Paper deep dive
ArtiFixer: Enhancing and Extending 3D Reconstruction with Auto-Regressive Diffusion Models
Riccardo de Lutio, Tobias Fischer, Yen-Yu Chang, Yuxuan Zhang, Jay Zhangjie Wu, Xuanchi Ren, Tianchang Shen, Katarina Tothova, Zan Gojcic, Haithem Turki
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 7/20/2026, 5:31:33 AM
Summary
The paper introduces ArtiFixer, a two-stage pipeline for enhancing 3D reconstruction (specifically 3D Gaussian Splatting) using auto-regressive diffusion models. It addresses scalability and quality issues in existing methods by first training a bidirectional generative model with an opacity mixing strategy to maintain consistency with observed regions while allowing extrapolation in unobserved areas. This model is then distilled into a causal auto-regressive generator capable of producing hundreds of frames in a single pass, outperforming baselines by 1-3 dB PSNR.
Entities (7)
Relation Signals (5)
ArtiFixer → outperforms → prior state-of-the-art methods
confidence 92% · we outperform all existing baselines by a wide margin, exceeding prior state-of-the-art methods by 1-3 dB PSNR.
ArtiFixer → uses → 3D Gaussian Splatting
confidence 90% · Our solution is a controllable auto-regressive video model that can either directly render arbitrary long novel-view renderings or provide pseudo-supervision to improve the underlying 3D reconstruction.
Opacity Mixing Strategy → enables → consistency with existing observations
confidence 88% · opacity mixing strategy that encourages consistency with existing observations while retaining the model's ability to extrapolate novel content
ArtiFixer → buildsupon → Wan 2.1
confidence 85% · We start from a pretrained text-to-video model (Wan 2.1 T2V-14B)
ArtiFixer → evaluatedon → DL3DV-10K
confidence 85% · generate paired reconstruction-ground truth samples from DL3DV-10K
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Per-scene optimization methods such as 3D Gaussian Splatting provide state-of-the-art novel view synthesis quality but extrapolate poorly to under-observed areas. Methods that leverage generative priors to correct artifacts in these areas hold promise but currently suffer from two shortcomings. The first is scalability, as existing methods use image diffusion models or bidirectional video models that are limited in the number of views they can generate in a single pass (and thus require a costly iterative distillation process for consistency). The second is quality itself, as generators used in prior work tend to produce outputs that are inconsistent with existing scene content and fail entirely in completely unobserved regions. To solve these, we propose a two-stage pipeline that leverages two key insights. First, we train a powerful bidirectional generative model with a novel opacity mixing strategy that encourages consistency with existing observations while retaining the model's ability to extrapolate novel content in unseen areas. Second, we distill it into a causal auto-regressive model that generates hundreds of frames in a single pass. This model can directly produce novel views or serve as pseudo-supervision to improve the underlying 3D representation in a simple and highly efficient manner. We evaluate our method extensively and demonstrate that it can generate plausible reconstructions in scenarios where existing approaches fail completely. When measured on commonly benchmarked datasets, we outperform all existing baselines by a wide margin, exceeding prior state-of-the-art methods by 1-3 dB PSNR.
Tags
Links
- Source: https://arxiv.org/abs/2603.00492v2
- Canonical: https://arxiv.org/abs/2603.00492v2
Trouble viewing inline? Open PDF directly →
Full Text
75,476 characters extracted from source content.
Expand or collapse full text
ArtiFixer: Enhancing and Extending 3D Reconstruction with Auto-Regressive Diffusion Models Riccardo de Lutio 0000-0002-2644-3876 NVIDIASanta ClaraUSA rdelutio@nvidia.com , Tobias Fischer 0000-0001-8227-001X NVIDIAZurichSwitzerland ETHZZurichSwitzerland tobiasf@nvidia.com , Yen-Yu Chang 0009-0006-4105-1770 NVIDIASanta ClaraUSA Cornell UniversityIthacaUSA yc2463@cornell.edu , Yuxuan Zhang 0000-0002-6409-5550 NVIDIANew YorkUSA alezhang@nvidia.com , Jay Zhangjie Wu 0009-0003-3684-7262 NVIDIATorontoCanada wjay@nvidia.com , Xuanchi Ren 0000-0001-6376-7100 NVIDIATorontoCanada University of TorontoTorontoCanada Vector InstituteTorontoCanada xuanchir@nvidia.com , Tianchang Shen 0000-0002-7133-2761 NVIDIATorontoCanada University of TorontoTorontoCanada Vector InstituteTorontoCanada frshen@nvidia.com , Katarína Tóthová 0000-0001-5864-179X NVIDIAZurichSwitzerland ktothova@nvidia.com , Zan Gojcic 0000-0001-6392-2158 NVIDIAZurichSwitzerland zgojcic@nvidia.com and Haithem Turki 0000-0001-5634-0918 NVIDIASeattleUSA hturki@nvidia.com Abstract. Per-scene optimization methods such as 3D Gaussian Splatting provide state-of-the-art novel view synthesis quality but extrapolate poorly to under-observed areas. Methods that leverage generative priors to correct artifacts in these areas hold promise but currently suffer from two shortcomings. The first is scalability, as existing methods use image diffusion models or bidirectional video models that are limited in the number of views they can generate in a single pass (and thus require a costly iterative distillation process for consistency). The second is quality itself, as generators used in prior work tend to produce outputs that are inconsistent with existing scene content and fail entirely in completely unobserved regions. To solve these, we propose a two-stage pipeline that leverages two key insights. First, we train a powerful bidirectional generative model with a novel opacity mixing strategy that encourages consistency with existing observations while retaining the model’s ability to extrapolate novel content in unseen areas. Second, we distill it into a causal auto-regressive model that generates hundreds of frames in a single pass. This model can directly produce novel views or serve as pseudo-supervision to improve the underlying 3D representation in a simple and highly efficient manner. We evaluate our method extensively and demonstrate that it can generate plausible reconstructions in scenarios where existing approaches fail completely. When measured on commonly benchmarked datasets, we outperform all existing baselines by a wide margin, exceeding prior state-of-the-art methods by 1-3 dB PSNR. †submissionid: 310†journal: TOG†ccs: Computing methodologies Computer vision†ccs: Computing methodologies Rendering Figure 1. ArtiFixer enhances and extends existing 3D reconstructions in a highly efficient and scalable manner. Given an initial reconstruction and optional reference views and text