battiato, TADM: Temporally-Aware Diffusion Model for Neurodegenerative Progression on Brain MRI, in: M.G. Linguraru, Q. Dou, A. Feragen, S. Giannarou, B. Glocker, K. Lekadir, J.A. Schnabel (Eds.), Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, Springer Nature Switzerland, Cham, 2024: p. 444–453. https://doi.org/10.1007/978-3-031-72069-7_42. [12] L. Puglisi, D.C. Alexander, D. Ravì, Brain Latent Progression: Individual-based spatiotemporal disease progression on 3D Brain MRIs via latent diffusion, Medical Image Analysis 106 (2025) 103734. https://doi.org/10.1016/j.media.2025.103734. [13] N.A. Disch, Y. Kirchhoff, R. Peretzke, M. Rokuss, S. Roy, C. Ulrich, D. Zimmerer, K. Maier-Hein, Temporal Flow Matching for Learning Spatio-Temporal Trajectories in 4D Longitudinal Medical Imaging, (2025). https://doi.org/10.48550/arXiv.2508.21580. [14] H. Chen, R. Yin, Y. Chen, Q. Chen, C. Li, Learning Patient-Specific Disease Dynamics with Latent Flow Matching for Longitudinal Imaging Generation, (2025). https://doi.org/10.48550/ARXIV.2512.09185. [15] J. Ho, A. Jain, P. Abbeel, Denoising Diffusion Probabilistic Models, (n.d.). [16] M. Ning, M. Li, J. Su, A.A. Salah, I.O. Ertugrul, ELUCIDATING THE EXPOSURE BIAS IN DIFFUSION MODELS, (2024). [17] G.-H. Liu, A. Vahdat, D.-A. Huang, E.A. Theodorou, W. Nie, A. Anandkumar, I2SB: Image-to-Image Schro ̈dinger Bridge, (n.d.). [18] A. Bansal, E. Borgnia, H.-M. Chu, J.S. Li, H. Kazemi, F. Huang, M. Goldblum, J. Geiping, T. Goldstein, Cold Diffusion: Inverting Arbitrary Image Transforms Without Noise, (n.d.). [19] M. Delbracio, P. Milanfar, Inversion by Direct Iteration: An Alternative to Denoising Diffusion for Image Restoration, (2024). https://doi.org/10.48550/arXiv.2303.11435. [20] O. Ronneberger, P. Fischer, T. Brox, U-Net: Convolutional Networks for Biomedical Image Segmentation, in: N. Navab, J. Hornegger, W.M. Wells, A.F. Frangi (Eds.), Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, Springer International Publishing, Cham, 2015: p. 234–241. https://doi.org/10.1007/978-3-319-24574-4_28. [21] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.N. Gomez, Ł. Kaiser, I. Polosukhin, Attention is All you Need, (n.d.). [22] B. Mildenhall, P.P. Srinivasan, M. Tancik, J.T. Barron, R. Ramamoorthi, R. Ng, NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis, (n.d.). [23] S. Elfwing, E. Uchibe, K. Doya, Sigmoid-weighted linear units for neural network function approximation in reinforcement learning, Neural Networks 107 (2018) 3–11. https://doi.org/10.1016/j.neunet.2017.12.012. [24] K. He, X. Zhang, S. Ren, J. Sun, Deep Residual Learning for Image Recognition, in: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, Las Vegas, NV, USA, 2016: p. 770–778. https://doi.org/10.1109/CVPR.2016.90. [25] T. Salimans, J. Ho, Progressive Distillation for Fast Sampling of Diffusion Models, (2022). https://doi.org/10.48550/arXiv.2202.00512. [26] J. Song, C. Meng, S. Ermon, DENOISING DIFFUSION IMPLICIT MODELS, (2021). [27] I. Loshchilov, F. Hutter, DECOUPLED WEIGHT DECAY REGULARIZATION, (2019). [28] I. Loshchilov, F. Hutter, SGDR: Stochastic Gradient Descent with Warm Restarts, (2017). https://doi.org/10.48550/arXiv.1608.03983. [29] P. Tanna, R.W. Strauss, K. Fujinami, M. Michaelides, Stargardt disease: clinical features, molecular genetics, animal models and therapeutic options, Br J Ophthalmol 101 (2017) 25–30. https://doi.org/10.1136/bjophthalmol-2016-308823. [30] Z. Wang, A.C. Bovik, H.R. Sheikh, E.P. Simoncelli, Image quality assessment: from error visibility to structural similarity, IEEE Trans Image Process 13 (2004) 600–612. https://doi.org/10.1109/tip.2003.819861. [31] W.R. Crum, O. Camara, D.L.G. Hill, Generalized overlap measures for evaluation and validation in medical image analysis, IEEE Trans Med Imaging 25 (2006) 1451–1461. https://doi.org/10.1109/TMI.2006.880587. [32] A.A. Taha, A. Hanbury, Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool, BMC Med Imaging 15 (2015) 29. https://doi.org/10.1186/s12880-015- 0068-x. [33] R. Zhang, P. Isola, A.A. Efros, E. Shechtman, O. Wang, The Unreasonable Effectiveness of Deep Features as a Perceptual Metric, in: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, IEEE, Salt Lake City, UT, 2018: p. 