Paper deep dive

Imagine How To Change: Explicit Procedure Modeling for Change Captioning

Jiayang Sun, Zixin Guo, Min Cao, Guibo Zhu, Jorma Laaksonen

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 105

Intelligence

Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%

Last extracted: 3/13/2026, 12:14:55 AM

Summary

ProCap is a two-stage framework for change captioning that shifts from static image pair comparison to dynamic procedure modeling. It uses a procedure encoder trained on sampled keyframes from synthesized intermediate frames to capture spatio-temporal dynamics, and employs learnable procedure queries to implicitly infer these dynamics during captioning, avoiding costly frame synthesis at inference.

Entities (5)

Change Captioning · task · 100%ProCap · framework · 100%Confidence-Based Frame Sampling Module · component · 95%Procedure Encoder · model · 95%Procedure Generation Module · component · 95%

Relation Signals (3)

ProCap → performs → Change Captioning

confidence 100% · We introduce ProCap, a novel framework that reformulates change modeling from static image comparison to dynamic procedure modeling.

Procedure Generation Module → partof → ProCap

confidence 95% · ProCap features a two-stage design... Procedure Generation Module: This component synthesizes intermediate frames

ProCap → utilizes → Procedure Encoder

confidence 95% · The first stage trains a procedure encoder to learn the change procedure from a sparse set of keyframes.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Change captioning generates descriptions that explicitly describe the differences between two visually similar images. Existing methods operate on static image pairs, thus ignoring the rich temporal dynamics of the change procedure, which is the key to understand not only what has changed but also how it occurs. We introduce ProCap, a novel framework that reformulates change modeling from static image comparison to dynamic procedure modeling. ProCap features a two-stage design: The first stage trains a procedure encoder to learn the change procedure from a sparse set of keyframes. These keyframes are obtained by automatically generating intermediate frames to make the implicit procedural dynamics explicit and then sampling them to mitigate redundancy. Then the encoder learns to capture the latent dynamics of these keyframes via a caption-conditioned, masked reconstruction task. The second stage integrates this trained encoder within an encoder-decoder model for captioning. Instead of relying on explicit frames from the previous stage -- a process incurring computational overhead and sensitivity to visual noise -- we introduce learnable procedure queries to prompt the encoder for inferring the latent procedure representation, which the decoder then translates into text. The entire model is then trained end-to-end with a captioning loss, ensuring the encoder's output is both temporally coherent and captioning-aligned. Experiments on three datasets demonstrate the effectiveness of ProCap. Code and pre-trained models are available at this https URL

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

104,543 characters extracted from source content.