prompt, it auto-regressively generates novel content that maintains a high degree of consistency with existing observations. ArtiFixer can directly produce hundreds of novel views in a single inference pass or serve as pseudo-supervision to improve the underlying 3D reconstruction. Project page: https://research.nvidia.com/labs/sil/projects/artifixer 1. Introduction High-quality novel view synthesis is essential for applications in virtual and augmented reality and closed-loop simulation for physical AI. These use cases require photorealistic rendering and the ability to navigate complex environments under unconstrained camera motion. In recent years, two paradigms have emerged as dominant approaches to novel view synthesis: explicit 3D neural reconstruction (Mildenhall et al., 2020; Kerbl et al., 2023), and camera-controlled image or video generation (Ren et al., 2025; Zhou et al., 2025). Neural reconstruction methods have matured significantly and now enable real-time rendering and high visual fidelity when trained from dense image collections with accurate camera poses. However, in the most widely used per-scene optimization setting, their performance remains fundamentally limited by the completeness and quality of the input observations. Regions that are sparsely observed or entirely missing during capture are poorly reconstructed, leading to artifacts, holes, or implausible geometry. While such deficiencies remain hidden near the training views, they are inevitably exposed during free navigation of the scene. Conversely, recent video generative models have demonstrated the ability to synthesize photorealistic and temporally coherent content that is often indistinguishable from real-world videos (Google DeepMind, 2024; OpenAI, 2024; NVIDIA et al., 2025). Despite this progress, precise camera control over extended sequences, long-term temporal consistency, and the accumulation of drift and hallucinations remain open challenges, limiting their applicability to interactive view synthesis. Instead of treating reconstruction and generation as standalone alternatives, we aim to combine their complementary strengths: generative models serve as powerful priors to repair and complete imperfect reconstructions, while the explicit—albeit noisy and partial—3D representation provides a strong conditioning signal that grounds generation, mitigates long-term drift, and suppresses hallucinations. Recent methods have taken initial steps in this direction by training generative models to map degraded novel-view renderings to clean images and distilling the resulting improvements back into an underlying 3D representation (Gao* et al., 2024; Yu et al., 2024; Wu et al., 2025a; Fischer et al., 2025). However, these approaches must navigate two fundamental trade-offs. First, they must balance temporal consistency and efficiency: some employ large bidirectional video generative models that provide strong temporal coherence but incur high computational cost (Gao* et al., 2024; Wu et al., 2025c; Fischer et al., 2025), while others rely on (multi-view) image-based generative models that are more efficient but limit temporal consistency and require progressive distillation strategies (Wu et al., 2024, 2025a). Second, they face the trade-off between conditioning strength and generative capacity. Approaches (Yu et al., 2024; Wu et al., 2025c) that condition generation on corrupted renderings via concatenation or cross-attention risk altering the observed scene content, whereas methods (Wu et al., 2025a; Fischer et al., 2025) trained to directly map corrupted renderings to clean images are incapable of synthesizing missing content, due to the mode collapse in fully unobserved regions where all input pixels are black. In our work, we follow this line of research by adapting a pretrained bidirectional video diffusion model into a camera-controllable generator that maps corrupted renderings to clean images. To overcome the aforementioned limitations, we introduce two key contributions: (i) an opacity-aware noise mixing strategy that injects Gaussian noise into low-opacity regions, preventing mode collapse and preserving generative capacity in unobserved areas; and (i) distillation of the bidirectional model into a few-step causal auto-regressive generator capable of producing arbitrarily long, temporally consistent videos while approaching the efficiency of prior image-based methods. In doing so, we demonstrate that even highly degraded 3D reconstructions provide sufficient conditioning signals to significantly simplify the distillation process. While recent work has begun incorporating explicit 3D representations as conditioning signals for auto-regressive video generation (Zhai et al., 2025; Wu et al., 2025d; Chen et al., 2025b), these approaches treat the 3D input as a fixed conditioning rather than an output to be improved. Our method closes this loop: the reconstruction conditions the generator, and the generator in turn enhances and extends the reconstruction, enabling both higher-quality video synthesis and improved 3D scene completeness. The resulting framework enables efficient improvement of the underlying 3D reconstruction and greatly outperforms a wide range of baselines across multiple benchmarks. 2. Related Work Novel view synthesis from 3D representations. Neural Radiance Fields (NeRFs) (Mildenhall et al., 2020) and, more recently, 3D Gaussian Splatting (3DGS) (Kerbl et al., 2023) have revolutionized the field of novel view synthesis by distilling sensor information (usually overlapping photos of a scene) into a 3D representation that can then be queried from arbitrary camera viewpoints. Because these representations are optimized on a per-scene basis, their ability to extrapolate beyond observed views is inherently limited, and they fail to render plausible content in sparsely observed or missing regions. A large body of work seeks to mitigate these limitations through handcrafted geometric priors (Niemeyer et al., 2022; Yang et al., 2023; Somraj et al., 2023), pretrained depth (Deng et al., 2022; Roessle et al., 2022; Wang et al., 2023; Zhu et al., 2024) and normal (Yu et al., 2022) estimators, and adversarial networks (Roessle et al., 2023). However, these approaches are sensitive to noise, difficult to balance with data terms, and yield only marginal improvements in denser captures. An alternative line of work trains feed-forward networks on large multi-scene datasets, which are used to enhance a scene-optimized NeRF/3DGS (Zhou et al., 2023; Lu et al., 2025b) or directly predict novel views (Yu et al., 2021; Chen et al., 2021; Ren et al., 2024; Lu et al., 2025a). While these deterministic methods perform well near reference views, they often produce blurry