586–595. https://doi.org/10.1109/CVPR.2018.00068. [34] Y. Blau, T. Michaeli, The Perception-Distortion Tradeoff, in: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, IEEE, Salt Lake City, UT, USA, 2018: p. 6228–6237. https://doi.org/10.1109/CVPR.2018.00652. [35] D.P. Kingma, M. Welling, Auto-Encoding Variational Bayes, (2022). https://doi.org/10.48550/arXiv.1312.6114. [36] L. Zhang, A. Rao, M. Agrawala, Adding Conditional Control to Text-to-Image Diffusion Models, in: 2023 IEEE/CVF International Conference on Computer Vision (ICCV), IEEE, Paris, France, 2023: p. 3813–3824. https://doi.org/10.1109/ICCV51070.2023.00355. [37] S. Geman, E. Bienenstock, R. Doursat, Neural Networks and the Bias/Variance Dilemma, Neural Computation 4 (1992) 1–58. https://doi.org/10.1162/neco.1992.4.1.1. [38] J. Liu, X. Li, Q. Wei, J. Xu, D. Ding, Semi-Supervised Keypoint Detector and Descriptor for Retinal Image Matching, (2022). https://doi.org/10.48550/arXiv.2207.07932. [39] D.G. Lowe, Distinctive Image Features from Scale-Invariant Keypoints, International Journal of Computer Vision 60 (2004) 91–110. https://doi.org/10.1023/B:VISI.0000029664.99615.94. [40] M.A. Fischler, R.C. Bolles, Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography, Commun. ACM 24 (1981) 381– 395. https://doi.org/10.1145/358669.358692. [41] Spatial Decomposition of Longitudinal RNFL Maps Reveals Distinct Modes of Glaucomatous Progression with Structure–Function and Genetic Signatures, (n.d.). Supplemental Figure 1. Pixel-wise prediction error maps for the three representative Stargardt eyes shown in Figure 2. Each row corresponds to one eye. The leftmost column shows the ground-truth target image (green border); subsequent columns show the signed difference (prediction − ground truth) for each evaluated method, rendered with a diverging colormap (blue: prediction darker than ground truth; red: prediction brighter; symmetric scale ±0.3). Smaller magnitude and spatially diffuse residuals indicate more accurate predictions. TRU produces the lowest overall error, with residuals concentrated at lesion boundaries, whereas baseline methods exhibit broader, higher-magnitude deviations particularly in regions of active atrophy expansion. Supplementary Materials Distributional Alignment Supersedes Generative Complexity in Longitudinal Retinal Image Prediction S1. Supplementary Methods S1.1 U-Net Backbone Architecture The backbone is a four-level encoder–decoder with channel widths [64, 128, 256, 512], producing feature maps at spatial resolutions 256², 128², 64², 32², and 16². The encoder consists of an initial 3×3 convolution mapping the single-channel input to 64 feature channels, followed by four downsampling stages. The first two stages (scales 1 and 2) each contain two ResBlocks before max-pooling; the third and fourth stages (scales 3 and 4) each contain one ResBlock before max-pooling. A bottleneck comprising two ResBlocks at 16² resolution with 512 channels follows the final downsampling. The decoder mirrors the encoder asymmetry: the first two upsampling stages (scales 4 and 3, at lower spatial resolution) each contain one ResBlock after transposed convolution and skip concatenation, while the final two stages (scales 2 and 1, at higher spatial resolution) each contain two ResBlocks, allocating greater capacity at the resolutions where fine-grained lesion boundary details reside. Skip connections link corresponding encoder and decoder levels at all four scales. ResBlock structure. Each ResBlock follows a standard two-convolution residual structure: GroupNorm (8 groups), SiLU activation, 3×3 convolution, with a 1×1 residual projection when channel dimensions change. Each ResBlock receives conditioning from the temporal delta embedding eΔ via learned affine modulation (scale and shift), following the standard conditioning mechanism used in diffusion architectures. Output initialization. The output layer is a 1×1 convolution mapping 64 channels to a single output channel, initialized with weight standard deviation 10 −3 and zero bias. This near-zero initialization ensures that at epoch 0, the network output is approximately zero. Because TRU predicts a residual, the initial prediction approximates the copy-last baseline I N . S1.2 Adaptive Intensity