Expand or collapse full text

Published as a conference paper at ICLR 2026 IMAGINE HOW TO CHANGE: EXPLICIT PROCEDURE MODELING FOR CHANGE CAPTIONING Jiayang Sun ∗ 1 Zixin Guo ∗2 Min Cao †1 Guibo Zhu †3 4 5 Jorma Laaksonen 2 1 School of Computer Science and Technology, Soochow University, Jiangsu, China 2 Department of Computer Science, Aalto University, Espoo, Finland 3 Foundation Model Research Center, Institute of Automation, Chinese Academy of Sciences, Beijing, China 4 School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China 5 Wuhan AI Research jysun02@stu.suda.edu.cn zixin.guo@aalto.fi mcao@suda.edu.cn gbzhu@nlpr.ia.ac.cn ABSTRACT Change captioning generates descriptions that explicitly describe the differences between two visually similar images. Existing methods operate on static im- age pairs, thus ignoring the rich temporal dynamics of the change procedure, which is the key to understand not only what has changed but also how it oc- curs. We introduce ProCap, a novel framework that reformulates change mod- eling from static image comparison to dynamic procedure modeling. ProCap features a two-stage design: The first stage trains a procedure encoder to learn the change procedure from a sparse set of keyframes. These keyframes are obtained by automatically generating intermediate frames to make the implicit procedural dynamics explicit and then sampling them to mitigate redundancy. Then the encoder learns to capture the latent dynamics of these keyframes via a caption-conditioned, masked reconstruction task. The second stage integrates this trained encoder within an encoder-decoder model for captioning. Instead of relying on explicit frames from the previous stage—a process incurring computa- tional overhead and sensitivity to visual noise—we introduce learnable procedure queries to prompt the encoder for inferring the latent procedure representation, which the decoder then translates into text. The entire model is then trained end-to-end with a captioning loss, ensuring the encoder’s output is both tem- porally coherent and captioning-aligned. Experiments on three datasets demon- strate the effectiveness of ProCap. Code and pre-trained models are available at https://github.com/BlueberryOreo/ProCap. 1INTRODUCTION Change captioning aims to generate textual descriptions that emphasize differences between two similar images. It has attracted growing interest due to its wide applications, like monitoring tem- poral changes in remote sensing (Chouaf et al., 2021), supporting medical diagnosis by leveraging comparisons between abnormal and normal medical images (Bian et al., 2025), supporting urban planning via intelligent surveillance (Sun et al., 2024), and improving industrial quality control (Xie et al., 2024). Despite its promise, the task remains challenging due to (1) subtle appearance changes often being obscured by variations in viewpoint, illumination, or background clutter, and (2) the dif- ficulty of transforming fine-grained visual differences into coherent, accurate language descriptions. To address them, existing methods follow an encoder-decoder framework, where the encoder cap- tures visual differences and the decoder generates descriptive captions. Early works (Park et al., 2019; Shi et al., 2020) model pixel-level differences via patch features, while later works (Qiu et al., 2021; Yao et al., 2022; Tu et al., 2023c) introduce intricate difference extractors with alignment ∗ Both authors contributed equally to this research. † Corresponding authors. 1 arXiv:2603.05969v1 [cs.CV] 6 Mar 2026 Published as a conference paper at ICLR 2026 mechanisms to better localize change regions. More recently, the field has seen a shift towards in- tegrating Large Language Models (LLMs) as decoders (Yang et al., 2023; Hu et al., 2024; Zhang et al., 2024), leading to substantial gains in caption quality. Furthermore, recent advancements in applying reinforcement learning to bolster LLM reasoning (Peng et al., 2025; Wu et al., 2025b) present a promising avenue for further enhancing change captioning. Although promising, these methods typically focus on static image pairs, neglecting dynamic context and temporal cues critical for robust change perception. In practice, the transition between images often involves intermedi- ate frames that capture rich spatio-temporal dynamics, explicitly revealing appearance and motion changes only implicitly encoded in the static pair (see Figure 1). Explicitly modeling this transition process thus offers a more principled basis for change understanding and captioning. The two scenes seem identical. The tiny yellow rubber cylinder changed its location. Static Image Pair Modeling Dynamic Procedure Modeling Figure 1: Comparison between static image pair modeling and our proposed dynamic procedure modeling. Dynamic procedures offer temporal cues: the yellow cylinder, initially partly obscured by the green cube, changes its location. In this work, we make the first attempt beyond change captioning on static image pairs by formulat- ing a procedure-modeling-then-captioning paradigm. We explicitly model the dynamic change pro- cedure between static image pairs and perform captioning on the modeled spatio-temporal change procedure. We present ProCap, an innovative two-stage framework: (1) explicit procedure mod- eling, which captures latent spatio-temporal dynamics between image pairs, and (2) implicit pro- cedure captioning, which generates rich descriptions by leveraging learnable queries to implicitly reason over the modeled change procedure. Explicit procedure modeling. Our framework models the underlying change procedure via three key components. Procedure Generation Module: This component synthesizes intermediate frames to transform the implicit transformation between input images into an explicit and observable tem- poral sequence. However, the generated sequence tends to be dense and temporally redundant, often containing low-information content that incurs unnecessary computational overhead. Confidence- Based Frame Sampling Module: To address this, we introduce a confidence-aware sampling module to distill the sequence into a sparse set of informative keyframes. Each frame is assigned a con- fidence score based on temporal and semantic importance. By retaining only the highest-scoring frames, our module focuses learning on pivotal transition moments, thereby improving efficiency and representational quality during training. Procedure Modeling Module: Finally, we employ a pro- cedure encoder to learn a compact latent representation of the sampled keyframe sequence. We cast this as a caption-conditioned masked frame reconstruction task, where a multi-granularity masking strategy—ranging from local patches to entire frames—is introduced. This encourages the model to capture aligned spatio-temporal dynamics across multiple scales, while mitigating overfitting to superficial visual cues and enhancing generalization in procedural understanding. Implicit procedure captioning. The key challenge for captioning lies in leveraging the procedu- ral knowledge learned by the procedure encoder for efficient and effective text generation. A naive approach—generating and encoding intermediate frames at inference—incurs high computational cost and introduces sensitivity to synthesis noise. To address them, we present the implicit proce- dure captioning, inserting a set of learnable procedure queries between static image pairs, acting as “slots” to replace explicit intermediate frames. By leveraging the understanding of spatio-temporal dynamics learned during the first stage, the procedure encoder is prompted to infer the latent change 2 Published as a conference paper at ICLR 2026 procedure implicitly encoded within the image pair. The resulting procedural representation is de- coded into a textual description, enabling end-to-end optimization via a captioning loss. This yields a temporally coherent, task-aligned representation without costly frame synthesis at inference. Our contributions are summarized: (1) We introduce ProCap, a two-stage framework reformulat- ing change captioning from static comparison to dynamic procedure modeling, directly addressing limitation of prior works: their reliance on static image pairs, overlooking rich temporal dynam- ics. (2) We propose explicit procedure modeling, where a procedure encoder is trained on sampled keyframes from a synthesized explicit procedure, with a caption-conditioned masked reconstruc- tion task to capture change dynamics. (3) We develop implicit procedure captioning, introducing learnable queries to enable the encoder to model the procedure implicitly, bypassing the costly and noise-prone frame synthesis at inference for efficient and effective captioning. 2RELATED WORK Existing change captioning methods primarily operate on static image pairs, treating the task as a spatial comparison problem. Pioneering works by Jhamtani & Berg-Kirkpatrick (2018) and Park et al. (2019) establish a foundational encoder-decoder framework. Subsequent efforts enhance this static comparison with two main paradigms: (1) designing intricate change encoders for fine-grained localization and robustness to distractors like viewpoint or illumination changes (Kim et al., 2021; Tu et al., 2023a; Yue et al., 2023; Tu et al., 2024a; Li et al., 2025; Hu et al., 2025; Zhong et al., 2025); and (2) adopting advanced training strategies, such as auxiliary retrieval tasks (Hosseinzadeh & Wang, 2021) or multi-stage alignment (Guo et al., 2022; Yao et al., 2022; Rahmanzadehgervi et al., 2025), to guide the learning process. Beyond model design, recent vision-language studies (Menon & Vondrick, 2022; Pratt et al., 2023; Guo et al., 2023) have increasingly explored prompt-driven methodologies to harness large-scale data. Inspired by these advancements, recent studies like Liu et al. (2025) and Di et al. (2025) have focused on constructing large-scale, high-quality datasets to further advance the field. Furthermore, to mitigate hallucinations arising from large-scale datasets, Guo et al. (2025) explores a noise-robust pre-training framework for change captioning. Although promising, these methods infer changes directly from “before” and “after” images, ignoring the un- derlying continuous and dynamic transition process. In contrast, we propose to explicitly model the change procedure, shifting the paradigm from spatial comparison to spatio-temporal procedure modeling. We argue that the intermediate sequence contains rich temporal dynamics critical for ro- bust change understanding—information inherently missing in static pairs. While recent advances in video understanding have focused on improving temporal grounding in LLMs through explicit frame identifiers (Wu et al., 2025a), the application of such dynamic modeling in change caption- ing remains underexplored. The most closely related work is Zhu et al. (2025), which implicitly models temporal dynamics in remote sensing using domain-specific change maps. Our approach differs fundamentally: (1) we explicitly generate and model intermediate transitions to reason about how changes unfold, enhancing dynamic representation; and (2) we eliminate reliance on domain- specific supervision, enabling generalization to complex, unconstrained natural scenes. Additional related work, particularly on frame interpolation, is included in Appendix B. 3METHODOLOGY Relying solely on two static images, existing methods neglect the rich spatio-temporal procedure that connects an image pair. Our key insight is that such procedure is crucial for understanding not only what has changed but also how it occurs, thereby improving the change dynamics mod- eling. Given an image pair (I bef ,I aft ) containing objects O = o 1 ,o 2 ,...,o n , each object o i is represented by three continuous attributes (p i ,a i ,w i ) corresponding to position, appearance, and existence, respectively. A valid change procedure with respect to a change caption T is formalized as a mapping γ T : [0, 1] → I, where I denotes the space of all possible images, satisfying: (1) boundary conditions γ T (0) = I bef and γ T (1) = I aft ; (2) continuous evolution of each object’s at- tributes (p i ,a i ,w i ) over time t, such that γ T (t) = (p i (t),a i (t),w i (t)) n i=1 ; (3) consistency with the semantic constraints imposed by caption T ; and (4) invariance of unchanged objects throughout the process. Our objective is to derive an informative sequence P ⊂ I that approximates γ T . To address this, we introduce procedure modeling for captioning (ProCap), illustrated in Figure 2. Our proposed ProCap formulates change captioning as a two-stage learning process: (1) explicit proce- 3 Published as a conference paper at ICLR 2026 Procedure Encoder Image Feature Extractor Text Embedding Layer Masked Sequence Modeling Head 0/10/1 2Confidence-based Frame Sampling Module Confidence Score Sample Training Stage 1: Explicit Procedure ModelingTraining Stage 2: Implicit Procedure Captioning Procedure Encoder 1 Procedure Generation Module the tiny yellow rubber cylinder changed its location 3Procedure Modeling Module Multi-granularity Masking Image Tokenizer [ALI] [CSY] Text Decoder Frame Interpolation Model 12 24 6755 Insert Procedure Queries Image Feature Extractor 34 1241 6725 79 the tiny yellow rubber cylinder changed its location [ALI] [CSY] I bef I aft Freeze Module Mask Embedding Text Embedding Image Embedding Alignment Token Consistency Token Figure 2:Our two-stage ProCap framework. In the first stage, Explicit Procedure Modeling, a procedure encoder learns change dynamics from keyframes sampled from the generated explicit procedure frames. In the second stage, Implicit Procedure Captioning, learnable procedure queries, instead of explicit frames, prompt the encoder to infer an implicit representation for captioning. dure modeling stage learning to capture the latent dynamics of the change procedure, and (2) implicit procedure captioning stage learning to generate descriptions based on the modeled procedure. 3.1EXPLICIT PROCEDURE MODELING This stage incorporates three key components: a procedure generation module that produces con- tinuous frames between given static image pairs, a confidence-based frame sampling module that identifies and selects keyframes from the produced continuous frames, and a procedure modeling module that models the latent change dynamics in these keyframes. 