results in ambiguous regions where the distribution of possible renderings is inherently multi-modal. Figure 2. Method overview. We first train a bidirectional flow matching model that transports degraded RGB renderings into clean outputs. We encode the input RGB into latent space and mix with Gaussian noise using the rendered opacity maps to avoid mode collapse in unseen regions. We inject fine-grained opacity information and camera control along with optional clean reference views and a text prompt. In the second phase of our pipeline, we distill the teacher into an auto-regressive causal model via Self Forcing-style DMD distillation (Huang et al., 2025), which can be directly used to render novel views or used as pseudo-supervision to distill back into the underlying 3D representation. Diffusion models for novel view synthesis. An alternative strategy is to leverage the priors learned by generative diffusion models trained on internet-scale data to enhance novel view synthesis. Early works (Poole et al., 2023; Sargent et al., 2024; Wu et al., 2024) use a diffusion model as a learned critic during reconstruction optimization, but this incurs substantial computational overhead. More recent approaches (Gao* et al., 2024; Liu et al., 2022, 2024; Wu et al., 2025a, c; Fischer et al., 2025) directly generate multi-view–consistent images that can be consumed by a downstream 3D reconstruction pipeline. While this strategy improves training efficiency, it typically relies on iterative generation and distillation, in which new views are progressively distilled back into the 3D representation to satisfy computational and consistency constraints. Lyra (Bahmani et al., 2026) sidesteps this iteration by distilling video diffusion knowledge into a feed-forward 3DGS generator, though it operates from a single image rather than enhancing an existing reconstruction. Recent work reverses this paradigm by building on the rapid progress of video generation (Blattmann et al., 2023; Wan et al., 2025b). Rather than distilling generative outputs into a 3D representation, these methods treat the 3D representation as a conditioning signal for a generative model that directly synthesizes novel views (Ren et al., 2025; Kong et al., 2025). Although this approach can improve the perceptual realism of novel views, it inherits limitations of the underlying generative models, including temporal inconsistencies, hallucinations, and imperfect camera control. Auto-regressive video generation. While bidirectional video generation models synthesize all frames jointly, auto-regressive models generate frames sequentially using block-causal attention. Auto-regressive generation improves scalability and generation efficiency compared to bidirectional models, but often suffers from quality degradation over time, as each frame is conditioned on previously generated outputs, causing errors to accumulate (Yin et al., 2025a). Several methods try to address the issue by better aligning the training scheme of these models with inference-time conditions, thereby reducing exposure bias (Huang et al., 2025; Cui et al., 2025; Liu et al., 2025). A complementary line of research focuses on improving generation speed and controllability by exploiting temporal and spatial cues to select per-frame context (Yang et al., 2025; Kong et al., 2025; Shin et al., 2025; Wan et al., 2025a; Li et al., 2025b), enabling interactive auto-regressive world models (Hong et al., 2025). Despite these advances, auto-regressive video models still lag behind explicit 3D representations in terms of spatial consistency, camera controllability, and rendering efficiency. 3. Preliminaries 3D Gaussian Splatting. 3DGS (Kerbl et al., 2023) represents a scene as a set of anisotropic 3D Gaussian primitives, each parameterized by a mean j μ_j, covariance j _j, opacity σj _j, and view-dependent color jc_j. Novel views are rendered by projecting the primitives onto the target image plane and compositing in front-to-back depth order: ()=∑iαii∏k<i(1−αk)C(p)= _i _ic_i _k<i(1- _k), where αi _i is the learned opacity scaled by the projected Gaussian evaluated at pixel p. Primitive parameters are optimized per scene with a photometric reconstruction loss. Video diffusion models. Diffusion models learn to transport samples between a data distribution pdata()p_data(x) and a tractable prior, typically (,)N(0,I) (Song et al., 2020; Ho et al., 2020). Most video diffusion models (Blattmann et al., 2023) operate in a lower-dimensional latent space for computational efficiency. Flow matching (Lipman et al., 2023a; Liu et al., 2023), the framework used by our method, learns an ODE flow between two arbitrary endpoint distributions psrcp_src and ptgtp_tgt by fitting a time-dependent vector field θ(t,t)v_θ(z_t,t) whose induced probability path ptt∈[0,1]\p_t\_t∈[0,1] satisfies p0=psrcp_0=p_src and p1=ptgtp_1=p_tgt. During training, we sample endpoint latents 0∼psrcz_0 p_src and 1∼ptgtz_1 p_tgt and a time t∈[0,1]t∈[0,1], construct an intermediate latent via t≔(1−t)0+t1z_t (1-t)z_0+tz_1 with target velocity t≔dtdt=1−0v_t dz_tdt=z_1-z_0, and fit the vector field using the conditional flow matching objective minθt,0,1‖θ(t,t)−t‖22 _θ\ E_t,z_0,z_1 _θ(z_t,t)-v_t _2^2. At inference, we draw 0∼psrcz_0 p_src and numerically integrate the learned ODE from t=0t=0 to t=1t=1 to obtain 1z_1 as a sample from ptgtp_tgt. 4. Method Given an initial 3D reconstruction of a scene created from a sparse set of images, our goal is to generate artifact-free renderings from arbitrary camera viewpoints, including regions unobserved by input images, at interactive rates. Our solution is a controllable auto-regressive video model that can either directly render arbitrary long novel-view renderings or provide pseudo-supervision to improve the underlying 3D reconstruction. We describe how to adapt a pretrained video diffusion model to serve as a bidirectional teacher in Sec. 4.1. We discuss causal distillation and the capabilities of the resulting model in Sec. 4.2. Fig. 2 illustrates our approach. 