Thresholding for Proxy Atrophy Metrics Proxy atrophy masks are generated by an adaptive intensity-thresholding algorithm applied independently to each image. Within a circular region of interest centered on the fovea (radius = 40% of the image dimension), the algorithm computes the mean (μ) and standard deviation (σ) of pixel intensities among valid fundus pixels (intensity > 10). The binarization threshold is set as min(μ − 1.5σ, 0.70μ), floored at 1.0. Pixels below this threshold within the ROI are classified as candidate atrophy. Morphological closing followed by opening (5×5 elliptical kernel) removes noise, and a foveal connectivity constraint retains only connected components of at least 20 pixels that overlap a central seed region (radius = 15% of image dimension). The same algorithm with identical parameters is applied to both predicted and ground-truth images, ensuring that any systematic bias in segmentation affects both sides of the comparison equally. S1.3 2D-BrLP Adaptation Details 2D-BrLP (2D Brain Latent Progression) adapts the BrLP pipeline, originally developed for 3D brain MRI synthesis, to 2D retinal FAF prediction. The method employs a three-stage pipeline: Stage 1 – Variational Autoencoder (VAE). A 2D VAE compresses 256×256 FAF images into a 32×32×4 latent space (12.96M parameters). Stage 2 – Diffusion U-Net. A diffusion U-Net (151.75M parameters) with epsilon-prediction learns the unconditional latent distribution. Stage 3 – ControlNet. A ControlNet (83.20M parameters), initialized from the U-Net encoder and trained with the U-Net frozen, provides subject-specific conditioning via concatenation of the baseline latent representation and a spatial time-delta map. Temporal conditioning uses the same log-sinusoidal time-delta embedding as TRU, producing a 256-dimensional context vector delivered via cross-attention. The total model comprises 247.91M trainable parameters, an order of magnitude larger than TRU (22.8M). Inference follows BrLP’s published protocol: 50-step DDIM sampling with Latent Average Stabilization (averaging m = 5 independent latent samples before decoding). Key adaptations from original BrLP. Conversion from 3D to 2D convolutions; replacement of the 8-dimensional clinical covariate vector (age, sex, diagnosis, regional brain volumes) with the time-delta embedding, since equivalent per-visit clinical covariates are unavailable for FAF; and omission of the auxiliary Leaspy disease progression model used in BrLP for extrapolating future brain volumes. The VAE and diffusion model were trained on the same FAF training partition as all other methods (excluding Stargardt), using matched hyperparameters (AdamW, learning rate 1×10 −4 , cosine decay, 150 diffusion epochs). S1.4 Longitudinal Image Registration of Optos FAF images All longitudinal sequences used for training and evaluation were produced by an automated registration and harmonization pipeline that aligns the visits of a single eye into a common spatial frame, normalizes intensities across acquisitions, and selects a single best-quality image per visit time stamp. Optos ultra-widefield acquisitions are produced at several native resolutions, each with a different ratio of usable retinal field to peripheral border. A fixed center-crop fraction is applied per native resolution before any further processing, removing the dark peripheral surround that the imaging device introduces and ensuring that downstream feature detection operates on physically meaningful retinal area. The crop fraction depends only on acquisition mode and not on image content. For every cropped image we estimate a binary mask of the usable retinal field by low-threshold border detection (intensity threshold 2), morphological closing with a 51 by 51 elliptic kernel, extraction of the largest connected component, and replacement of the contour with its convex hull. The convex hull step is important because it guarantees that intrinsically dark structures inside the retinal disc, including the optic disc, vessels, and atrophic lesions, are included in the mask even when their intensity approaches the background level. The resulting mask is used both as a region of interest for downstream histogram matching and, after warping into the registered