3.1.1PROCEDURE GENERATION MODULE The first step is to make the change procedure explicit. To achieve this, we employ a pre-trained, off-the-shelf Frame Interpolation (FI) model (Lu et al., 2022) to synthesize the procedure. Given an image pair (I bef ,I aft ), the FI model first uses a CNN to predict bidirectional optical flows O t→bef , O t→aft , which are applied to the images and their features to generate warped image pairs ( ̃ I bef , ̃ I aft ) and warped feature pairs ( ̃ F bef , ̃ F aft ). These warped pairs, along with the original images, are then fed into a Transformer that produces a soft mask H and an image residual ∆I t . The intermediate frame I t is synthesized as I t = H⊙ ̃ I bef + (1− H)⊙ ̃ I aft + ∆I t , where⊙ denotes the Hadamard product. Typically, I t represents an intermediate state within the overall change pro- cedure. To construct a sequence of l pseudo-frames, we recursively apply the FI model, yielding an explicit procedure: P FI = FI(I bef ,I aft ) =I 1 ,I 2 ,...,I l .(1) However, the generated dense sequenceP FI is not optimal for direct modeling (Appendix M.2, Fig- ure 9). The primary challenge is inherent temporal redundancy. Owing to the recursive nature of the synthesis process, redundancy in a single intermediate frame—particularly when it closely resem- bles the input images—propagates to adjacent regions of the sequence, thereby providing minimal novel information about the change. Modeling this entire sequence is not only computationally in- efficient but also risks diluting the critical moments of the change with trivial, redundant frames. Therefore, distilling the sequence into a sparse set of keyframes that are relatively more informative about the change dynamics is critical for efficient and effective procedure modeling. 3.1.2CONFIDENCE-BASED FRAME SAMPLING MODULE To achieve this, we introduce a confidence-based frame sampling module. It identifies and selects the keyframes using a “score-then-sample” strategy. Specifically, each frame in P FI is assigned a confidence score quantifying its informativeness, which then guides the sampling process. 4 Published as a conference paper at ICLR 2026 Score. To quantify a frame’s informativeness, we formalize the intuition that the more critical frames are those that represent the semantic midpoint of the change—the point where a frame is semantically equidistant from the initial (I bef ) and final (I aft ) states. Conversely, frames that are highly similar to either endpoint are information-redundant. Given the image pair (I bef ,I aft ) and the generated setP FI , we compute the confidence score vector w as: w = 1− σ( [s(I bef ,P FI )− s(I aft ,P FI )] 2 ),(2) where σ(·) is the softmax function, and s(·,·) is a semantic similarity function. The squared differ- ence term ensures that frames are penalized regardless of which endpoint they are closer to. This score w assigns high values to frames that are semantically equidistant from the start and end im- ages, effectively identifying the “peak” of the change. We explore two strategies to compute the similarity s(·,·), leveraging different sources of informa- tion: (1) visual-only: based solely on visual frames, and (2) visual-text: incorporating both visual frames and textual change caption. Further details of these strategies are presented in Appendix C. Sample. Guided by the confidence score vector w, we then sample a sparse subset of k keyframes, P s = I s 1 ,I s 2 ,...,I s k ⊂ P FI . This sampled set is prepended with I bef and appended with I aft to construct the procedureP : P =I bef ,I s 1 ,I s 2 ,...,I s k ,I aft .(3) This sequence serves as the input for the procedure modeling module, which learns to encode the dynamics of the change. 3.1.3PROCEDURE MODELING MODULE The core of our method is the procedure modeling module, designed to learn a rich, unified repre- sentation of the spatio-temporal dynamics within the procedure listP . Inspired by Han et al. (2022), we employ a Transformer-based encoder as the backbone to model the change procedure, and utilize a pre-trained image tokenizer (Esser et al., 2021) to quantizeP into discrete tokens, which serve as the targets for our masked multi-frame reconstruction objective. This encourages the model to infer missing spatio-temporal information, ensuring a deep understanding of the procedural dynamics. Input representation. To prepare the encoder input, we first create a multi-modal sequence. • Visual stream: Each frame in the sequence P (length k+2) is passed through a frozen CNN backbone (Esser et al., 2021) to extract a grid of n I patch-level visual features with d-dimensional vectors, yielding a sequential embeddings e I ∈R (k+2)n I ×d . • Textual stream: Concurrently, the corresponding change caption T is tokenized into n T tokens and embedded into a sequential embeddings e T ∈R n T ×d , where each token is embedded in a d-dimensional space. • Special tokens: To structure this sequence for joint modeling, we prepend two learnable token embeddings: e csy ∈R d to e I to capture frame consistency, and e align ∈R d to e T to facilitate visual-textual alignment. The concatenated input embeddings aree align , e T , e csy , e I . Multi-granularity masking. To learn both coarse-grained semantics and fine-grained details, we introduce a multi-granularity masking strategy. This strategy is applied to the visual patch embed- dings e I , while leaving the caption fully visible. This encourages the encoder to learn the underlying spatio-temporal dynamics by reconstructing masked regions under textual guidance. The strategy comprises four distinct masking schemes. The first operates at a coarse, frame-level granularity, while the remaining three focus on fine-grained, patch-level details: • Entire masking masks the entire frame embeddings. It forces the encoder to reconstruct them using cross-modal context from the change caption. • Random patch masking masks individual patches across the frames. This strategy encourages the encoder to learn distributed visual representations. • In-block masking (Tan et al., 2021) masks a contiguous rectangular block of patches. This forces the encoder to learn the appearance and texture of local regions by “filling in” the masked area from its surrounding context. 5 Published as a conference paper at ICLR 2026 • Out-of-block masking (Tong et al., 2022) masks all patches “outside” a specific block. This encourages the encoder to learn how to represent a region while understanding its relationship to the broader surrounding scene. During each training step, every visual stream within the batch is independently masked using one of four randomly selected multi-granularity strategy. This selected strategy then generates a binary mask index set M ∈R (k+2)n I , where a value of 1 indicates a patch to be masked. The chosen patch embeddings in e I are replaced with a learnable mask embedding e m ∈R d , creating the masked visual sequence e I msk . This sequence is then fed into the procedure encoder, which outputs a sequence of contextualized hidden states H msk for subsequent optimization: H msk =h align ,h T ,h csy ,h I msk .(4) Details on the formulation of these masking operations are provided in Appendix D. 3.1.4OPTIMIZATION The objective of procedure modeling, L PRO , comprises three components: (1) masked sequence modeling for reconstructing the masked regions in each frame, L msm ; (2) cross-modal alignment between the visual frames and the textual change caption,L align ; and (3) temporal consistency within the procedure sequence,L csy . The overall training objective is formulated as: L PRO =L msm +L align +L csy .(5) Masked sequence modeling. We leverage a pre-trained image tokenizer (Esser et al., 2021) to tokenize each frame in procedure sequence P into n I discrete tokens. This tokenization process yields a corresponding discrete token sequence z ∈R (k+2)n I , serving as the ground truth. Given the modeled masked frame representations h I msk from Eq. (4), we apply a linear projection layer as the masked sequence modeling head to map each position in the representation to a vocabulary- sized logits vector, yielding the predicted token sequence y msm ∈R (k+2)n I . Given the masked frame embeddings e I msk and the change caption T , the masked sequence modeling loss is defined as: L msm =− 1 |I msk | X i∈I msk logp(y msm i = z i | e I msk , e T ),(6) where I msk =i|M i = 1 denotes the index set of positions masked in multi-granularity masking. Cross-modal alignment. We incorporate an alignment loss to bridge the visual change procedure and its corresponding textual change caption. Using the special token representation h align , which captures the relevance between visual and linguistic modalities as defined in Eq. (4), we optimize the encoder to effectively differentiate between aligned and non-aligned caption-procedure pairs: L align =− logp(1| e I msk , e T )− logp(0| e I msk , e ̄ T ),(7) where T is the change caption paired with frame embedding e I msk , and ̄ T is a negative sample not aligned with e I msk . Temporal consistency. To mitigate the impact of temporal incoherence on the modeled procedure, we incorporate a consistency loss that encourages coherent representations across frames in the sequence. Using the special token representation h csy that captures frame consistency from Eq. (4), we optimize the encoder to differentiate between consistent and non-consistent frame sequences: L csy =− logp(1| e I msk , e T )− logp(0| e ̄ I msk , e T ),(8) where ̄ I denotes the temporally warped version of I , intentionally disrupting the temporal consis- tency of the sequence. This warped version serves as a negative sample, encouraging the model to learn to distinguish between temporally coherent and incoherent sequences and generate more temporal coherent sequences. More details about constructing ̄ I are provided in Appendix E. 6 Published as a conference paper at ICLR 2026 3.2IMPLICIT PROCEDURE CAPTIONING The captioning stage employs an encoder-decoder architecture, with the encoder sharing weights with the procedure encoder from the prior stage. Directly leveraging the synthesized intermediate frames for captioning may introduce additional computational overhead as well as irrelevant noise. Therefore, we propose learnable procedure queries as dynamic “slots” inserted between start and end image features, which guide the encoder to implicitly infer change dynamics from the image pair. This design supports end-to-end training via captioning loss and yields encoder outputs that are both temporally coherent and task-relevant. Processing. First, we process the image pair (I bef ,I aft ) with a CNN backbone to extract their respective visual patch features, e I bef ∈R n I ×d and e I aft ∈R n I ×d . To bridge these two static representations, we introduce learnable procedure queries that replace the k sampled intermediate frames within the previous stage. Since each frame is represented by n I patch features, we insert k sets of queries, where each set contains n I learnable embeddings. This results in a total of k· n I queries. Each of these queries is a learnable masked embedding (e m ) used in the previous stage. The input sequence for the procedure encoder is constructed as follows: e inp =e I bef , e m ,· , e m , e I aft .(9) The procedure encoder processes e inp to produce representations that capture the underlining dy- namic change procedure. Given these encoded representations, a Transformer-based textual decoder then learns to generate the change caption. Optimization. The objective of captioning, L CAP , is an autoregressive language modeling loss that trains the entire model using ground-truth change captions as supervision: L CAP =− X i logp(T i | T <i , e inp ),(10) where T i denotes i-th word in the caption sequence. 3.3PROCAP INFERENCE FOR CAPTIONING For inference on incoming image pair, the procedure encoder takes e inp in Eq. (9) as the input. The output, a latent procedural representation, is then translated into the text caption by the textual decoder. Compared to other approaches, ProCap introduces only an additional k·n I matrix for pro- cessing. As n I remains fixed throughout our experiments, the variation in computational overhead is primarily governed by k. With k = 2, this overhead is negligible, and we further analyze its effect in the following experiments. 4EXPERIMENTS 4.1DATASETS AND METRICS Datasets. We conduct experiments on three widely-used benchmark datasets:CLEVR- Change (Park et al., 2019), Spot-the-Diff (Jhamtani & Berg-Kirkpatrick, 2018), and Image-Editing- Request (Tan et al., 2019). These datasets cover a diverse range of change domains, from synthetic changes (CLEVR-Change) to subtle differences in natural scenes (Spot-the-Diff and Image-Editing- Request), allowing for a comprehensive evaluation of our model’s capabilities. Additional details about dataset introduction are presented in Appendix G. Metrics. To evaluate the quality of the generated captions, we report four standard metrics: BLEU- 4 (B) (Papineni et al., 2002), METEOR (M) (Banerjee & Lavie, 2005), ROUGE-L (R) (Lin, 2004), and CIDEr (C) (Vedantam et al., 2015). All scores are obtained using the official Microsoft COCO evaluation toolkit (Chen et al., 2015). To evaluate the trade-off between captioning accuracy and our procedure modeling efficiency, we also measure inference efficiency in Tokens Per Second (TPS). 