4.1. Bidirectional Training Figure 3. Transformer block. We start from a pretrained text-to-video model (Wan et al., 2025b) and inject camera and opacity information into each transformer block via linear layers after applying self-attention and layer normalization. We patchify reference views into visual tokens, apply relative camera conditioning via PRoPE (Li et al., 2025a), and add KnK_n and VnV_n projections to the cross-attention operation. We zero-initialize frf_r, fof_o, and VnV_n to ensure compatibility with the pretrained initialization. Architecture. We start from a pretrained text-to-video model (Wan 2.1 T2V-14B (Wan et al., 2025b)), freeze its VAE and text encoder, and finetune the remaining components. Degraded renderings are encoded by the frozen VAE and 3D-patchified with (t,h,w)=(1,2,2)(t,h,w)=(1,2,2), where (t,h,w)(t,h,w) is the temporal/vertical/horizontal patch size in latent voxels. We guide where to generate scene content through rendered opacity maps O and enable camera control in completely unobserved areas via per-pixel Plücker raymaps R, which assign each pixel the six-vector (,×)(d,\,o×d) formed from its ray direction d (unprojected through the camera intrinsics/extrinsics) and the camera center o. Both signals bypass the VAE entirely – we downscale their spatial dimensions to match the spatial compression factor of the VAE via the PixelUnshuffle operation (Paszke et al., 2019), encode them via per-block linear layers fof_o and frf_r (Fig. 3), and add the embeddings to the visual tokens: (1) Tr:=Ts+fr(PixelUnshuffle()) T_r:=T_s+f_r(PixelUnshuffle(R)) (2) To:=Tr+fo(PixelUnshuffle()), T_o:=T_r+f_o(PixelUnshuffle(O)), where TsT_s denotes the token set after applying self-attention and layer-normalization. We found this strategy to be more computationally efficient than alternatives such as VAE encoding R and O while providing camera control even when the input rendering is entirely empty. To provide additional scene context, we encode clean reference views with the frozen VAE, patchified per-image along the batch dimension (no temporal compression). Each transformer block then cross-attends from target tokens (Q) to the concatenated reference tokens, which are mapped to keys and values via additional linear projections KnK_n and VnV_n; the cross-attention output is added back to the target tokens, following the image-to-video variant of Wan 2.1. We apply PRoPE (Li et al., 2025a) only within this cross-attention, using target intrinsics/extrinsics for Q and reference intrinsics/extrinsics for Kn/VnK_n/V_n. frf_r, fof_o, and VnV_n are all zero-initialized to ensure compatibility with the pretrained initialization. Figure 4. Opacity mixing. Given a degraded rendering and optional reference views and text prompt (left), we predict an artifact-free rendering at a target viewpoint. Starting from Gaussian noise and channel concatenating the degraded rendering as in prior work (Wu et al., 2025c; Yin et al., 2025b) produces renderings that are semantically similar to the reference views, but with notable inconsistencies (such as the table in the top row). Directly starting from the degraded rendering instead of Gaussian noise improves consistency, but degrades quality noticeably when extrapolating to areas outside those covered by the degraded renderings (bottom row). Instead, we mix Gaussian noise into the rendering based on its opacity map. The resulting input retains the consistency benefits of the original while enabling a strong generative capability in entirely novel regions. Opacity mixing. Most generative models start from Gaussian noise ϵ∼(,) ε (0,I) which is iteratively transformed into a latent video representation z. Most prior work similarly starts from such noise, conditioning the generation process on the initial degraded rendering latent degz_deg via channel-concatenation (Wu et al., 2025c; Yin et al., 2025b) or classifier-free guidance (Liu et al., 2022). Although the resulting latent enhz_enh tends to be semantically similar to its degraded counterpart, notable inconsistencies remain, especially in high-artifact regions (Fig. 4). Several methods start directly from degz_deg instead of noise (Wu et al., 2025a; Fischer et al., 2025), validating the insight that the source distribution should reflect what can already be rendered. While this encourages stronger consistency guarantees, it suffers from mode collapse in completely unseen areas: the source distribution collapses to a Dirac mass at zero in empty regions, hindering the ability to extrapolate high-quality renderings (Fig. 4). To address this, we mix Gaussian noise into low-opacity regions by downscaling O into zO_z through max pooling to match degz_deg’s spatial dimensions (we retain fine-grained information via Eq. 2) and deriving mix=zdeg+(1−z)ϵz_mix=O_zz_deg+(1-O_z) ε as the source distribution for our model. As no source information is lost from the max-pooling, this approach preserves the consistency benefits of starting from degz_deg while gracefully interpolating to the standard Gaussian prior in entirely novel regions. This strategy is conceptually linked to inpainting methods (Avrahami et al., 2022; Kim et al., 2025; Mayet et al., 2025) that preserve known regions at low noise while pushing unknown regions toward the generative prior, though we operate with a continuous opacity signal rather than a binary mask. We formally derive compatibility with flow matching in Appendix A. Data curation. Our goal is to not only correct artifacts in under-observed areas as in prior work (Wu et al., 2025a; Fischer et al., 2025) but also generate plausible content in entirely unseen areas. To do so, we generate paired reconstruction-ground truth samples from DL3DV-10K (Ling et al., 2024) with a camera selection strategy that encourages highly sparse reconstructions with large empty regions that the model must learn to inpaint. Given a set of camera poses with rotations iR_i and translations it_i, we first measure the camera pose distance dij=θij/π+λt∥i−j∥2/r¯d_ij= _ij/π+ _t\, _i-t_j _2/ r, where θij∈[0,π] _ij∈[0,π] is the SO(3) geodesic angle (in radians) between iR_i and jR_j, r¯=1N∑k∥k∥2 r= 1N _k _k _2 is the mean L2 norm of the camera positions in the scene, and λt=1 _t=1; this puts both terms on the same order of magnitude (the rotation term lies in [0,1][0,1], and translations are normalized to unit mean radius). We then find the camera pair (P1,P2)(P_1,P_2) with the largest distance, and seed groups G1G_1 and G2G_2. We assign the remaining cameras to G1G_1 or G2G_2 based on their distance to P1P_1 and P2P_2, and then sample 2-12 cameras with the largest inter-camera distance within each group to generate reconstructions of differing sparsity. We roughly align the camera scales of each reconstruction with a pretrained metric depth estimator (Wang et al., 2025) and prompt a vision-language model (Bai et al., 2025) for scene descriptions. We provide more details in Appendix G of the supplement. Optimization. Given an initial latent-encoded rendering degz_deg, which we transform into mixz_mix, we train our model to predict its enhanced counterpart enhz_enh via conditional flow matching loss ℒcfmL_cfm (Lipman et al., 2023b). We construct batches of paired reconstruction-ground truth data by sampling N=81N=81 frames along with the corresponding camera poses, text prompt (dropped with 10% probability), and a uniformly varying number of reference views (0-12). To enhance the model’s generative abilities and viewpoint controllability, we drop the last K≤NK≤ N frames of the input (K is randomly chosen) by zeroing both the RGB rendering and opacity map while retaining the Plücker raymaps, so that the model must rebuild the ground truth from the prompt, reference views, and camera conditions alone. 