frame, as a validity mask that excludes blank borders introduced by the registration rotation. This warped registration-border mask is the validity mask referenced in the main-text masked MSE loss (Section 3). Each cropped image is letterboxed to 1024 by 1024 with bilinear scaling and zero padding, with the scale and padding parameters stored so that geometric transforms estimated in this model space can be mapped analytically back to the full crop resolution. Keypoints and 256- dimensional descriptors are extracted using SuperRetina [38], a self-supervised fundus-specific keypoint network, with non-maximum suppression at a 10-pixel window and a detection score threshold of 0.10. Border keypoints within 4 pixels of the image edge are discarded. Pairwise descriptor matching uses brute-force mutual nearest neighbours with Lowe ratio test [39] at 0.85, and image pairs producing fewer than 10 good matches are dropped from further processing. For each eye with multiple visits, one visit is designated as the geometric reference (the anchor) and all other visits are warped onto its frame. The anchor is selected as the visit maximizing the product of its in-FOV keypoint count and its FOV area fraction, with a fivefold penalty applied to visits whose FOV area fraction falls below 0.35. Selecting a global anchor rather than chaining pairwise alignments avoids the drift accumulation characteristic of chained registration on long longitudinal sequences. For each non-anchor visit we fit three geometric models in order of increasing complexity, all using the same RANSAC [40] configuration (reprojection threshold 2.0 pixels, confidence 0.999). The candidates are a four-parameter similarity transform, a six-parameter affine transform, and an eight-parameter homography. Each candidate is scored by a composite metric combining the inlier ratio, the convex-hull spread of inlier matches, and the median reprojection error of inliers, penalized by an anisotropy guard on the linear part of the transform and a magnitude guard on the projective component. A more complex model is promoted over a simpler one only if its score exceeds the simpler model by at least 10 percent. The homography candidate is additionally rejected unless its linear-part condition number is at most 3 and its projective-component magnitude is below 1e-3. These guards prevent the well-known instability of full projective estimation on small overlap regions and bias the selection toward the simplest geometric model consistent with the data. A registered visit is retained only if it satisfies all of the following criteria simultaneously. The number of good ratio-test matches is at least 20, the number of RANSAC inliers is at least 10, the median inlier reprojection error is at most 1.5 pixels, the convex-hull spread of inliers covers at least 5 percent of the image area, and the final composite score (combining inlier ratio, mask overlap fraction, inlier spread, and reprojection accuracy) is at least 0.03. When several images at the same time stamp pass the gate, the one with the highest composite score is retained, ensuring exactly one image per visit per eye. Eyes contributing fewer than two surviving visits are dropped. Binary FOV masks are warped together with their images using bilinear interpolation followed by rebinarization at 0.5, which avoids the staircase artefacts of nearest-neighbour warping. After warping, an additional center crop at fraction 0.80 is applied to remove the residual blank margins introduced by the geometric warp. After all geometric processing, surviving images are intensity-normalized to a shared distribution by histogram matching. The reference cumulative distribution is constructed from a fixed three- component Gaussian mixture parameterized by target intensity percentiles (5th = 50, 50th = 128, 95th = 190), corresponding to a bulk healthy retinal component (70 percent of pixels), a dark tail component for vessels and shadows (15 percent), and a bright tail component for hyperautofluorescence (15 percent). The same reference is used for every image in the dataset, ensuring that all eyes are mapped to a single common intensity distribution. For each image, the source histogram is computed from in-mask pixels only (excluding registration-border pixels), and an intensity look-up table mapping each source level to the reference level with the closest cumulative-distribution value is applied to the whole image. Restricting the source histogram to in-mask pixels prevents the blank registration borders from biasing the matching. Despite this harmonization step, residual inter-acquisition intensity differences persist in the output and contribute to the time-invariant acquisition variability. Left eyes are horizontally flipped immediately before saving, with masks flipped consistently, so that all eyes in the dataset share the same chirality and the macula and optic disc occupy consistent positions across the cohort. S1.5 Longitudinal Image Registration of Scanning Laser Ophthalmoscopy (SLO) Images Feature extraction was performed on the green channel of images resized to 200×200 pixels using the SuperRetina network with non-maximum suppression and L2-normalized descriptors. For each longitudinal sequence, a fixed reference frame was established by selecting the candidate image across all timepoints with the highest raw keypoint count, and moving images were subsequently paired based on the density of ratio-test matches. Geometric alignment was achieved through a hierarchical estimation process as described above in S1.4. Following the geometric warp and intensity renormalization, a final quality gate discarded any registered pairs where the invalid area outside the field of view exceeded 10% of the total image region. This dataset was previously used in a spatial-decomposition analysis of glaucomatous progression[41]. S2. Supplementary Results S2.1 Sensitivity of Atrophy Dice Rankings to Segmentation Parameters To confirm that model rankings on atrophy metrics are not artifacts of the specific threshold parameters chosen for the proxy atrophy segmentation algorithm (Section S1.2), we performed a systematic sensitivity analysis sweeping three key hyperparameters: threshold sigma (σ ∈ 1.0, 1.5, 2.0), intensity cap fraction (∈ 0.60, 0.70, 0.80), and foveal seed radius fraction (∈ 0.10, 0.15, 0.20), yielding 3×3×3 = 27 parameter combinations evaluated on the full FAF hold-out set (1,001 eyes). Model rankings on Dice are highly stable (Table S1): TRU ranks first or second in 24 of 27 combinations (89%), and the relative ordering among learned methods (TRU > 2D-BrLP > ImageFlowNet > TADM) is preserved across the majority of settings. In the three combinations where TRU drops below rank 2, all at the most aggressive threshold (σ = 2.0, cap fraction = 0.60) which detects minimal atrophy, it is displaced only by copy-last, which trivially matches when near-empty masks produce uniformly high Dice scores. This confirms that the reported lesion-level conclusions are robust to the choice of segmentation parameters. Table S1. Sensitivity of atrophy Dice rankings to adaptive thresholding parameters. Mean rank and rank frequency across 27 parameter combinations on the FAF hold-out (N = 1,001 eyes). Default parameters: σ = 1.5, cap fraction = 0.70, seed radius = 0.15. Method Mean Rank Rank ≤ 2 Rank 4–5 TRU 1.63 24/27 (89%) 0/27 (0%) Copy-last 1.89 19/27 (70%) 0/27 (0%) 2D-BrLP 2.93 5/27 (19%) 3/27 (11%) ImageFlowNet 3.93 0/27 (0%) 18/27 (67%) TADM 4.63 0/27 (0%) 24/27 (89%) S2.2 Distributional Alignment Analysis on the Stargardt Cohort To verify that the distributional alignment findings from Section 5 of the main text generalize beyond the primary FAF benchmark, we replicated the alignment progression analysis on the zero-shot Stargardt cohort (288 eyes). All models were trained on the mixed-disease FAF training set with no Stargardt cases. Table S2 reports prediction performance for all five conditioning configurations plus the copy-last baseline; Table S3 reports the corresponding pairwise statistical tests. The Stargardt cohort reproduces the alignment pattern observed on the primary FAF benchmark. Inference-side correction (Std-DDIM-50step → Std-DDIM-1step) produces large, highly significant improvements on all metrics (all p < 10⁻⁴). Training-side alignment (Std-DDIM-1step → IA-Nonlinear) yields further significant gains on all primary metrics. Once alignment is achieved, the three aligned configurations (IA-Nonlinear, IA-Linear, and