7 Published as a conference paper at ICLR 2026 Table 1: Comparison with SOTA on CLEVR-Change, Spot-the-Diff and Image-Editing-Request. Methods CLEVR-ChangeSpot-the-DiffImage-Editing-Request B↑M↑R↑C↑B↑M↑R↑C↑B↑M↑R↑C↑ LLM-based Methods Qwen-VL (2023)48.936.071.2119.8– LLaVA-1.5 (2023)49.735.470.8122.4– VIXEN-C (2024) –8.615.442.538.1 FINER (2024)55.636.672.5137.212.914.735.561.813.314.639.650.5 LLaVA-1.5+RP (2025)–9.713.030.843.216.219.546.760.9 Non-LLM-based Methods DUDA (2019)47.333.9–112.39.111.829.132.56.512.437.322.8 DUDA+Aux (2021)51.237.770.5115.48.112.431.338.1– IFDC (2021)49.232.569.1118.78.711.730.237.0– NCT (2023b)55.140.273.8124.1–8.115.038.834.2 VARD-Trans (2023a)55.440.173.8126.4–10.014.839.035.7 SCORER+CBR (2023c)56.341.274.5126.810.212.2–38.910.015.039.633.4 MURAT+GCM (2024)–10.213.133.139.4– SMART (2024b)56.140.874.2127.0–13.531.639.410.515.239.137.8 DIRL+CCR (2024a)–10.313.832.840.910.915.041.034.1 RDD+ACR (2025)56.141.375.0128.19.213.931.043.6– MCT-CCDiff (2025)57.540.675.6131.710.814.535.541.710.215.441.238.3 ProCap (Ours)56.741.774.7135.611.013.633.742.711.715.943.240.6 4.2PERFORMANCE COMPARISON 4.2.1BASELINES We compare ProCap against a set of state-of-the-art methods, which are grouped into two cate- gories: 1) Non-LLM-based methods are the conventional paradigm, where a pre-trained CNN extracts visual features from input image pairs. These features are then fed into a Transformer- based encoder-decoder for captioning. We compare our method with: DUDA (Park et al., 2019), DUDA+Aux (Hosseinzadeh & Wang, 2021), IFDC (Huang et al., 2021), NCT (Tu et al., 2023b), VARD-Trans (Tu et al., 2023a), SCORER+CBR (Tu et al., 2023c), MURAT+GCM (Yue et al., 2024), SMART (Tu et al., 2024b), DIRL+CCR (Tu et al., 2024a), RDD+ACR (Li et al., 2025) and MCT-CCDiff (Hu et al., 2025). 2) LLM-based methods leverage LLMs as powerful decoders, capitalizing on their vast knowledge and strong generative capabilities to improve caption quality. We compare our method with: Qwen-VL (Bai et al., 2023), LLaVA-1.5 (Liu et al., 2023), VIXEN- C (Black et al., 2024), FINER (Zhang et al., 2024) and LLaVA-1.5+RP (Jiao et al., 2025). Our ProCap falls into the non-LLM-based category. While LLM-based methods benefit from rich prior knowledge, they typically entail heavy computation and large parameter sizes. In contrast, ProCap achieves strong performance with a lightweight, efficient architecture, avoiding reliance on large external language models. 4.2.2RESULTS We analyze ProCap’s performance across three challenging scenarios, each testing a specific ca- pability, in Table 1. Additional qualitative comparisons with state-of-the-art methods, extensive visualizations and case studies, are provided in Appendix M to illustrate ProCap effectiveness. Robustness to viewpoint changes. First, we evaluate ProCap’s robustness to viewpoint shifts on the CLEVR-Change dataset. As shown in Table 1, ProCap substantially outperforms all non-LLM methods on CIDEr and achieves competitive results on other metrics, indicating stronger semantic understanding. This improvement stems from our procedure modeling, which disentangles object transformations (the “what” of change) from camera movements (distractors) by analyzing the full transition path. Compared with LLM-based methods, ProCap surpasses Qwen-VL and LLaVA-1.5, and even outperforms FINER on most metrics, demonstrating strong reasoning capability without relying on large-scale decoders. A detailed comparison across different change categories is pro- vided in Appendix I. 8 Published as a conference paper at ICLR 2026 Application to multiple changes in complex scenes. Next, we evaluate ProCap on the Spot-the- Diff dataset, a more challenging real-world benchmark with cluttered scenes and multiple subtle changes. As shown in Table 1, ProCap achieves a competitive CIDEr score of 42.7. This demon- strates a key strength of our approach: by modeling change as a stepwise procedure, ProCap can “re- play” the transformation process to disentangle concurrent changes and generate accurate captions. To better capture the rich dynamics in this setting, the frame interpolation module is pre-trained on a specialized video dataset (Oh et al., 2011) before ProCap’s main training. Generalization to open-ended scenarios. Finally, we assess ProCap’s generalization abilities on the Image-Editing-Request dataset, which is characterized by its open-ended nature with largely un- seen vocabulary. The results in Table 1 show that ProCap consistently outperforms all non-LLM baselines across all metrics. This suggests that by learning the “how” of a change (the procedure), our model develops a core understanding of the transformation itself, making it more resilient to variations in vocabulary and phrasing. While the LLM-based LLaVA-1.5+RP, with its vast knowl- edge base, still leads in overall accuracy, ProCap significantly narrows the performance gap. This demonstrates that procedural modeling is a powerful strategy for achieving robust generalization. It highlights a key distinction: whereas LLM-based methods obtain generalization by infusing external knowledge, ProCap’s ability stems directly from its architectural innovation. 4.3ABLATION STUDY We study the impact of key components within the procedure modeling stage. Additional ablations on other components are detailed in Appendices K–L. Table 2: Ablation study for explicit procedure modeling (EPM) and implicit procedure cap- tioning (IPC) on CLEVR-Change dataset. EPM IPCkB↑M↑ R↑C↑ 047.2 35.8 68.6 108.4 ✓ 052.6 38.0 70.1 112.7 ✓ 147.3 36.3 68.8 106.2 ✓156.5 41.9 75.5 128.5 Table 3: Effectiveness and performance com- parison on CLEVR-Change dataset with vary- ing procedure query set length k. MethodskTPS↑B↑M↑R↑C↑ ProCap 1766.0256.541.975.5128.5 2699.0456.741.774.7135.6 4461.2457.442.375.5128.7 7270.5556.841.875.5130.5 Impact of introducing explicit procedure modeling and implicit procedure captioning. Ta- ble 2 analyzes the introduction of the explicit procedure modeling stage. We begin with a baseline encoder-decoder model trained on static image pairs from scratch. We then compare two enhance- ments to this baseline: (1) applying a pre-training stage (explicit procedure modeling), and (2) introducing a set of learnable procedure queries to enable implicit procedure captioning (implicit procedure captioning). Finally, we extend the model with both pre-training and learnable procedure queries. Compared to the baseline initialized randomly, applying the learnable queries directly (line 3) introduces random vectors of learnable queries, therefore lacking any temporal or procedural con- text. In this case, the model cannot effectively reason about the evolution from the “before” to the “after” image. Besides, applying explicit procedure modeling without the learnable queries (line 2) demonstrates that pre-training alone provides only limited gains, far smaller than the improvement observed when both pre-training and learnable queries are used together (line 4), with the CIDEr score significantly increasing to 128.5. This remarkable gain highlights our key insight: explicitly modeling the procedural dynamics of change is far more effective than simply comparing static im- age pairs. Notably, Table 14 in Appendix L further presents the advantages of implicit procedure captioning on reducing computational overhead and exhibiting greater robustness to visual noise, once the explicit modeling stage has provided the rich temporal understanding. Impact of procedure query set length k. Table 3 shows the effect of varying the procedure query set length k on both accuracy and computational efficiency (using one NVIDIA A40 GPU). Overall, efficiency decreases as the sequence length increases due to the heavier computational load. Consid- ering the overall performance across the four evaluation metrics, C reaches its peak value of 135.6 at k = 2, while the other metrics exhibit a non-monotonic trend. Although the model achieves its best scores on B, M, and R at k = 4, the TPS drops substantially. Therefore, we select k = 2 as 9 Published as a conference paper at ICLR 2026 it offers the optimal balance between capturing sufficient procedural detail for accuracy and main- taining computational efficiency. We further compare the performance and effectiveness of different procedure query set lengths with LLM-based methods in Appendix K.4. Table 4: Ablation study for combinations of training objectives in explicit procedure modeling on CLEVR-Change and Spot-the-Diff. CLEVR-ChangeSpot-the-Diff L msm L align L csy B↑M↑R↑C↑B↑M↑R↑C↑ ✓55.140.673.9127.58.111.828.129.7 ✓55.540.673.8127.17.911.728.028.9 ✓56.140.974.5128.69.312.531.236.3 ✓56.741.774.7135.611.013.633.742.7 Integration of all the objectives. Table 4 presents the contribution of each objective function from Eq. (5) within this stage. Building on the foundation ofL msm , the full model—jointly optimized with all objectives—achieves peak performance, reaching a CIDEr score of 135.6 on the CLEVR-Change dataset and 42.7 on the Spot-the-Diff dataset. This corresponds to improvements of 8.5 on CLEVR- Change and 13.8 on Spot-the-Diff when removingL align , and gains of 7.0 and 6.4 on the two datasets when removing L csy . This further improvement highlights the integration of the other two losses, each targeting a specific aspect of the procedure representation: The alignment loss (L align ) acts as a crucial bridge, grounding the visual procedure representation in the linguistic domain. It explicitly enforces that the learned procedure is not just visually coherent, but also semantically aligned with its corresponding textual description. Meanwhile, the consistency loss (L csy ) ensures the temporal order of the procedure, specifically penalizing temporally incoherent (e.g., shuffled) sequences. This forces the model to be sensitive to the correct order of events within the change. 5CONCLUSION In this paper, we introduce ProCap, a novel two-stage paradigm that shifts change captioning from modeling static image comparison to the dynamic change procedure. The first stage learns a pro- cedure encoder that models change dynamics by performing caption-conditioned masked recon- struction on a sparse set of intermediate frames, distilled from the synthesized explicit procedure. The second stage, captioning, introduces efficient and learnable procedure queries to represent the implicit process within the image pair. This design enables end-to-end training without costly inter- mediate frame synthesis during inference. Experiments across diverse datasets demonstrate ProCap effectiveness. ACKNOWLEDGEMENTS This work is supported by the National Natural Science Foundation of China under Grants 62476188, the National Key R&D Program of China (No. 2022ZD0160601), and the Key Lab- oratory of Computing Power Network and Information Security, Ministry of Education under Grant No.2024PY024. REFERENCES Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, local- ization, text reading, and beyond, 2023. Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, p. 65–72, 2005. Minghao Bian, Kun Zhang, Dexin Zhao, and S Kevin Zhou. DiffRGennet: Difference-aware medi- cal report generation. In Medical Imaging with Deep Learning, 2025. 10 Published as a conference paper at ICLR 2026 Alexander Black, Jing Shi, Yifei Fan, Tu Bui, and John Collomosse. Vixen: Visual text comparison network for image difference captioning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, p. 846–854, 2024. Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server, 2015. Seloua Chouaf, Genc Hoxha, Youcef Smara, and Farid Melgani. Captioning changes in bi-temporal remote sensing images. In 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, p. 2891–2894. IEEE, 2021. Zonglin Di, Jing Shi, Yifei Fan, Hao Tan, Alexander Black, John Collomosse, and Yang Liu. Difftell: A high-quality dataset for describing image manipulation changes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 24580–24590, 2025. Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszko- reit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, p. 12873–12883, 2021. Tsu-Jui Fu, Licheng Yu, Ning Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what happened: Unifying text-guided video completion via multimodal masked video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 10681–10692, 2023. Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In European Conference on Computer Vision, p. 102–118. Springer, 2022. Zixin Guo, Tzu-Jui Wang, and Jorma Laaksonen. Clip4idc: Clip for image difference captioning. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computa- tional Linguistics and the 12th International Joint Conference on Natural Language Processing, p. 33–42, 2022. Zixin Guo, Tzu-Jui Julius Wang, Selen Pehlivan, Abduljalil Radman, and Jorma Laaksonen. Pitl: Cross-modal retrieval with weakly-supervised vision-language pre-training via prompting. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 2261–2265, 2023. Zixin Guo, Jiayang Sun, Tzu-Jui Julius Wang, Abduljalil Radman, Selen Pehlivan, Min Cao, and Jorma Laaksonen. Learning to describe implicit changes: Noise-robust pre-training for image difference captioning. In Findings of the Association for Computational Linguistics: EMNLP 2025, p. 10125–10145, 2025. Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dim- itris Metaxas, and Sergey Tulyakov. Show me what and tell me how: Video synthesis via mul- timodal conditioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 3615–3625, 2022. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog- nition. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 770–778, 2016. Mehrdad Hosseinzadeh and Yang Wang. Image change captioning by learning from an auxiliary task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 2725–2734, 2021. Erdong Hu, Longteng Guo, Tongtian Yue, Zijia Zhao, Shuning Xue, and Jing Liu. Onediff: A generalist model for image difference captioning. In Proceedings of the Asian Conference on Computer Vision, p. 2439–2455, 2024. 11 Published as a conference paper at ICLR 2026 Jinhong Hu, Guojin Zhong, Jin Yuan, Wenbo Pan, and Xiaoping Wang. Mct-ccdiff: Context-aware contrastive diffusion model with mediator-bridging cross-modal transformer for image change captioning. IEEE Transactions on Image Processing, 2025. Qingbao Huang, Yu Liang, Jielong Wei, Cai Yi, Hanyu Liang, Ho-fung Leung, and Qing Li. Image difference captioning with instance-level fine-grained feature representation. IEEE Transactions on Multimedia, 2021. Junhwa Hur, Charles Herrmann, Saurabh Saxena, Janne Kontkanen, Wei-Sheng Lai, Yichang Shih, Michael Rubinstein, David J Fleet, and Deqing Sun. High-resolution frame interpolation with patch-based cascaded diffusion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, p. 3868–3876, 2025. Tobias H ̈ oppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling, 2022. Harsh Jhamtani and Taylor Berg-Kirkpatrick. Learning to describe differences between pairs of sim- ilar images. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, p. 4024–4034, 2018. Qirui Jiao, Daoyuan Chen, Yilun Huang, Bolin Ding, Yaliang Li, and Ying Shen. Img-diff: Con- trastive data synthesis for multimodal large language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 9296–9307, 2025. Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 2901–2910, 2017. Hoeseong Kim, Jongseok Kim, Hyungseok Lee, Hyunsung Park, and Gunhee Kim. Agnostic change captioning with cycle consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 2095–2104, 2021. Rong Li, Liang Li, Jiehua Zhang, Qiang Zhao, Hongkui Wang, and Chenggang Yan. Region-aware difference distilling with attribute-guided contrastive regularization for change captioning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, p. 4887–4895, 2025. Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, p. 74–81, 2004. Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (eds.), Advances in Neural Information Processing Systems, volume 36, p. 34892–34916, 2023. Yuan Liu, Saihui Hou, Saijie Hou, Jiabao Du, Shibei Meng, and Yongzhen Huang. Omnidiff: A comprehensive benchmark for fine-grained image difference captioning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 21440–21449, October 2025. Liying Lu, Ruizheng Wu, Huaijia Lin, Jiangbo Lu, and Jiaya Jia. Video frame interpolation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, p. 3532–3542, 2022. Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. Sangmin Oh, Anthony Hoogs, Amitha Perera, Naresh Cuntoor, Chia-Chih Chen, Jong Taek Lee, Saurajit Mukherjee, Jake K Aggarwal, Hyungtae Lee, Larry Davis, et al. A large-scale benchmark dataset for event recognition in surveillance video. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 3153–3160. IEEE, 2011. Maxime Oquab, Timoth ́ e Darcet, Th ́ eo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 12 Published as a conference paper at ICLR 2026 Enrico Pallotta, Sina Mokhtarzadeh Azar, Shuai Li, Olga Zatsarynna, and Juergen Gall. Syncvp: Joint diffusion for synchronous multi-modal video prediction. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 13787–13797, 2025. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: A method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, p. 311–318, 2002. Dong Huk Park, Trevor Darrell, and Anna Rohrbach. Robust change captioning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 4624–4633, 2019. Yingzhe Peng, Gongrui Zhang, Miaosen Zhang, Zhiyuan You, Jie Liu, Qipeng Zhu, Kai Yang, Xingzhong Xu, Xin Geng, and Xu Yang. Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl. arXiv preprint arXiv:2503.07536, 2025. Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? gener- ating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 15691–15701, 2023. Yue Qiu, Shintaro Yamamoto, Kodai Nakashima, Ryota Suzuki, Kenji Iwata, Hirokatsu Kataoka, and Yutaka Satoh. Describing and localizing multiple changes with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 1971–1980, 2021. Pooyan Rahmanzadehgervi, Hung Huy Nguyen, Rosanne Liu, Long Mai, and Anh Totti Nguyen. Tab: Transformer attention bottlenecks enable user intervention and debugging in vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 22551–22562, 2025. Xiangxi Shi, Xu Yang, Jiuxiang Gu, Shafiq Joty, and Jianfei Cai. Finding it at another side: A viewpoint-adapted matching encoder for change captioning. In European Conference on Com- puter Vision, p. 574–590. Springer, 2020. Yanjun Sun, Yue Qiu, Mariia Khan, Fumiya Matsuzawa, and Kenji Iwata. The stvchrono dataset: Towards continuous change recognition in time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 14111–14120, 2024. Hao Tan, Franck Dernoncourt, Zhe Lin, Trung Bui, and Mohit Bansal. Expressing visual relation- ships via language. In Proceedings of the 57th Annual Meeting of the Association for Computa- tional Linguistics, p. 1873–1883, 2019. Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning, 2021. Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data- efficient learners for self-supervised video pre-training. Advances in neural information process- ing systems, 35:10078–10093, 2022. Yunbin Tu, Liang Li, Li Su, Junping Du, Ke Lu, and Qingming Huang. Adaptive representation disentanglement network for change captioning. IEEE Transactions on Image Processing, 32: 2620–2635, 2023a. Yunbin Tu, Liang Li, Li Su, Ke Lu, and Qingming Huang. Neighborhood contrastive transformer for change captioning. IEEE Transactions on Multimedia, 2023b. Yunbin Tu, Liang Li, Li Su, Zheng-Jun Zha, Chenggang Yan, and Qingming Huang. Self-supervised cross-view representation reconstruction for change captioning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 2805–2815, 2023c. Yunbin Tu, Liang Li, Li Su, Chenggang Yan, and Qingming Huang. Distractors-immune represen- tation learning with cross-modal contrastive regularization for change captioning. In European Conference on Computer Vision, p. 311–328. Springer, 2024a. 13 Published as a conference paper at ICLR 2026 Yunbin Tu, Liang Li, Li Su, Zheng-Jun Zha, and Qingming Huang. Smart: Syntax-calibrated multi- aspect relation transformer for change captioning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4926–4943, 2024b. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.), Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 4566–4575, 2015. Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. Advances in neural information processing systems, 35:23371–23385, 2022. Yongliang Wu, Xinting Hu, Yuyang Sun, Yizhou Zhou, Wenbo Zhu, Fengyun Rao, Bernt Schiele, and Xu Yang. Number it: Temporal grounding videos like flipping manga. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 13754–13765, 2025a. Yongliang Wu, Yizhou Zhou, Zhou Ziheng, Yingzhe Peng, Xinyu Ye, Xinting Hu, Wenbo Zhu, Lu Qi, Ming-Hsuan Yang, and Xu Yang. On the generalization of sft: A reinforcement learning perspective with reward rectification. arXiv preprint arXiv:2508.05629, 2025b. Jiayuan Xie, Zhiping Zhou, Zihan Wu, Xinting Zhang, Jiexin Wang, Yi Cai, and Qing Li. Automated defect report generation for enhanced industrial quality control. Proceedings of the AAAI Confer- ence on Artificial Intelligence, 38(17):19306–19314, Mar. 2024. doi: 10.1609/aaai.v38i17.29900. Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers, 2021. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al.Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025. Xu Yang, Yongliang Wu, Mingzhuo Yang, Haokun Chen, and Xin Geng. Exploring diverse in- context configurations for image captioning. Advances in Neural Information Processing Systems, 36:40924–40943, 2023. Linli Yao, Weiying Wang, and Qin Jin. Image difference captioning with pre-training and contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, p. 3108– 3116, 2022. Shengbin Yue, Yunbin Tu, Liang Li, Ying Yang, Shengxiang Gao, and Zhengtao Yu. I3n: Intra-and inter-representation interaction network for change captioning. IEEE Transactions on Multime- dia, 2023. Shengbin Yue, Yunbin Tu, Liang Li, Shengxiang Gao, and Zhengtao Yu. Multi-grained represen- tation aggregating transformer with gating cycle for change captioning. ACM Transactions on Multimedia Computing, Communications and Applications, 2024. Guozhen Zhang, Yuhan Zhu, Yutao Cui, Xiaotong Zhao, Kai Ma, and Limin Wang. Motion-aware generative frame interpolation. arXiv preprint arXiv:2501.03699, 2025a. Xian Zhang, Haokun Wen, Jianlong Wu, Pengda Qin, Hui Xue’, and Liqiang Nie. Differential- perceptive and retrieval-augmented mllm for change captioning. In Proceedings of the 32nd ACM International Conference on Multimedia, p. 4148–4157, 2024. Zihao Zhang, Haoran Chen, Haoyu Zhao, Guansong Lu, Yanwei Fu, Hang Xu, and Zuxuan Wu. Eden: Enhanced diffusion for high-quality large-motion video frame interpolation. In Proceed- ings of the Computer Vision and Pattern Recognition Conference, p. 2105–2115, 2025b. 14 Published as a conference paper at ICLR 2026 Guojin Zhong, Jinhong Hu, Jiajun Chen, Jin Yuan, and Wenbo Pan. Decider: Difference-aware contrastive diffusion model with adversarial perturbations for image change captioning. In Pro- ceedings of the AAAI Conference on Artificial Intelligence, volume 39, p. 10662–10670, 2025. Duowang Zhu, Xiaohu Huang, Haiyan Huang, Hao Zhou, and Zhenfeng Shao. Change3d: Revis- iting change detection and captioning from a video modeling perspective. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 24011–24022, 2025. 15 Published as a conference paper at ICLR 2026 AAPPENDIX OVERVIEW The appendix provides the following details: B. More related work about frame interpolation. C. Semantic Similarity Function: A detailed description of the function s(·,·) used in our Confidence-based Frame Sampling Module (see Eq. (2) in the main paper). D. Multi-granularity Masking Schemes: An overview of the four masking schemes employed in our Procedure Modeling Module. E. Warping Strategies: The description of the warping strategies to enhance temporal consis- tency in Explicit Procedure Modeling. F. Asymptotic Upper Bound: The derivation of the asymptotic upper bound for ProCap. G. Introduction of Datasets: The details of three datasets evaluated in our experiment. H. Implementation Details: The description of hyperparameters and settings used in our ex- periments. I. Comparison on Varied Change Categories: The performance comparison of different change categories on CLEVR-Change with SOTA methods. J. Extended Comparison with MCT-CCDiff: An extended analysis on Spot-the-Diff and com- parison with MCT-CCDiff on effectiveness and inference efficiency. K. Ablation on Explicit Procedure Modeling: An analysis of component contributions to our Explicit Procedure Modeling. L. Ablation on Implicit Procedure Captioning: An analysis of component contributions to our Implicit Procedure Captioning. M. Qualitative comparisons with SOTA methods and visualization of procedure modeling. N. Limitation and Future Work: The discussion of limitations in ProCap, and future work. O. The statement of ethics. P. The statement of reproducibility. Q. The statement of using LLMs in the paper. BRELATED WORK B.1FRAME INTERPOLATION Frame Interpolation (FI) aims to synthesize a dynamic visual transition between a given start and end frame. Existing approaches have achieved remarkable progress with powerful generative models, in- cluding denoising diffusion models that generate intermediate frames from noise (Voleti et al., 2022; H ̈ oppe et al., 2022; Pallotta et al., 2025; Zhang et al., 2025b; Hur et al., 2025) and Transformer-based architectures that predict missing content autoregressively (Yan et al., 2021; Ge et al., 2022). A no- table solution is text-conditioned interpolation (Han et al., 2022; Fu et al., 2023), which uses textual descriptions to guide the synthesis in a controllable manner. However, existing FI research primar- ily focuses on generating visually realistic videos, rather than supporting reasoning for downstream tasks such as captioning. To enhance change captioning, we draw inspiration from FI techniques to explicitly synthesize a procedural sequence and model the underlying change dynamics, thus providing a richer foundation for downstream reasoning. CSEMANTIC SIMILARITY FUNCTION To quantify the informativeness of the intermediate frame, we investigate two strategies for com- puting the similarity metric, s(·,·) in Eq. (2) in the main paper. These strategies are defined by the modalities they incorporate: (1) visual-only, which relies solely on visual frame information, and (2) visual-text, which integrates both visual frames and the corresponding textual change caption. The effectiveness of these strategies is experimentally presented in Appendix K.2. 16 Published as a conference paper at ICLR 2026 C.1VISUAL-ONLY Modeling fine-grained semantic similarity in images—a task that requires detailed comparison of object attributes and context—poses a challenge for conventional feature extractors. Extractors like ResNet (He et al., 2016), which are pre-trained on classification tasks, tend to produce coarse, global feature representations that overlook subtle semantic distinctions. To capture them, we employ DI- NOv2 (Oquab et al., 2023), a powerful Vision Transformer (ViT) (Dosovitskiy et al., 2021) pre- trained through self-supervision. Its attention-based architecture and training objective encourage the extraction of features that are highly sensitive to local details and object-level semantics. Conse- quently, we employ DINOv2 to extract features from each image and compute their cosine similarity, providing a robust measure of their semantic alignment. We formalize visual similarity using features from a pretrained DINOv2 model with a ViT-L/14 backbone, denoted as the encoderE DINO (·). Given a target image I t (where t ∈ bef, aft) and the generated frame setP FI , we define the visual similarity score set s vis (I t ,P FI ) as: s vis (I t ,P FI ) =s(I t ,I i )| I i ∈P FI ,(11) s(I t ,I i ) = sim[E DINO (I t ),E DINO (I i )], where sim[·,·] represents the cosine similarity between the extracted features. C.2VISUAL-TEXT While visual similarity with P FI serves to measure information redundancy, it is insufficient for verifying the semantic correctness of the change transformation. A purely visual metric is text- agnostic; thus, a pseudo-frame can be a visually plausible interpolation yet fail to represent the specific change conveyed by the ground-truth caption. To resolve this issue and enforce semantic validity, we incorporate the ground-truth change caption to explicitly model the informativeness of each pseudo-frame. To this end, we employ the pretrained CLIP-based model from Guo et al. (2022), which is specif- ically designed to measure semantic alignment between an image-pair transformation and a tex- tual description. The model provides a dedicated image-pair encoderE CLIP I (·,·), and a text encoder E CLIP T (·). The similarity function s vis-text (·,·) between a target image I t and pseudo-frame candidates P FI under caption T is defined as: s vis-text (I t ,P FI | T ) =s(I t ,I i ,T )| I i ∈P FI ,(12) s(I t ,I i ,T ) = sim[E CLIP I (I t ,I i ),E CLIP T (T )], where T is the change caption corresponding to the image pair (I bef ,I aft ). If a pseudo-frame I i is semantically misaligned with the caption T , it will receive a lower similarity score, indicating that it contains incorrect or irrelevant information about the change transformation. DMULTI-GRANULARITY MASKING SCHEMES We adopt four masking strategies as illustrated in Figure 3 during the training of Explicit Procedure Modeling: (1) entire masking, (2) random patch masking, (3) in-block masking (Tan et al., 2021) and (4) out-of-block masking (Tong et al., 2022). During training, one masking strategy is randomly selected with a probability of 0.1, 0.7, 0.1, 0.1, respectively, and applied to each sample in a batch. Given an input image embedding e I ∈R (k+2)n I ×d , the binary mask index set is denoted asM ∈ R (k+2)n I , where a value of 1 on index i indicates the i-th patch to be masked. Entire Masking. This strategy masks all embeddings in the process sequence, forcing the model to reconstruct the entire process solely based on the accompanying text sequences in the alignment setting. Formally, the masking probability is defined as: p(e I i = e I msk | e I i ∈ e I ) = 1.