4.2. Causal Distillation Initialization. We initialize the causal model from the weights of the bidirectional teacher. To stabilize training, we follow a simpler strategy than the ODE initialization protocol of prior work (Yin et al., 2025a; Huang et al., 2025; Shin et al., 2025), which requires generating a dataset of ODE trajectories from the teacher model. Instead, we simply apply a block-causal mask, perturb each input frame with differing noise levels as in Diffusion Forcing (Chen et al., 2025a), and otherwise use the same inputs and training protocol as in Sec. 4.1. Autoregressive rollout. After initialization, we adopt a training strategy similar to Self Forcing (Huang et al., 2025), where we generate video chunks sequentially and condition on previously generated chunks via KV caching, except that we continue applying dropout as in Sec. 4.1 as camera control and generation from pure noise otherwise degrade. We apply Distribution Matching Distillation (DMD) (Yin et al., 2024) to convert the model into a few-step generator (N=4N=4 in our experiments, although, outside of entirely novel regions, this can often be reduced to fewer steps with little noticeable difference as discussed in Appendix C of the supplement). Long video generation. Existing methods rely on long-horizon training (Yang et al., 2025; Hong et al., 2025) to minimize error accumulation in long video rollouts. Although these strategies can be applied to our method, in practice we find our conditioning signals (notably the degraded rendering and reference views) sufficient to prevent error accumulation. We thus train with the same number of frames as in Sec. 4.1 and use a rolling KV cache during inference. Although simple, this approach accelerates training convergence (due to training on a more diverse set of shorter videos for a given computational budget) and generalizes to arbitrary length videos, as shown in our experiments. Table 1. Artifact removal on Nerfbusters and DL3DV. All ArtiFixer variants outperform prior methods by a considerable margin, improving PSNR by 2 dB. Nerfbusters (Warburg et al., 2023) DL3DV (Ling et al., 2024) Method PSNR↑ SSIM↑ LPIPS↓ FID↓ PSNR↑ SSIM↑ LPIPS↓ FID↓ Nerfacto (Tancik et al., 2023) 17.29 0.621 0.402 134.65 17.16 0.581 0.430 112.30 3DGS (Kerbl et al., 2023) 17.66 0.678 0.327 113.84 17.18 0.588 0.384 107.23 Nerfbusters (Warburg et al., 2023) 17.72 0.647 0.352 116.83 17.45 0.606 0.370 96.61 GANeRF (Roessle et al., 2023) 17.42 0.611 0.354 115.60 17.54 0.610 0.342 81.44 NeRFLiX (Zhou et al., 2023) 17.91 0.656 0.346 113.59 17.56 0.610 0.359 80.65 Difix3D (Nerfacto) (Wu et al., 2025a) 18.08 0.653 0.328 63.77 17.80 0.596 0.327 50.79 Difix3D (3DGS) (Wu et al., 2025a) 18.14 0.682 0.287 51.34 17.80 0.598 0.314 50.45 Difix3D+ (Nerfacto) (Wu et al., 2025a) 18.32 0.662 0.279 49.44 17.82 0.613 0.283 41.77 Difix3D+ (3DGS) (Wu et al., 2025a) 18.51 0.686 0.264 41.77 17.99 0.602 0.293 40.86 ArtiFixer 19.83 0.701 0.254 37.78 19.73 0.672 0.231 20.85 ArtiFixer3D 20.24 0.729 0.267 39.67 20.14 0.705 0.256 24.27 ArtiFixer3D+ 20.12 0.713 0.264 41.17 20.06 0.686 0.242 22.61 3D distillation. Prior work distills diffusion model outputs into 3D representations (Kerbl et al., 2023) for consistency purposes, as they otherwise exhibit temporal instability (Wu et al., 2025a) or are limited by number of frames bidirectional models can generate in a single pass (Wu et al., 2025c; Fischer et al., 2025). As our auto-regressive model can sequentially generate arbitrary-length renderings, we are not limited by these constraints. However, 3D distillation is still sometimes desirable from an efficiency perspective, as these representations render orders of magnitude faster. To do so, existing methods require a progressive distillation process that alternates between view generation and 3D reconstruction, incurring significant training time overhead. In our case, as we can generate an arbitrary number of frames in a consistent manner, we adopt a more efficient approach by simply generating all desired novel views in a single pass before applying standard 3D reconstruction. 5. Experiments We evaluate three variants of our method: ArtiFixer, which directly renders novel views from the auto-regressive generator, ArtiFixer3D, which distills its outputs back into the underlying 3D representation, and ArtiFixer3D+, which re-applies the auto-regressive model as a post-processing step on top of ArtiFixer3D (as in (Wu et al., 2025a)). We assess their ability to enhance in-the-wild captures against a wide range of prior work in Sec. 5.2 and their capacity to synthesize unobserved regions on a more challenging dataset split against a smaller set of relevant baselines in Sec. 5.3. We validate the contribution of individual components in Sec. 5.4. Table 2. Sparse view reconstruction methods on the Mip-NeRF 360 dataset. We exceed existing work by a wide margin across every metric. PSNR ↑ SSIM ↑ LPIPS ↓ Method 3-view 6-view 9-view 3-view 6-view 9-view 3-view 6-view 9-view Zip-NeRF (Barron et al., 2023) 12.77 13.61 14.30 0.271 0.284 0.312 0.705 0.663 0.633 3DGS (Kerbl et al., 2023) 13.06 14.96 16.79 0.251 0.355 0.447 0.576 0.505 0.446 2DGS (Huang et al., 2024) 13.07 15.02 16.67 0.243 0.338 0.423 0.580 0.506 0.449 FSGS (Zhu et al., 2024) 14.17 16.12 17.94 0.318 0.415 0.492 0.578 0.517 0.468 FreeNeRF (Yang et al., 2023) 12.87 13.35 14.59 0.260 0.283 0.319 0.715 0.717 0.695 SimpleNeRF (Somraj et al., 2023) 13.27 13.67 15.15 0.283 0.312 0.354 0.741 0.721 0.676 DiffusioNeRF (Wynn and Turmukhambetov, 2023) 11.05 12.55 13.37 0.189 0.255 0.267 0.735 0.692 0.680 ZeroNVS (Sargent et al., 2024) 14.44 15.51 15.99 0.316 0.337 0.350 0.680 0.663 0.655 DNGaussian (Li et al., 2024) 14.00 15.21 16.72 0.301 0.356 0.397 0.620 0.604 0.603 FlowR (Fischer et al., 2025) 14.46 16.18 17.53 0.347 0.409 0.456 0.587 0.520 0.467 ReconFusion (Wu et al., 2024) 15.50 16.93 18.19 0.358 0.401 0.432 0.585 0.544 0.511 GenFusion (Wu et al., 2025c) 15.29 17.16 18.36 0.369 0.447 0.496 0.585 0.500 0.465 GSFixer (Yin et al., 2025b) 15.61 17.27 18.63 0.370 0.426 0.481 0.559 0.478 0.420 CAT3D (Gao* et al., 2024) 16.62 17.72 18.67 0.377 0.425 0.460 0.515 0.482 0.460 ArtiFixer 17.06 18.64 19.96 0.420 0.476 0.518 0.437 0.390 0.353 ArtiFixer3D 17.29 18.95 20.24 0.451 0.526 0.598 0.440 0.382 0.327 ArtiFixer3D+ 17.51 18.95 20.16 0.444 0.498 0.537 0.441 0.396 0.359 5.1. Implementation We implement our method in PyTorch (Paszke et al., 2019) and train it on 128 H100 GPUs, using a batch size of one per GPU (128 total). We use FlashAttention-3 (Shah et al., 2024) for acceleration. In our main experiments, we finetune the bidirectional model described in Sec. 4.1 for 15,000 iterations using AdamW (Loshchilov and Hutter, 2019) with a learning rate of 1×10−51× 10^-5. We then initialize the causal model for 5,000 iterations with the same learning rate, followed by 2,000 iterations of auto-regressive rollout and DMD training (≈ 15k GPU-hours total), using learning rates of 2×10−62× 10^-6 for the generator and 4×10−74× 10^-7 for the fake score function. For the ablations, we use a truncated schedule of 10,000 + 2,000 + 600 iterations on 64 GPUs to reduce computational cost (≈ 4k GPU-hours). We use 3DGUT (Wu et al., 2025b) with MCMC densification (Kheradmand et al., 2024) for the initial reconstructions used by our model. At test time, we use K=6K\!