TRU) produce statistically indistinguishable results on the primary clinical measures (ΔSSIM, SSIM, Dice), confirming the equivalence principle across cohorts. Table S2. Distributional alignment analysis on the Stargardt cohort (N = 288 eyes). All models trained on mixed- disease FAF with no Stargardt cases. Values are mean ± SD. Bold method names indicate inference-aligned configurations. Method MAE ↓ PSNR ↑ SSIM ↑ ΔSSIM ↑ Dice ↑ HD95 ↓ Copy-last 0.058±0.021 22.53±2.92 0.568±0.113 0.294±0.129 0.702±0.255 13.73±17.92 Std-DDIM-50step 0.064±0.019 21.82±2.23 0.532±0.111 0.247±0.117 0.661±0.259 17.68±21.15 Std-DDIM-1step 0.055±0.020 23.14±2.88 0.611±0.113 0.320±0.127 0.710±0.246 13.78±18.15 IA-Nonlinear 0.052±0.016 23.65±2.45 0.616±0.108 0.350±0.138 0.712±0.257 15.21±19.53 IA-Linear 0.052±0.016 23.63±2.44 0.618±0.109 0.345±0.140 0.719±0.247 13.85±17.74 TRU 0.052±0.016 23.66±2.41 0.615±0.110 0.351±0.147 0.704±0.254 15.39±20.27 Table S3. Pairwise statistical comparisons on the Stargardt cohort (N = 288 eyes). Wilcoxon signed-rank tests, two- sided. p-values below 0.05 are highlighted. N columns report the number of eyes with valid atrophy segmentation for each comparison. Comparison MAE PSNR SSIM ΔSSIM Dice HD95 Std-50step vs Std-1step 1.44E−27 7.16E−28 5.77E−49 1.01E−16 1.82E−5 2.93E−4 IA-Nonlinear vs Std-DDIM- 1step 6.87E−4 6.48E−9 3.69E−6 6.49E−4 4.60E−3 0.054 TRU vs IA-Nonlinear 0.669 0.993 0.117 0.999 0.339 0.646 TRU vs IA-Linear 0.303 0.263 1.48E−5 0.251 0.035 0.255 The first two rows confirm that both inference-side correction and training-side alignment produce significant improvements on the Stargardt cohort (all primary metrics p < 0.02), replicating the findings from the primary FAF benchmark (Section 5.1 of the main text). The third and fourth rows confirm the equivalence principle: TRU versus IA-Nonlinear shows no significant difference on any metric (all p > 0.10), and TRU versus IA- Linear shows significance only on SSIM (p = 1.48×10 −5 ), an isolated finding on a secondary metric consistent with the pattern observed on the FAF benchmark (Section 6.2 of the main text). S2.3 Distributional Alignment and Posterior Concentration Analyses on the Heidelberg Spectralis Cohort To assess whether the distributional alignment findings (main text Section 6.1), the framework equivalence (Section 6.2), and the mechanistic account of both (Section 6.3) reproduce under simultaneous vendor and disease shift, we replicated all three analyses on the zero-shot Heidelberg Spectralis Stargardt cohort (125 eye-pairs, Section 5.3 of the main text). All models were trained exclusively on the Optos mixed-disease FAF training set, with no Spectralis images or Stargardt cases seen during training. S2.3.1 Alignment progression Table S4 reports the prediction performance of the five conditioning configurations plus the copy-last reference on the Spectralis cohort. The alignment progression observed on the Optos FAF hold-out (Section 6.1) and on the Optos Stargardt cohort (Supplementary Table S2) reproduced on Spectralis. Table S5 reports the corresponding pairwise Wilcoxon signed-rank tests. Every alignment- progression comparison on primary metrics reaches high statistical significance (all p < 0.005), consistent with the Optos results. The two-stage additive decomposition of the alignment effect (inference-side correction, then training-side alignment), established on Optos in Section 6.1, therefore reproduced on Spectralis. Table S4. Alignment progression on the Heidelberg Spectralis Stargardt cohort (N = 125 eye-pairs). Five conditioning configurations plus the copy-last reference, ordered by degree of distributional alignment. All models trained exclusively on Optos mixed-disease FAF with no Spectralis data. Values are mean ± SD. Method MAE ↓ PSNR ↑ SSIM ↑ ΔSSIM ↑ Dice ↑ HD95 ↓ copy_last 0.0761 ± 0.032 20.2 ± 3.7 0.671 ± 0.11 0.245 ± 0.11 0.659 ± 0.24 18.2 ± 15 Std-DDIM-50step 0.0850 ± 0.032 19.6 ± 2.7 0.545 ± 0.08 0.201 ± 0.09 0.623 ± 0.23 22.8 ± 19 Std-DDIM-1step 0.0701 ± 0.027 20.8 ± 2.5 0.678 ± 0.09 0.247 ± 0.11 0.653 ± 0.21 19.3 ± 15 IA-Nonlinear 0.0671 ± 0.020 21.8 ± 2.5 0.699 ± 0.10 0.288 ± 0.14 0.678 ± 0.21 18.0 ± 14 IA-Linear 0.0698 ± 0.022 21.5 ± 2.6 0.696 ± 0.10 0.269 ± 0.15 0.661 ± 0.22 19.8 ± 16 TRU 0.0670 ± 0.020 21.8 ± 2.4 0.695 ± 0.10 0.272 ± 0.15 0.669 ± 0.20 18.7 ± 14 Table S5. Pairwise statistical comparisons on the Heidelberg Spectralis Stargardt cohort (N = 125 eye-pairs). Wilcoxon signed-rank tests, two-sided. The top block reports alignment-progression comparisons (Section 6.1); the bottom block reports equivalence comparisons among aligned configurations (Section 6.2). Comparison MAE PSNR SSIM ΔSSIM Dice HD95 Std-DDIM-50step vs Std-DDIM-1step 2.41E−15 6.68E−14 8.08E−21 1.01E−8 2.26E−4 2.73E−4 IA-Nonlinear vs Std-DDIM-1step 3.67E−3 3.47E−4 2.59E−4 4.46E−3 4.76E−3 3.60E-3 TRU vs IA-Nonlinear 0.905 0.856 0.036 0.206 0.178 0.411 TRU vs IA-Linear 0.018 0.007 0.395 0.768 0.811 0.594 S2.3.2 Equivalence among aligned configurations The lower block of Table S5 reports the equivalence comparisons among the three aligned configurations on Spectralis. On the primary clinical metrics, TRU versus IA-Nonlinear is non- significant on ΔSSIM (p = 0.21), atrophy Dice (p = 0.18), and HD95 (p = 0.41); SSIM reaches borderline uncorrected significance (p = 0.036) with an absolute effect size of 0.004 SSIM units, which would not survive correction for the multiple comparisons in this table and is clinically imperceptible. TRU versus IA-Linear is non-significant on SSIM, ΔSSIM, Dice, and HD95 (all p > 0.39); isolated significance appears on the pixel-level secondary metrics MAE (p = 0.018) and PSNR (p = 0.007), with effect sizes of 0.003 MAE units and 0.33 dB respectively — the same pattern of sporadic pixel-metric significance with clinically imperceptible effect sizes observed on the Optos FAF hold-out (Section 6.2 of the main text, where IA-Linear showed PSNR p = 0.003 with an effect size of 0.10 dB). The equivalence principle therefore reproduces on Spectralis: once distributional alignment is achieved, no aligned configuration separates from any other on any primary clinical metric at any of the three evaluation cohorts. S2.3.3 Task entropy on Spectralis We recomputed the task-entropy statistics of main text Section 6.3.1 on the Spectralis cohort, treating each (I_N, I*) pair as a consecutive-visit pair and computing |δ| = |I* − I_N| across all valid fundus-mask pixels. Table S6 reports the Spectralis statistics alongside the Optos statistics for direct comparison. The task-entropy signature reproduces closely across the two platforms. Median absolute pixel change is identical to four decimal places (0.0471 on both), the 99th percentile is identical (0.2706), and the fraction of pixels changing by less than 5% between consecutive visits is 50.2% on Spectralis versus 51.2% on Optos. The correlation between inter-visit interval and changed-pixel fraction per eye is r = 0.138 on Spectralis versus r = 0.116 on Optos, both indicating that the dominant source of inter-visit pixel change is time-invariant acquisition variability rather than temporally structured disease progression. The one statistic that differs meaningfully between cohorts is copy-last SSIM (0.684 on Spectralis consecutive-visit pairs versus 0.525 on Optos consecutive-visit pairs). This difference is attributable to the narrower Spectralis field of view (approximately 30° centered on the macula, versus approximately 200° for Optos ultra-widefield), which after macula-centered cropping and resampling to 256×256 places proportionally more temporally stable macular anatomy in each frame; the elevated copy-last SSIM on Spectralis therefore reflects field-of-view geometry rather than a different task-entropy regime. The reproduction of the task-entropy signature across confocal narrow-field Spectralis FAF and non-confocal ultra-widefield Optos FAF — two platforms with different optical principles, different intensity dynamics, and processed by independent invocations of the registration and intensity-harmonization pipeline described in Section S1.4 — indicates that the low-entropy characterization of inter-visit change in slowly progressing FAF is a property of the prediction task rather than an artifact of any specific acquisition platform or pipeline configuration. Table S6. Task-entropy statistics on Heidelberg Spectralis Stargardt versus Optos FAF training pairs. All statistics computed on consecutive-visit (I_N, I*) pairs within