(13) 17 Published as a conference paper at ICLR 2026 Raw Frames Masking Patch Masking BlockKeeping Block (1) Entire Masking (2) Random Patch Masking (3) In-block Masking (4) Out-of-block Masking Figure 1: Caption Figure 3: Four masking schemes in the proposed multi-granularity strategy. We mask visual patch embeddings for reconstruction during training; the masks are visualized at the patch level for clarity. Random Patch Masking. Given an interval (a,b), the masking probability for index i is sampled from a uniform distribution over this interval, denoted asU (a,b), where a and b are set to 0.2 and 0.5 respectively in experiments. Specifically, for e I i ∈ e I , the probability of replacing e I i with a mask token e I msk is given by: p(e I i = e I msk | e I i ∈ e I ) = p i where p i ∼U (a,b).(14) In-block and Out-of-block Masking. Given an unflattened image embedding e I k ∈R h×w×d , a rectangular region, whose area ratio to the whole image is randomly sampled within [0.2, 0.8] (with an expected value of approximately 0.5), is randomly selected with bottom-left corner at (x 1 ,y 1 ) and top-right corner at (x 2 ,y 2 ): R =(i,j)| x 1 ≤ i≤ x 2 ,y 1 ≤ j ≤ y 2 ,(15) where 0 < x 1 < x 2 < w and 0 < y 1 < y 2 < h. In In-block masking, all embeddings within this region are masked: p e m,n k = e msk | e m,n k ∈ e I k , k ∈1,...,k + 2, (m,n)∈R = 1.(16) Conversely, in Out-of-block masking, all embeddings outside the selected region are masked: p e m,n k = e msk | e m,n k ∈ e I k , k ∈1,...,k + 2, (m,n) /∈ R = 1.(17) EWARPING STRATEGIES We apply four widely used warping strategies for disrupting the temporal consistency of the frame sequence for training in Explicit Procedure Modeling stage: (1) batch procedure frame shuffle, (2) frame shuffle, (3) color shifting, and (4) affine transformation. Batch Procedure Frame Shuffle. Given a batch of procedures P 1 ,P 2 ,...,P B , the sequence frame shuffle strategy randomly selects two positions i and j from two different procedureP b 1 and P b 2 , respectively. It then replaces the frame I i ∈P b 1 with I j ∈P b 2 . Frame Shuffle. Given a procedureP , a random permutation is applied to its frames to produce a shuffled sequenceP ′ , which serves as the augmented data. 18 Published as a conference paper at ICLR 2026 Color Shifting. Given a procedureP ∈R T×H×W×3 , we randomly select a single RGB channel and add a random scalar value a to all the pixels in that channel across the entire sequence. This results in a color shifting augmentation: I c shift = I c + a,(18) where I c ∈R T×H×W represents the selected RGB channel of all images inP . Affine Transformation. We apply a random affine transformation to the input image I i ∈ P . Specifically, we sample: • a rotation angle θ ∼U (−α,α), • horizontal and vertical transitions t x ,t y ∼U (−τ,τ ), • and a scaling factor s∼U (1− μ, 1 + μ). where α, τ and μ are user-defined hyperparameters, which is set to 30, 0.1 and 0.1 in out experiments respectively. An affine transformation matrix is defined as: A = s· cosθ −s· sinθ t x s· sinθ s· cosθ t y .(19) For each position of the input image [x y], the augmented output can be denoted as: x ′ y ′ = s· cosθ −s· sinθ s· sinθ s· cosθ x y + t x t y .(20) FASYMPTOTIC UPPER BOUND In this section, we will discuss the asymptotic upper bound in inference. Let n I denote the length of image embeddings and n T the length of text embeddings. For simplicity, we assume all embeddings have a uniform dimensionality d. The asymptotic upper bound of the entire model in inference can be divided into two compnents: one corresponding to procedure encoder, and the other to the text decoder. Analysis for Procedure Encoder. Before sending to the Transformer-based procedure encoder, image pairs are first encoded into embeddings via a CNN. These embeddings are then concatenated with masked embeddings to form a sequence of shape (k + 2)n I × d. For clarity in complexity analysis, we let K = k + 2 and denote K· n I as n P . The time complexity of CNN can be denoted as O(n I × channels 2 × kernels). Assuming a constant kernel size and fixed number of channels, the time complexity of a convolutional layer scales linearly with the number of output pixels, i.e., O(n I ). For each layer in the Transformer architecture, the input embeddings are linearly projected to obtain queries, keys, and values, incurring a time complexity of O(n P × d 2 ). The self-attention mechanism, as introduced by Vaswani et al. (2017), computes attention as follows: Attention(Q,K,V ) = softmax QK ⊤ √ d V.(21) This step dominates the computational cost of the attention mechanism, with a time complexity of O(n 2 P × d). As a result, the final asymptotic upper bound of procedure encoder of l e layers can be denoted as: O(n I + l e × (n P × d 2 + n 2 P × d)).(22) Given that d ≫ 1, the lower-order term O(n I ) becomes negligible, and the complexity can be approximated by: O(l e × (n P × d 2 + n 2 P × d)).(23) 19 Published as a conference paper at ICLR 2026 Analysis for Text Decoder. The text decoder for captioning is a l d -layer Transformer decoder, which includes both self-attention and cross-attention mechanisms. For the self-attention mecha- nism, the time complexity per layer is given by O(n T × d 2 + n 2 T × d). For the cross-attention mechanism, where attention is computed between the change procedure sequence and the text se- quence, the time complexity can be expressed as: O(n P × n T × d + n P × d 2 + n T × d 2 ),(24) accounting for the projections of both input sequences and the attention computation. As n P ≫ n T and d≫ N c in our experiments, Eq. (24) can be approximated by: O(n P × d + n P × d 2 + d 2 )≈ O(n P × d 2 ).(25) Therefore, the final asymptotic upper bound of a l d -layer text decoder can be denoted as: O(l d × (n T × d 2 + n 2 T × d + n P × d 2 )),(26) which can be approximated by: O(l d × n P × d 2 ).(27) Asymptotic Upper Bound in Inference. Comprising Eq. (23) and Eq. (27), the final asymptotic upper bound of the entire model in inference can be denoted as: O(l e × (n P × d 2 + n 2 P × d) + l d × n P × d 2 ).(28) Since l d is a small constant, it can be omitted from the asymptotic expression. Therefore, the asymp- totic upper bound can be denoted as: O(l e × (n P × d 2 + n 2 P × d)).(29) Substituting n P = K× n I into the above expression yields: O(l e × (K× n I × d 2 + K 2 × n 2 I × d)).(30) It can be noted that the inference computation scales quadratically with respect to procedure length K. Therefore, it is necessary to reach a balance between performance and inference computation cost. GINTRODUCTION OF DATASETS We conduct experiments on three widely used benchmark datasets: Spot-the-Diff (Jhamtani & Berg- Kirkpatrick, 2018), CLEVR-Change (Park et al., 2019), and Image-Editing-Request (Tan et al., 2019). In this section, we provide a detailed overview of each dataset. Spot-the-Diff is the first dataset specifically designed for change captioning. It is constructed by sampling from VIRAT (Oh et al., 2011), a realistic video surveillance dataset. The dataset comprises 13,192 pairs of similar images, each paired with a human-annotated change caption. Since the image pairs are derived from surveillance videos, they are well-aligned, and each pair contains at least one semantic change. The dataset is split to training, validation and testing sets with an 8:1:1 distribution. CLEVR-Change is a synthetic dataset generated using CLEVR (Johnson et al., 2017), a rendering engine capable of producing images of objects with complex relationships. It consists of 79,606 pairs of similar images with 493,735 change caption annotations, which is split into 67,660, 3,976, and 7,970 training/validation/test image pairs, respectively. Unlike Spot-the-Diff, CLEVR-Change introduces distractors alongside semantic changes—for example, variations in viewpoint that do not alter object positions. This design poses greater challenges for change captioning, requiring models to distinguish genuine semantic changes from irrelevant visual differences and to be more robust in reasoning about visual transformations. Image-Editing-Request provides similar image pairs with image editing approaches guided by instructions. It comprises 3,939 similar image pairs with 5,695 human-annotated instructions as change captions. The dataset is segmented into 3,061 training pairs, 383 validation pairs, and 495 testing pairs. 20 Published as a conference paper at ICLR 2026 Table 5: Evaluation on CLEVR-Change with varied change categories by METEOR. MethodColor Texture Add Drop Move DUDA (2019)32.827.333.431.423.5 DUDA+Aux (2021)36.130.437.836.727.0 IFDC (2021) 33.127.936.231.431.2 NCT (2023b)39.136.339.037.230.5 SMART (2024b)40.237.839.338.131.5 DIRL+CCR (2024a)40.738.240.037.933.5 ProCap (Ours)39.737.641.039.038.1 Table 6: Extended comparison with MCT-CCDiff on the Spot-the-Diff dataset, where † denotes model training with LLM-augmented captions. MethodSpeed (s/caption)↓B↑M↑R↑C↑ MCT-CCDiff (2025)0.9110.814.535.541.7 ProCap (Ours)0.0411.013.633.742.7 ProCap † (Ours) 0.0411.714.234.644.6 HIMPLEMENTATION DETAILS We employ a pre-trained frame interpolation model, VFIformer (Lu et al., 2022), to synthesize pseudo change procedures, with the process length set as l = 7. To balance captioning quality and inference efficiency, we sample k = 2 intermediate frames. For image representation, we fine-tune a pre-trained VQGAN on the change captioning datasets via an image reconstruction task. The VQGAN is configured with a codebook size of K = 1024 and a latent dimension d z = 256. Input images are resized to 224× 224 and encoded into a latent resolution of 14× 14. The procedure encoder is configured with l e = 12 layers on CLEVR-Change and Image-Editing-Request, and l e = 4 layers on Spot-the-Diff. The hidden size is fixed at 768. The caption decoder consists of l d = 2 layers on CLEVR-Change and Image-Editing-Request datasets, and consists of l d = 3 layers on Spot-the-Diff dataset, with a common hidden size of 512. In the Explicit Procedure Modeling stage, we train our model for 200,000 steps on 2 NVIDIA A40 GPUs using a warm-up strategy that linearly increases the learning rate from 1× 10 −6 to 1× 10 −4 over the first 5,000 steps. The total batch size is set to 8. In the Implicit Procedure Captioning stage, we train our model for 40 epochs with the total batch size of 16 on 1 NVIDIA A40 GPU. The procedure encoder is optimized with a fixed learning rate of 5× 10 −5 on the CLEVR-Change and Image-Editing-Request datasets, and 2× 10 −5 on the Spot-the-Diff dataset. Meanwhile, the caption decoder adopts a warm-up schedule that linearly increases the learning rate from 0 to 5 × 10 −5 during the first 10% of total training steps for all datasets. Code and data for our experiments will be made publicly available. ICOMPARISON ON VARIED CHANGE CATEGORIES In this section, we present a detailed comparison of performance across different change categories on CLEVR-Change, evaluated with METEOR against SOTA methods. Table 5 shows that our ap- proach achieves competitive results on color and texture changes, and attains the best performance on addition, removal, and movement changes. In particular, it significantly outperforms the cur- rent SOTA method on movement changes, indicating a superior ability to distinguish action-related changes in the presence of environmental distractors. JEXTENDED COMPARISON WITH MCT-CCDIFF To better understand the performance characteristics of ProCap on the Spot-the-Diff dataset, we conducted an extended analysis comparing our method with the current SOTA approach, MCT- 21 Published as a conference paper at ICLR 2026 CCDiff (Hu et al., 2025). We observed that MCT-CCDiff reports notably higher METEOR and ROUGE scores on this dataset, while ProCap achieves superior CIDEr performance. Upon exami- nation, we found that this discrepancy is primarily attributable to differences in the richness of the training captions rather than limitations of the model architecture itself. As documented in MCT-CCDiff, their training pipeline expands the original Spot-the-Diff training set with GPT-generated captions, substantially enriching the linguistic diversity of the supervision. In contrast, our primary experiments strictly follow the original, unaugmented annotations. Since METEOR and ROUGE are highly sensitive to caption diversity and surface-level phrasing, this difference in training data preparation naturally affects these metrics. To isolate the effect of caption richness, we conducted an additional experiment in which we aug- mented the Spot-the-Diff training captions using Qwen3 (Yang et al., 2025), following the strategy introduced in MCT-CCDiff. As shown in Table 6, under this matched setting, ProCap achieves com- parable METEOR and ROUGE scores and surpasses MCT-CCDiff on the more semantically aligned measures, including CIDEr and BLEU-4 (with improvements of +7% and +8%, respectively). These results indicate that the gap previously observed on METEOR and ROUGE stems largely from the linguistic properties of the training set rather than from the robustness of the model. In addition to accuracy, we also compare inference efficiency. Under identical conditions, Pro- Cap is 22× faster than MCT-CCDiff while maintaining superior CIDEr performance. This demon- strates that ProCap offers not only competitive captioning quality but also a significantly better efficiency–effectiveness trade-off compared to existing non-LLM SOTA approaches. KABLATION ON EXPLICIT PROCEDURE MODELING This section extends the ablation study from Sec. 4.3 of the main paper with a detailed compo- nent analysis on the three commonly used datasets. We specifically evaluate the contributions of individual components within the Procedure Generation, Confidence-based Frame Sampling, and Procedure Modeling Modules. K.1MORE ABLATION ON SPOT-THE-DIFF DATASET Tables 7 and 8 present additional ablation studies on the Spot-the-Diff dataset, which contains more realistic scenarios compared with CLEVR-Change. Consistent patterns emerge across these experi- ments, further demonstrating the effectiveness of our method and its strong generalization ability in real-world settings. Table 7: Ablation study for explicit procedure modeling (EPM) and implicit procedure cap- tioning (IPC) on Spot-the-Diff dataset. EPM IPCkB↑ M↑ R↑C↑ 07.9 11.7 28.0 28.9 ✓08.5 12.1 27.8 30.6 ✓18.3 12.1 27.5 29.8 ✓ 18.6 12.5 32.2 36.0 Table 8: Effectiveness and performance com- parison on Spot-the-Diff dataset with varying procedure query set length k. MethodskB↑M↑R↑C↑ ProCap 18.612.532.236.0 211.013.633.742.7 48.512.427.731.3 77.511.825.729.2 K.2PROCEDURE GENERATION MODULE We investigate the interaction between the number of generated pseudo-frames, l, and the choice of semantic similarity function for keyframe sampling. To this end, we evaluate the two functions (see Appendix C) within our Confidence-based Frame Sampling Module, benchmarking them against a random sampling baseline that selects frames uniformly. Varying number of generated pseudo-frames l. Figure 4 examines how varying the number of generated frames l affects captioning performance, while keeping the number of sampled keyframes in the Procedure Modeling Module fixed at k = 2. The results highlight a clear trade-off: increas- ing l enriches spatio-temporal cues but simultaneously introduces substantial redundancy and noise. 