=\!6 uniformly sampled reference views for experiments matching the Difix3D+ protocol (Table 1) and all available input views otherwise (Tables 2 and 3). We use prompts generated by a vision-language model (Appendix G). Baselines are evaluated following their standard protocols. 5.2. Enhancing In-the-Wild Captures Datasets. We run comparisons on Nerfbusters (Warburg et al., 2023) and DL3DV (Ling et al., 2024) using the splits provided by (Wu et al., 2025a), and on Mip-NeRF 360 (Barron et al., 2022) with the splits proposed by (Wu et al., 2024) and used in subsequent work (Gao* et al., 2024; Wu et al., 2025c). Baselines. We compare ArtiFixer to an extensive set of baselines, including the original 3DGS (Kerbl et al., 2023) and 2DGS (Huang et al., 2024), NeRF variants (Tancik et al., 2023; Barron et al., 2023), non-generative sparse reconstruction methods (Zhu et al., 2024; Yang et al., 2023; Somraj et al., 2023; Li et al., 2024), and other diffusion-based work (Warburg et al., 2023; Wu et al., 2025a; Wynn and Turmukhambetov, 2023; Sargent et al., 2024; Wu et al., 2024, 2025c; Gao* et al., 2024; Yin et al., 2025b; Fischer et al., 2025). Metrics. We calculate PSNR, SSIM (Wang et al., 2004), LPIPS (Zhang et al., 2018), and FID (Heusel et al., 2017) on Nerfbusters and DL3DV using the same protocol and metric implementations as Difix3D+ (Wu et al., 2025a). On Mip-NeRF 360, we calculate PSNR, SSIM, and LPIPS across the 3-, 6-, and 9-view splits using the same implementations as GenFusion (Wu et al., 2025c). Results. We present quantitative results for Nerfbusters and DL3DV in Table 1 and Mip-NeRF 360 in Table 2. We provide visual comparisons in Fig. 9 and Fig. 10. All ArtiFixer variants outperform all baselines by a substantial margin. Although the different variants produce similar renderings, ArtiFixer’s are slightly sharper, while ArtiFixer3D’s are even more consistent with the source images at the cost of some blurriness due to its explicit 3D representation, leading to a minor increase in PSNR and SSIM and a small degradation in LPIPS and FID in Table 1. Re-applying the generator to the improved 3D reconstruction (ArtiFixer3D+) restores some of this sharpness, leading to renderings that are crisper than ArtiFixer3D and slightly more consistent than ArtiFixer (Fig. 5). Table 3. Novel content generation. We reconstruct DL3DV scenes following a protocol that creates large areas unobserved by training views. We outperform the next-best method (GenFusion) by almost 3 dB in PSNR. Method PSNR↑ SSIM↑ LPIPS↓ FID↓ 3DGUT (Wu et al., 2025b) 16.12 0.537 0.445 92.94 Difix3D (Nerfacto) (Wu et al., 2025a) 14.16 0.453 0.545 74.59 Difix3D (3DGS) (Wu et al., 2025a) 16.60 0.599 0.405 52.70 Difix3D+ (Nerfacto) (Wu et al., 2025a) 13.74 0.434 0.483 30.07 Difix3D+ (3DGS) (Wu et al., 2025a) 16.34 0.564 0.382 21.77 Fixer (offline) (NVIDIA, 2025) 13.09 0.355 0.584 135.43 Fixer (online) (NVIDIA, 2025) 13.93 0.443 0.535 79.44 Gen3C (Ren et al., 2025) 15.50 0.491 0.476 68.36 GenFusion (Wu et al., 2025c) 17.03 0.624 0.392 132.91 ArtiFixer 19.75 0.643 0.303 12.22 ArtiFixer3D 19.92 0.673 0.306 16.28 ArtiFixer3D+ 20.15 0.662 0.307 13.91 5.3. Novel Content Generation Dataset. We evaluate novel content generation by following the sparse reconstruction protocol described in Appendix G on scenes from DL3DV, resulting in numerous “holes” that must be corrected in a manner consistent with existing observations. Baselines. We compare to a smaller set of baselines most relevant to our work, notably 3DGUT (Wu et al., 2025b) as the base representation we provide as initial renderings to our method, image-based diffusion methods via Difix3D+ (Wu et al., 2025a) and Fixer (NVIDIA, 2025), and approaches that build upon bidirectional video models (Wu et al., 2025c; Ren et al., 2025). Figure 5. ArtiFixer variants. Most visible differences occur in highly corrupted regions. ArtiFixer3D’s explicit 3D consistency improves fidelity with the source images and mitigates transient corruption (middle), at the cost of some sharpness, which ArtiFixer3D+ restores. Nonetheless, all variants outperform prior work by a substantial margin. Results. We present quantitative results, using the same metrics as Table 1, in Table 3. We provide qualitative results in Fig. 8. All ArtiFixer variants outperform the next-best method (GenFusion (Wu et al., 2025c)) by almost 3 dB in PSNR. Gen3C (Ren et al., 2025) gives the next-best visually appealing results, but its conditioning often does not respect the source content, and its quality is upper-bounded by the depth estimator it uses to generate its 3D cache (in contrast to our purely data-driven approach). Difix3D+ (Wu et al., 2025a) and Fixer (NVIDIA, 2025) generally fail to inpaint plausible context due to their deterministic conditioning. Table 4. Diagnostics. We evaluate reconstruction quality on Mip-NeRF 360. Denoising input renderings instead of conditioning via channel concatenation is crucial to producing outputs consistent with source images. Method Direct Input Opacity Mixing Diffusion Forcing PSNR↑ SSIM↑ LPIPS↓ FID↓ Channel Concatenation ✗ ✗ ✓ 14.52 0.391 0.490 87.551 w/o Opacity Mixing ✓ ✗ ✓ 17.34 0.440 0.429 87.058 w/o Initialization ✓ ✓ ✗ 17.58 0.450 0.416 74.924 Full Method ✓ ✓ ✓ 17.99 0.461 0.408 69.43 5.4. Diagnostics Ablations. We ablate the effectiveness of our opacity mixing strategy by comparing it to variants that instead use channel concatenation or omit the opacity mixing. We also measure the impact of the causal model weight initialization described in Sec. 4.2. We report results on the Mip-NeRF 360 dataset averaged over all splits in Table 4 and show that our design choice of starting from the initial rendering instead of conditioning on it via channel concatenation is essential to rendering consistently