each cohort, restricted to pixels within the fundus validity mask. The Optos column reports the statistics from main text Table 5a, recomputed over all 9,009 training pairs for direct comparison. Statistic Optos (9,009 pairs) Spectralis (125 pairs) Fraction of pixels with |δ| < 1% 11.9% 12.0% Fraction of pixels with |δ| < 5% 51.2% 50.2% Fraction of pixels with |δ| < 10% 79.6% 78.0% Median |δ| 0.0471 0.0471 95th percentile |δ| 0.1804 0.1843 99th percentile |δ| 0.2706 0.2706 Median changed-pixel fraction per eye 0.490 0.505 Copy-last SSIM 0.525 0.684 r(Δt, changed fraction) 0.116 0.138 S2.3.4 Posterior concentration on Spectralis We recomputed the posterior-concentration analysis of main text Section 6.3.2 on the Spectralis cohort, generating K = 10 independent predictions per eye for both IA-Nonlinear and IA-Linear, varying only the random seed that determines the noise realization at each model's operating point while holding all conditioning inputs fixed. Table S7 reports inter-sample agreement and the bias–variance decomposition of prediction error on Spectralis alongside the Optos numbers from main text Table 5b. The posterior-concentration pattern reproduces on Spectralis. Inter-sample SSIM is 0.99967 for IA-Nonlinear and 0.99996 for IA-Linear, corresponding to a Var/MSE ratio of 0.01% and a bias² fraction of 99.99% for IA-Nonlinear, and Var/MSE ≤ 0.01% with bias² ≥ 99.99% for IA-Linear. Because these Spectralis evaluations use models that received no Spectralis images during training, the collapse of the conditional posterior to an effective point mass is a property of the learned posterior rather than of the training distribution. Ten independent noise realizations drawn at inference on a held-out vendor still produce essentially the same output, consistent with the interpretation in main text Section 6.3 that the predictable component of the conditional posterior is narrow as a property of the prediction task. Table S7. Posterior-concentration analysis on Heidelberg Spectralis Stargardt versus Optos FAF hold-out. K = 10 independent samples per eye on each cohort, varying only the inference-time noise realization. Var/MSE ratio: fraction of prediction mean squared error attributable to inter-sample stochastic variance. Metric IA-Nonlinear (Optos) IA-Nonlinear (Spectralis) IA-Linear (Optos) IA-Linear (Spectralis) Inter-sample SSIM 0.99984 ± 0.00006 0.99967 ± 0.000483 0.99998 ± 0.00001 0.99996 ± .000049 Prediction MSE 5.5 × 10⁻³ 7.8 × 10⁻³ 5.7 × 10⁻³ 8.4 × 10⁻³ Var / MSE ratio 0.015% 0.01% 0.002% ≤ 0.01% Bias² fraction 99.99% 99.99% 100.00% ≥ 99.99% S2. Supplementary Tables Table S8: Effect of history length on prediction quality. TRU advantage: difference between TRU and the next-best method on each metric. Metric Subset TRU Next-best TRU advantage p (vs next-best) PSNR ↑ All 23.2 22.71 (IFN) +0.52 4.33 × 10 −29 ≥3 23.5 22.79 (2D-BrLP) +0.74 1.48 × 10 −15 ≥4 23.7 22.64 (IFN) +1.02 1.31 × 10 −10 SSIM ↑ All 0.601 0.577 (2D-BrLP) +0.025 3.23 × 10 −141 ≥3 0.614 0.573 (2D-BrLP) +0.042 1.76 × 10 −43 ≥4 0.601 0.552 (2D-BrLP) +0.049 2.48 × 10 −17 ΔSSIM ↑ All 0.327 0.303 (2D-BrLP) +0.024 1.83 × 10 −6 ≥3 0.347 0.302 (2D-BrLP) +0.045 3.78 × 10 −6 ≥4 0.371 0.293 (2D-BrLP) +0.078 3.18 × 10 −6 Dice ↑ All 0.671 0.657 (copy_last) +0.014 p = 0.061 ≥3 0.700 0.661 (2D-BrLP) +0.039 p = 2.77 × 10 −4 ≥4 0.711 0.675 (2D-BrLP) +0.036 p = 0.147 Table S9: Performance of inference-aligned configurations on the FAF hold-out (N = 1,001 eyes). All three configurations share the same architecture and training data, differing only in their training framework. Method ΔSSIM ↑ SSIM ↑ PSNR ↑ MAE ↓ Dice ↑ HD95 ↑ IA-Nonlinear 0.328 ± 0.14 0.602 ± 0.09 23.3 ± 2.4 0.056 ± 0.017 0.671 ± 0.24 11.0 ± 15 IA-Linear 0.321 ± 0.14 0.602 ± 0.09 23.1 ± 2.4 0.057 ± 0.017 0.676 ± 0.24 11.0 ± 15 TRU 0.327 ± 0.14 0.601 ± 0.09 23.2 ± 2.4 0.056 ± 0.017 0.671 ± 0.24 11.1 ± 15 Table S10: Pairwise statistical comparisons between TRU and stochastic aligned configurations on the FAF hold-out (N = 1,001 eyes). Wilcoxon signed-rank tests, two-sided. Comparison ΔSSIM p- value SSIM p- value PSNR p- value MAE p- value Dice p- value HD95 p- value TRU vs IA- Nonlinear 0.815 0.356 0.365 0.370 0.396 0.499 TRU vs IA-Linear 0.053 0.119 0.003 0.115 0.167 0.093