22 Published as a conference paper at ICLR 2026 This trade-off is most pronounced in the random sampling strategy on the CLEVER-Change dataset and in the visual-only sampling strategy on the Spot-the-Diff dataset. In both cases, performance improves as l increases from 3 to 7, but then noticeably degrades when l rises to 15. Although our proposed sampling strategy also experiences a slight decline on the Spot-the-Diff dataset as l contin- ues to grow, it consistently outperforms the other two strategies. This suggests that, without semantic guidance, redundant and irrelevant frames can easily overwhelm the model, reinforcing the need for more robust sampling mechanisms capable of isolating truly informative temporal cues while filter- ing out misleading ones. Based on these observations, we set l = 7 as the default configuration in our experiments. 371531 l 120 125 130 135 140 CIDEr Random Visual Only Visual + Text 371531 l 20 25 30 35 40 45 50 CIDEr Random Visual Only Visual + Text Figure 4: Comparison of CIDEr scores across four sampling strategies with respect to the number of pseudo-frames l on CLEVR-Change dataset (left) and Spot-the-Diff dataset (right). Each strategy is set to sample two key frames from the pseudo-frames. Measure of constraint in FI model. As defined in Sec. 3, we formalized the change procedure as a mapping γ T : [0, 1] → I, where T is a referred change caption andI denotes the space of all possible images. As the mapping is non-bijective, without additional constraints, there exist infinite procedures for the same image pair. In our experiment, to restrict the solution space, we adopt an off- the-shelf optical-flow-based frame interpolation method to synthesize change procedures, where the optical flow serves as a strong constraint: the intermediate frame is obtained by warping the before and after images according to the linearly interpolated optical flow, rather than being generated from scratch. To empirically demonstrate the necessity of these constraints, we compared our approach on the Image-Editing-Request dataset with one diffusion-based frame interpolation, which operates within a significantly less constrained solution space. As shown in Table 9, relaxing the constraints leads to noticeable performance degradation compared to the optical-flow-based approach. We at- tribute this drop to the stochastic nature of diffusion models. Unlike optical-flow-based methods that enforce strict pixel-wise correspondence, diffusion models inherently introduce unpredictable and uncontrollable visual variations in the intermediate frames (as shown in Figure 5). These unintended variations make procedure modeling more difficult, hindering effective model training. Table 9: Performance comparison with different constraints of the FI model on the Image-Editing- Request dataset. FI ModelsB↑M↑R↑C↑ Ours (diffusion-based (2025a))9.915.341.337.8 Ours (optical-flow-based (2022)) 11.715.943.240.6 K.3CONFIDENCE-BASED FRAME SAMPLING MODULE Impact of semantic similarity functions. Figure 4 illustrates the comparative performance of three distinct semantic similarity functions for keyframe selection. Our analysis yields the fol- lowing observations. (1) Random Sampling vs Visual Only Strategies: Compared with random sampling, Visual Only demonstrates benefits, particularly when sampling a larger number of pseudo- frames, such as l = 15. This highlights the effectiveness of filtering out redundant frames in long frame sequences. However, Visual Only strategy still exhibits a clear performance decline as l in- creases, indicating its sensitivity to irrelevant visual content when textual grounding is absent. (2) Visual+Text Strategy: In contrast, Visual+Text strategy consistently outperforms other strategies 23 Published as a conference paper at ICLR 2026 (a) (b) (c) (d) Figure 5: Uncontrollable predicted intermediate frames examples of diffusion-based FI models. Samples (a) and (b) show an unexpected object prediction, while samples (c) and (d) show an unex- pected motion generation. across most evaluated values of l. Its performance remains robust even as l increases, suggesting that the integration of textual cues provides a strong guiding signal for identifying informative and relevant frames. This makes Visual+Text strategy resilient to noisy or redundant frames within the temporal sequence. (3) Overall: These results collectively highlight the effectiveness of leveraging multimodal signals—particularly textual grounding—for key frame selection under varying tempo- ral resolutions. As a result, we select Visual+Text strategy for our model. 0100200300400500600700800 TPS (Tokens per Second, is better) 115 120 125 130 135 140 CIDEr ( is better) LLaVA-1.5 Qwen-VL FINER ProCap (k=1) ProCap (k=2) ProCap (k=4) ProCap (k=7) LLM-based ProCap (Ours) ProCap variants Pareto frontier Figure 6: Effectiveness and performance comparison with LLM-based methods on CLEVR-Change dataset. K.4PROCEDURE MODELING MODULE Comparison with LLM-based methods on different query set lengths k. Figure 6 presents the performance comparison with LLM-based approaches on the CLEVR-Change dataset, using the same query set configurations as in Sec. 4.3. Our method achieves clear improvements over general multi-modal large language models Qwen-VL and LLaVA-1.5, demonstrating its strong capabil- 24 Published as a conference paper at ICLR 2026 0.10.250.40.7 20 25 30 35 40 45 50 CIDEr Entire Masking Random Patch Masking In-block Masking Out-of-block Masking 0.71.0 Random Patch Masking Mask Probability Setting Figure 7: Comparison of CIDEr scores on the Spot-the-Diff dataset under different masking strate- gies across varying probability settings. When the probability of one strategy is set to p, the proba- bilities of the remaining three strategies are each set to (1− p)/3. ity in change captioning. Although the specifically trained LLM-based method FINER performs well on CLEVR-Change, it suffers from substantial computational cost due to its large number of parameters. In contrast, our approach attains competitive overall performance while maintaining remarkable effectiveness at k = 2. Impact of caption-conditioning. Table 10 presents the benefit of incorporating ground-truth cap- tions as a condition during procedure modeling. A significant performance boost is observed when the model is conditioned on the text, compared to using visual inputs alone. This shows the power of cross-modal learning in our procedure modeling. The caption acts as a powerful semantic prior, achieving two key objectives: (1) helping understand the nature of visual changes, and (2) achieving an early alignment between visual dynamics and linguistic contents. By learning to generate a pro- cedure that is consistent with the target description, the model produces a representation that is not only visually coherent but also semantically aligned with the captioning stage. Therefore, we utilize ground-truth captions as conditional guidance for training the procedure modeling module. Table 10: Ablation study for caption-conditioning in explicit procedure modeling on CLEVR- Change and Spot-the-Diff. CLEVR-ChangeSpot-the-Diff Settings B↑M↑R↑C↑B↑M↑R↑C↑ w/o caption57.0 40.9 74.7 128.88.011.6 28.1 28.9 w/ caption56.7 41.7 74.7 135.611.0 13.6 33.7 42.7 Table 11: Ablation study for multi-granularity masking strategy in explicit procedure modeling stage on Spot-the-Diff. SettingsB↑M↑R↑C↑ w/o Entire Masking8.811.930.232.5 w/o Random Patch Masking10.312.532.740.7 w/o In-block Masking7.912.028.030.0 w/o Out-of-block Masking8.012.127.630.5 w/ All Masking Strategies11.013.633.742.7 Impact of multi-granularity masking strategy.Table 11 shows the contribution of each mask- ing strategy described in Sec 3.1.3. Without the entire masking strategy, the model cannot ade- quately learn to reconstruct intermediate frames solely from change captions, thereby weakening its cross-modal understanding ability. In contrast, incorporating random patch masking yields bet- ter performance by promoting the learning of distributed visual representations. Furthermore, the 25 Published as a conference paper at ICLR 2026 significant performance drop observed when either in-block or out-of-block masking is removed highlights the crucial role of these strategies in facilitating spatial-temporal understanding. Figure 7 illustrates the performance comparison across different probability configurations of the four mask- ing strategies. Together with Table 11, the observations consistently reveal three key findings: (1) Stronger learning of distributed visual representations leads to better performance. Random patch masking plays a central role by providing broad and dense visual coverage, and therefore receives the highest probability. (2) Entire masking, in-block masking, and out-of-block masking are es- sential for modeling global context and localized structural cues. However, overemphasizing any of these structured strategies removes too many fine-grained visual details, which hampers detailed feature learning and ultimately degrades change-detection performance. This is evident from the steady performance drop observed when the probability of any of these three strategies is increased. (3) The four masking strategies work synergistically, jointly supporting both coarse-grained and fine-grained representation learning. In contrast, relying solely on random patch masking yields only marginal improvements. Impact of the procedure encoder’s depth. We investigate the impact of the procedure encoder’s depth on the CLEVR-Change and Spot-the-Diff datasets, with results presented in Tables 12 and 13. The results reveal that the optimal encoder depth is dataset-dependent. On CLEVR-Change, per- formance consistently improves with a deeper encoder, peaking with a 12-layer architecture. This suggests that modeling the changes in CLEVR-Change benefits from a higher-capacity encoder. In contrast, a shallower 4-layer encoder is optimal for Spot-the-Diff, as an overfitting is observed with deeper encoders. Table 12: Ablation results of using different procedure encoder layers on CLEVR-Change. LayersBMRC 252.6 38.9 71.8 117.4 453.9 39.9 73.1 124.3 854.2 41.0 73.9 133.2 12 56.7 41.7 74.7 135.6 Table 13: Ablation results of using different procedure encoder layers on Spot-the-Diff. LayersBMRC 27.413.0 28.4 30.2 411.0 13.6 33.7 42.7 89.412.0 32.1 42.2 12 7.413.5 27.8 30.2 LABLATION ON IMPLICIT PROCEDURE CAPTIONING We further evaluate the contributions of two key components within the implicit procedure caption- ing on the CLEVR-Change dataset and the Spot-the-Diff dataset. Explicit and implicit procedure captioning. We compare our proposed Implicit Procedure Cap- tioning (using learnable queries) against a baseline that performs Explicit Procedure Captioning (directly encoding synthesized frames). Table 14 shows that our implicit approach with procedure queries achieves superior performance on the CLEVR-Change dataset. The explicit baseline, which relies on synthesized frames, not only incurs higher computational costs but also suffers in perfor- mance. We attribute the lower accuracy of explicit procedure modeling to the redundant and noisy temporal information in the generated frames. In contrast, our learnable queries provide a more robust representation of procedural dynamics, leading to more accurate change descriptions. Table 14: Impact of implicit procedure captioning using procedure queries. The first line denotes explicit procedure captioning using synthetic pseudo-frames generated from Procedure Generation Module directly. SettingsTPSBMRC Explicit procedure captioning421.0356.5 40.8 74.4 128.5 Implicit procedure captioning 699.0456.7 41.7 74.7 135.6 26 Published as a conference paper at ICLR 2026 Impact of the text decoder’s depth. We analyze the effect of decoder depth on the CLEVR- Change and Spot-the-Diff datasets (Tables 15 and 16), observing a general trend of overfitting with excessive layers. The optimal decoder depth for Spot-the-Diff (3 layers) is greater than for CLEVR- Change (2 layers). We attribute this to the nature of the target change descriptions. Unlike the highly structured descriptions for CLEVR-Change, Spot-the-Diff requires more descriptive power. Its surveillance-style scenes feature non-canonical object poses and complex background clutter, demanding greater linguistic capacity from decoder. Table 15: Ablation results of using different text decoder layers on CLEVR-Change. LayersBMRC 256.7 41.7 74.7 135.6 356.7 41.4 74.7 129.5 456.5 40.7 74.6 129.7 556.8 41.0 74.7 130.4 Table 16: Ablation results of using different text decoder layers on Spot-the-Diff. LayersBMRC 29.412.0 32.6 37.1 311.0 13.6 33.7 42.7 47.110.7 28.1 31.7 58.111.7 27.5 28.5 MQUALITATIVE RESULTS M.1COMPARISON OF CAPTIONING GENERATIONS Figure 8 presents the qualitative results of our ProCap. We compare our model with two non-LLM- based approaches (DIRL (Tu et al., 2024a) and SCORER (Tu et al., 2023c)) and one LLM-based method (FINER (Zhang et al., 2024)) to highlight its generation capabilities. Our model demon- strates robust performance across a variety of change scenarios. Moreover, by incorporating tempo- ral information into the change captioning process, our model better captures the temporal order of events, enabling it to generate more accurate and coherent captions, as exemplified in Figure 8 (j). M.2VISUALIZATION OF CHANGE PROCEDURES Figures 9-12 present qualitative visualizations of the explicit change procedures generated by our model on three datasets: CLEVR-Change, Spot-the-Diff, and Image-Editing-Request. Our model leverages the synthetic procedures from the Procedure Generation Module and the key frames se- lected by the Confidence-based Frame Sampling Module to effectively capture the transformation process between image pairs. Notably, it remains robust even when the synthesized procedures ex- hibit temporal redundancy in the third and fourth samples, which is a critical prerequisite for the subsequent Implicit Procedure Captioning. M.3CASES WITH SIGNIFICANT VIEWPOINT SHIFT Figure 13 shows several cases exhibiting significant viewpoint shifts in the CLEVR-Change dataset. Following Park et al. (2019), we use the IoU between similar image pairs to quantify the degree of viewpoint change. The mean IoU in CLEVR-Change is 0.51 with a variance of 0.02; therefore, an IoU around 0.2 is regarded as indicating a substantial viewpoint shift (as illustrated in the first two rows). Notably, even under such drastic viewpoint differences, our model is able to reconstruct a plausible intermediate process, demonstrating the robustness of our procedure modeling module. We attribute this robustness to our proposed consistency loss, which explicitly promotes spatial- temporal consistency in the reconstructed intermediate frames. M.4FAILURE CASES Figure 14 presents several failure cases produced by our proposed ProCap. For most failures on the CLEVR-Change dataset, the modifications are extremely subtle, which makes it difficult for the model to reliably detect the change throughout the procedure. In contrast, the primary source of errors in the Image-Editing-Request and Spot-the-Diff datasets lies in inaccurate reconstruction of the intermediate procedure, which subsequently leads to incorrect change captions. We attribute this issue to overfitting, as these two datasets are more open and unconstrained compared with the 27 Published as a conference paper at ICLR 2026 (a) Color Change beforeafter Ground Truth:the tiny yellow cube turned gray ProCap (Ours):the small yellow rubber cube that is behind the big yellow matte object changed to gray DIRL:the scene remains the same SCORER:the scene remains the same FINER:there is no change (b) Texture Change before after Ground Truth:the small blue shiny cube behind the large gray rubber object be- came rubber ProCap (Ours):the small blue metal cube that is behind the large gray matte ball changed to rubber DIRL:the scene remains the same SCORER:the scene is the same as before FINER:there is no change (c) Object Add before after Ground Truth:the yellow ball has been newly placed ProCap (Ours):the small yellow metal ball that is in front of the big blue object has been newly placed DIRL:the small cyan matte ball that is in front of the big red metal thing became yellow FINER:the small cyan matte ballthat is in front of the big red cubechanged to metallic (d) Object Remove before after Ground Truth:the tiny green cylinder is no longer there ProCap (Ours):the small green matte cylinder that is behind the big gray matte cylinder is no longer there DIRL:the scene remains the same SCORER:the small yellow rubber cube that is behind the small brown rubber block became red FINER:there is no change (e) Object Move before after Ground Truth:the shiny cylinder changed its location ProCap (Ours):the small green metal cylinder that is in front of the small red matte object changed its location DIRL:the scene remains the same SCORER:the scene is the same as before FINER:there is no change (f) Complex Scenes beforeafter Ground Truth:the white vehicle driving in the lot on the left is gone ProCap (Ours):a silver carhas left the parking lot DIRL:there isa person walkingin the parking lot FINER:there isa person walkingin the parking lot (g) Complex Scenes before after Ground Truth:a black suv has appeared ProCap (Ours):a black vehicle is driving down the road in the second image DIRL:there is a person walkingin the parking lot FINER:there isa person walkingin the parking lot (h) Fine-Grained Attribute Change before after Ground Truth:Remove the text from the entire image ProCap (Ours):remove the textfrom the photo DIRL:Change the snake on the snake FINER:change the background to pink (i) Composite Change beforeafter Ground Truth:Change the wall color from white to blue ProCap (Ours):change the color of the image to a blue DIRL:change the background from a lit- tle bit FINER:zoom in on the girl (j) General Attribute Change before after Ground Truth:Darken the entire image ProCap (Ours):darken the image DIRL:Brighten the image FINER:brighten the entire image (k) General Attribute Change before after Ground Truth:crop photo ProCap (Ours):crop the image DIRL:Change the backgroundfrom that it looks like a little FINER:remove the people in the back- ground (l) Pattern Change before after Ground Truth:remove the text “louis vuitton” and the “LV” sign, add the same area that is behind and above instead of the text and the sign ProCap (Ours):remove the text DIRL:remove all background FINER:remove the background Figure 8: Comparison of captioning generations. We compare our model against two non-LLM- based approaches (DIRL and SCORER) and one LLM-based method (FINER). The examples are grouped into 10 change types, and (a)-(e) are from the CLEVR-Change dataset, (f)-(g) from Spot- the-Diff, and (h)-(l) from Image-Editing-Request. CLEVR-Change dataset. In future work, we plan to further investigate the generation and modeling of more coherent and semantically reasonable intermediate transformation processes to improve the robustness of change captioning. 28 Published as a conference paper at ICLR 2026 Caption: the big matte cube changed to purple Caption: the tiny rubber cube changed to metal Caption: the tiny brown thing has been added Caption: the small sphere is no longer there Figure 9: Visualization of change procedures on CLEVR-Change. For each sample, the top row displays the synthetic procedure generated by the Procedure Generation Module. The bottom- left shows key frames selected from this synthetic procedure using the Confidence-based Frame Sampling Module, while the bottom-right visualizes the reconstructed procedural representation produced by the Procedure Encoder within the Procedure Modeling Module. 29 Published as a conference paper at ICLR 2026 Caption: the large cylinder is in a different location Caption: there is no difference Caption: the scene is the same as before Caption: no change has occurred Figure 10: Additional visualizations of change procedures on CLEVR-Change. 30 Published as a conference paper at ICLR 2026 Caption: there’s 3 fewer boys on left photo Caption: there is one more person Caption: a car is now parked in the spot next to the building Caption: there is only one person shown Figure 11: Visualization of change procedures on Spot-the-Diff. 31 Published as a conference paper at ICLR 2026 Caption: color his hair blonde Caption: lighten the image Caption: make this a solid white background brighten up the whole image of him Caption: sharpen this image Figure 12: Visualization of change procedures on Image-Editing-Request. 32 Published as a conference paper at ICLR 2026 Caption: the blue object is in a different location Caption: the large brown block became metallic Caption: the big sphere moved Caption: the green matte cylinder changed to shiny Figure 13: Visualization of cases with significant viewpoint shift. The left shows the original image pair with the overlaid image. The right visualizes the reconstructed procedural representation pro- duced by the Procedure Encoder within the Procedure Modeling Module. Reference: the tiny brown shiny sphere right of the tiny yellow thing became yellow ProCap: no change has occurred Reference: the cube changed its location ProCap: no change has occurred Reference: Adjust the brightness ProCap: change the background to a Reference: a red car has left the scene ProCap: there is a black car in the middle of the lot Figure 14: Visualization of failure cases generated by ProCap. The left shows key frames selected from the synthetic procedure using the Confidence-based Frame Sampling Module, while the right visualizes the reconstructed procedural representation produced by the Procedure Encoder within the Procedure Modeling Module. 33 Published as a conference paper at ICLR 2026 NLIMITATION AND FUTURE WORK In this work, we propose a novel two-stage framework, ProCap, which reformulates change cap- tioning from static comparison to dynamic procedure modeling. While experiments demonstrate that our method achieves strong performance across three widely-used benchmark datasets, certain challenges remain in specific scenarios. For instance, when scenes exhibit dramatic changes, for example, where transformations exceed the variations in position, appearance, and existence defined in Sec. 3, or drastic viewpoint changes hap- pen, generating perfectly physically grounded intermediate frames becomes inherently ill-posed for any current generative model, as pixel-level correspondence is no longer preserved. In such cases, 2D generative models, such as optical-flow-based approaches (Lu et al., 2022), face fundamental limitations due to the lack of explicit geometric depth reasoning. We believe that a paradigm shift toward 3D scene modeling to maintain geometric consistency is beneficial to maintain geometric consistency and produce physically grounded intermediate frames under such extreme variations. Consequently, we identify 3D-aware representation as a critical direction to extreme geometric dis- continuities in future exploration. Another open problem lies in defining what constitutes a theoretically optimal informative point. While our current formulation provides a practical solution, a more rigorous theoretical definition remains unexplored. Future work could investigate a principled mathematical characterization of this optimal point within the broader context of change analysis, potentially leading to more robust and generalizable criteria. Finally, integrating LLMs represents a natural and valuable extension of our framework. We plan to explore LLM-based architectures—such as instruction-tuning strategies—to combine the high- level reasoning capability of LLMs with the explicit dynamic modeling strengths of ProCap. Such integration may enable richer semantic guidance and more adaptive dynamic understanding in future systems. Collectively, we believe these limitations highlight several promising avenues for continued devel- opment. With more refined model design and deeper theoretical grounding, ProCap can be extended to address these challenges more effectively. OETHICS STATEMENT This work adheres to the ICLR Code of Ethics. No human subjects or animal experiments were involved in this study. All datasets used, including CLEVR-Change, Spot-the-Diff, and Image- Editing-Request, were obtained in accordance with their respective usage guidelines, ensuring full compliance with privacy standards. We have taken care to minimize potential biases and avoid discriminatory outcomes throughout the research process. No personally identifiable information was utilized, and no experiments were conducted that could raise privacy or security concerns. We are committed to upholding transparency, fairness, and integrity in all aspects of this research. PREPRODUCIBILITY STATEMENT We have taken extensive measures to ensure the reproducibility of our results. All code and data used in the experiments will be released publicly to facilitate replication and independent verifica- tion. The experimental setup—including training procedures, model configurations, and hardware specifications—is detailed in Appendix H. In addition, we provide a comprehensive description of ProCap to further support reproducibility. Furthermore, the three change captioning datasets used in our work—CLEVR-Change, Spot-the- Diff, and Image-Editing-Request—are publicly available, ensuring consistent and reproducible eval- uation. We believe these efforts will enable other researchers to faithfully reproduce our findings and con- tribute to advancing the field. 34 Published as a conference paper at ICLR 2026 QSTATEMENT OF USING LLMS IN THE PAPER Large Language Models (LLMs) were employed to assist in writing and refining this manuscript, specifically for grammar checking and sentence polishing, with the aim of enhancing overall read- ability. Importantly, the LLM was not involved in the ideation, research methodology, experimental design, or data analysis. All research concepts, ideas, and analyses were independently developed and carried out by the authors. The role of the LLM was strictly limited to improving the linguistic quality of the text, without contributing to the scientific content. The authors take full responsibility for the manuscript, including any portions refined with LLM assistance. We have ensured that the use of LLMs complies with ethical standards and does not involve plagiarism or scientific misconduct. 35

Instant research discovery