with the source imagery. Our causal initialization method is not essential as the model still converges to a competitive level of quality, but provides a modest boost. Conditioning. To probe which inputs drive output quality, we progressively strip conditioning signals. First, we drop the initial rendering, forcing the model to rely solely on reference views and camera rays. Although fidelity decreases, the model still recovers the high-level scene structure (Fig. 6). Next, we remove all conditioning except the text prompt, reverting to standard text-to-video generation; output quality remains comparable to the base Wan 2.1 model (Fig. 7). Figure 6. Reference views. Without the initial rendering condition, ArtiFixer can generate predictions from the reference views. Although fidelity drops somewhat, the high-level structure of the scene remains intact. Figure 7. Text-to-video generation. To illustrate our model’s generative ability, we generate videos from text prompts alone. With opacity mixing, it retains similar quality to its base model (Wan et al., 2025b). Table 5. Inference speed. Causal distillation yields a 70×70× speedup over the bidirectional Wan 2.1 backbones. ArtiFixer3D renders directly from 3DGUT. Additional configurations are reported in Table 7. Method FPS ↑ Wan 2.1 T2V-14B (Wan et al., 2025b) 0.12 Wan 2.1 T2V-1.3B (Wan et al., 2025b) 0.49 ArtiFixer /ArtiFixer3D+ (14B) 8.36 ArtiFixer /ArtiFixer3D+ (1.3B) 34.38 ArtiFixer3D 268 Model scale. To disentangle model scale from our other contributions, we train with Wan 2.1 T2V-1.3B and report results in Appendix D. Timing. We report inference speed in Table 5 on a single GB300 GPU. Causal distillation with KV caching and few-step sampling yields a 70×70× speedup over the bidirectional Wan 2.1 14B and 1.3B backbones. With the 14B backbone, ArtiFixer and ArtiFixer3D+ reach 8.36 FPS. Our 1.3B variant reaches 34.38 FPS. ArtiFixer3D renders at native 3DGUT speed (268 FPS). Fewer denoising steps and context parallelism provide further gains (Appendix C). 6. Conclusion Neural reconstruction and camera-controlled video generation provide complementary strengths for novel view synthesis. In this work, we introduced ArtiFixer, an auto-regressive video diffusion model that seeks to combine the advantages of both paradigms. ArtiFixer transforms corrupted renderings of reconstructed scenes into clean, temporally consistent frames, while retaining sufficient generative capacity to inpaint unobserved regions and the efficiency required for interactive use. The strong conditioning signal from the reconstructed scene significantly simplifies distillation and conversion to an auto-regressive formulation, enabling ArtiFixer to generate long video sequences with less quality degradation. 7. Acknowledgments We thank Zian Wang and Nicholas Sharp for their helpful advice and feedback throughout this project. References M. Asim, C. Wewer, T. Wimmer, B. Schiele, and J. E. Lenssen (2025) MEt3R: measuring multi-view consistency in generated images. In CVPR, Cited by: Table 11, Appendix D. O. Avrahami, D. Lischinski, and O. Fried (2022) Blended diffusion for text-driven editing of natural images. In CVPR, Cited by: §4.1. S. Bahmani, T. Shen, J. Ren, J. Huang, Y. Jiang, H. Turki, A. Tagliasacchi, D. B. Lindell, Z. Gojcic, S. Fidler, H. Ling, J. Gao, and X. Ren (2026) Lyra: generative 3d scene reconstruction via self-distillation with video diffusion models. In ICLR, Cited by: §2. S. Bai, Y. Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, W. Ge, Z. Guo, Q. Huang, J. Huang, F. Huang, B. Hui, S. Jiang, Z. Li, M. Li, M. Li, K. Li, Z. Lin, J. Lin, X. Liu, J. Liu, C. Liu, Y. Liu, D. Liu, S. Liu, D. Lu, R. Luo, C. Lv, R. Men, L. Meng, X. Ren, X. Ren, S. Song, Y. Sun, J. Tang, J. Tu, J. Wan, P. Wang, P. Wang, Q. Wang, Y. Wang, T. Xie, Y. Xu, H. Xu, J. Xu, Z. Yang, M. Yang, J. Yang, A. Yang, B. Yu, F. Zhang, H. Zhang, X. Zhang, B. Zheng, H. Zhong, J. Zhou, F. Zhou, J. Zhou, Y. Zhu, and K. Zhu (2025) Qwen3-vl technical report. External Links: 2511.21631, Link Cited by: Appendix G, §4.1. J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman (2022) Mip-nerf 360: unbounded anti-aliased neural radiance fields. In CVPR, Cited by: §5.2. J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman (2023) Zip-nerf: anti-aliased grid-based neural radiance fields. In ICCV, Cited by: §5.2, Table 2. A. Blattmann, T. Dockhorn, S. Kulal, D. Mendelevitch, M. Kilian, D. Lorenz, Y. Levi, Z. English, V. Voleti, A. Letts, et al. (2023) Stable video diffusion: scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127. Cited by: §2, §3. A. Chen, Z. Xu, F. Zhao, X. Zhang, F. Xiang, J. Yu, and H. Su (2021) Mvsnerf: fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, p. 14124–14133. Cited by: §2. B. Chen, D. Martí Monsó, Y. Du, M. Simchowitz, R. Tedrake, and V. Sitzmann (2025a) Diffusion forcing: next-token prediction meets full-sequence diffusion. NeurIPS 37, p. 24081–24125. Cited by: §4.2. L. Chen, Z. Zhou, M. Zhao, Y. Wang, G. Zhang, W. Huang, H. Sun, J. Wen, and C. Li (2025b) FlexWorld: progressively expanding 3d scenes for flexiable-view synthesis. In NeurIPS, Cited by: §1. J. Cui, J. Wu, M. Li, T. Yang, X. Li, R. Wang, A. Bai, Y. Ban, and C. Hsieh (2025) Self-forcing++: towards minute-scale high-quality video generation. arXiv preprint arXiv:2510.02283. Cited by: §2. K. Deng, A. Liu, J. Zhu, and D. Ramanan (2022) Depth-supervised nerf: fewer views and faster training for free. In CVPR, p. 12882–12891. Cited by: §2. T. Fischer, S. R. Bulò, Y. Yang, N. Keetha, L. Porzi, N. Müller, K. Schwarz, J. Luiten, M. Pollefeys, and P. Kontschieder (2025) FlowR: flowing from sparse to dense 3d reconstructions. In ICCV, Cited by: §1, §2, §4.1, §4.1, §4.2, §5.2, Table 2. R. Gao*, A. Holynski*, P. Henzler, A. Brussee, R. Martin-Brualla, P. P. Srinivasan, J. T. Barron, and B. Poole* (2024) CAT3D: create anything in 3d with multi-view diffusion models. Cited by: Table 8, Appendix D, §1, §2, §5.2, §5.2, Table 2. Google DeepMind (2024) Veo: a generative model for high-quality video. Note: https://deepmind.google/technologies/veo/Accessed: 2025 Cited by: §1. M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter (2017) Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS 30. Cited by: §5.2. J. Ho, A. Jain, and P. Abbeel (2020) Denoising diffusion probabilistic models. NeurIPS. Cited by: §3. Y. Hong, Y. Mei, C. Ge, Y. Xu, Y. Zhou, S. Bi, Y. Hold-Geoffroy, M. Roberts, M. Fisher, E. Shechtman, K. Sunkavalli, F. Liu, Z. Li, and H. Tan (2025) RELIC: interactive video world model with long-horizon memory. External Links: 2512.04040, Link Cited by: Appendix G, §2, §4.2. B. Huang, Z. Yu, A. Chen, A. Geiger, and S. Gao (2024) 2D gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH Asia, Cited by: §5.2, Table 2. X. Huang, Z. Li, G. He, M. Zhou, and E. Shechtman (2025) Self forcing: bridging the train-test gap in autoregressive video diffusion. Cited by: Figure 2, §2, §4.2, §4.2. B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis (2023) 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42 (4). External Links: Link Cited by: Table 10, §1, §2, §3, §4.2, Table 1, §5.2, Table 2. S. Kheradmand, D. Rebain, G. Sharma, W. Sun, Y. Tseng, H. Isack, A. Kar, A. Tagliasacchi, and K. M. Yi (2024) 3d gaussian splatting as markov chain monte carlo. Advances in Neural Information Processing Systems 37, p. 80965–80986. Cited by: Appendix G, §5.1. S. Kim, S. Suh, and M. Lee (2025) RAD: region-aware diffusion models for image inpainting. In CVPR, Cited by: §4.1. A. Knapitsch, J. Park, Q. Zhou, and V. Koltun (2017) Tanks and temples: benchmarking large-scale scene reconstruction. ACM Transactions on Graphics 36 (4). Cited by: Appendix D. H. Kong, X. Yang, X. Zheng, and X. Wang (2025) WorldWarp: propagating 3d geometry with asynchronous video diffusion. arXiv preprint arXiv:2512.19678. Cited by: §2, §2. V. Leroy, Y. Cabon, and J. Revaud (2024) Grounding image matching in 3d with mast3r. In ECCV, Cited by: Appendix D. J. Li, J. Zhang, X. Bai, J. Zheng, X. Ning, J. Zhou, and L. Gu (2024) DNGaussian: optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In CVPR, Cited by: Table 10, §5.2, Table 2. R. Li, B. Yi, J. Liu, H. Gao, Y. Ma, and A. Kanazawa (2025a) Cameras as relative positional encoding. Cited by: Figure 3, §4.1. R. Li, P. Torr, A. Vedaldi, and T. Jakab (2025b) VMem: consistent interactive video scene generation with surfel-indexed view memory. In ICCV, Cited by: §2. L. Ling, Y. Sheng, Z. Tu, W. Zhao, C. Xin, K. Wan, L. Yu, Q. Guo, Z. Yu, Y. Lu, et al. (2024) Dl3dv-10k: a large-scale scene dataset for deep learning-based 3d vision. In CVPR, p. 22160–22169. Cited by: §4.1, Table 1, §5.2. Y. Lipman, \. Chen, H. Ben-Hamu, M. Nickel, and M. Le (2023a) Flow matching for generative modeling. In ICLR, Cited by: Appendix A, §3. Y. Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le (2023b) Flow matching for generative modeling. In ICLR, Cited by: §4.1. F. Liu, W. Wu, H. Tan, Y. Yuan, Y. Zhou, J. Liu, K. Duan, H. Xie, J. Pei, H. Wang, et al. (2026) ReconX: reconstruct any scene from sparse views with video diffusion model. IEEE Transactions on Image Processing. Cited by: Table 10, Appendix D. K. Liu, W. Hu, J. Xu, Y. Shan, and S. Lu (2025) Rolling forcing: autoregressive long video diffusion in real time. arXiv preprint arXiv:2509.25161. Cited by: §2. X. Liu, C. Zhou, and S. Huang (2022) 3DGS-enhancer: enhancing unbounded 3d gaussian splatting with view-consistent 2d diffusion priors. Cited by: §2, §4.1. X. Liu, C. Gong, and Q. Liu (2023) Flow straight and fast: learning to generate and transfer data with rectified flow. In ICLR, Cited by: §3. X. Liu, J. Chen, S. Kao, Y. Tai, and C. Tang (2024) Deceptive-nerf: enhancing nerf reconstruction using pseudo-observations from diffusion models. Cited by: §2. I. Loshchilov and F. Hutter (2019) Decoupled weight decay regularization. In ICLR, Cited by: §5.1. Y. Lu, X. Ren, J. Yang, T. Shen, Z. Wu, J. Gao, Y. Wang, S. Chen, M. Chen, S. Fidler, et al. (2025a) InfiniCube: unbounded and controllable dynamic 3d driving scene generation with world-guided video models. In ICCV, Cited by: §2. Y. Lu, J. Zhang, T. Fang, J. Nahmias, Y. Tsin, L. Quan, X. Cao, Y. Yao, and S. Li (2025b) Matrix3D: large photogrammetry model all-in-one. CVPR. Cited by: §2. T. Mayet, P. Shamsolmoali, S. Bernard, E. Granger, R. Hérault, and C. Chatelain (2025) TD-paint: faster diffusion inpainting through time aware pixel conditioning. In ICLR, Cited by: §4.1. B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng (2020) NeRF: representing scenes as neural radiance fields for view synthesis. In ECCV, Cited by: §1, §2. M. Niemeyer, J. T. Barron, B. Mildenhall, M. S. M. Sajjadi, A. Geiger, and N. Radwan (2022) RegNeRF: regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, Cited by: §2. NVIDIA, N. Agarwal, A. Ali, M. Bala, Y. Balaji, E. Barker, T. Cai, P. Chattopadhyay, Y. Chen, Y. Cui, Y. Ding, D. Dworakowski, J. Fan, M. Fenzi, F. Ferroni, S. Fidler, D. Fox, S. Ge, Y. Ge, J. Gu, S. Gururani, E. He, J. Huang, J. Huffman, P. Jannaty, J. Jin, S. W. Kim, G. Klár, G. Lam, S. Lan, L. Leal-Taixe, A. Li, Z. Li, C. Lin, T. Lin, H. Ling, M. Liu, X. Liu, A. Luo, Q. Ma, H. Mao, K. Mo, A. Mousavian, S. Nah, S. Niverty, D. Page, D. Paschalidou, Z. Patel, L. Pavao, M. Ramezanali, F. Reda, X. Ren, V. R. N. Sabavat, E. Schmerling, S. Shi, B. Stefaniak, S. Tang, L. Tchapmi, P. Tredak, W. Tseng, J. Varghese, H. Wang, H. Wang, H. Wang, T. Wang, F. Wei, X. Wei, J. Z. Wu, J. Xu, W. Yang, L. Yen-Chen, X. Zeng, Y. Zeng, J. Zhang, Q. Zhang, Y. Zhang, Q. Zhao, and A. Zolkowski (2025) Cosmos world foundation model platform for physical ai. External Links: Link Cited by: §1. NVIDIA (2025) NVIDIA fixer. Note: https://huggingface.co/nvidia/FixerAccessed: 2026-01-26 Cited by: Table 11, §5.3, §5.3, Table 3, Table 3, Figure 8. OpenAI (2024) Sora: creating video from text. Note: https://openai.com/soraAccessed: 2025 Cited by: §1. A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, et al. (2019) Pytorch: an imperative style, high-performance deep learning library. Advances in neural information processing systems 32. Cited by: §4.1, §5.1. B. Poole, A. Jain, J. T. Barron, and B. Mildenhall (2023) DreamFusion: text-to-3d using 2d diffusion. In ICLR, Cited by: §2. X. Ren, Y. Lu, H. Liang, J. Z. Wu, H. Ling, M. Chen, F. Fidler, and J. Huang (2024) SCube: instant large-scale scene reconstruction using voxsplats. In NeurIPS, Cited by: §2. X. Ren, T. Shen, J. Huang, H. Ling, Y. Lu, M. Nimier-David, T. Müller, A. Keller, S. Fidler, and J. Gao (2025) GEN3C: 3d-informed world-consistent video generation with precise camera control. In CVPR, Cited by: Table 11, Table 9, Appendix D, §1, §2, §5.3, §5.3, Table 3, Figure 8. B. Roessle, J. T. Barron, B. Mildenhall, P. P. Srinivasan, and M. Nießner (2022) Dense depth priors for neural radiance fields from sparse input views. In CVPR, p. 12892–12901. Cited by: §2. B. Roessle, N. Müller, L. Porzi, S. R. Bulò, P. Kontschieder, and M. Nießner (2023) Ganerf: leveraging discriminators to optimize neural radiance fields. ACM Transactions on Graphics (TOG) 42 (6), p. 1–14. Cited by: §2, Table 1. K. Sargent, Z. Li, T. Shah, C. Herrmann, H. Yu, Y. Zhang, E. R. Chan, D. Lagun, L. Fei-Fei, D. Sun, and J. Wu (2024) ZeroNVS: zero-shot 360-degree view synthesis from a single image. In CVPR, Cited by: §2, §5.2, Table 2. J. Shah, G. Bikshandi, Y. Zhang, V. Thakkar, P. Ramani, and T. Dao (2024) FlashAttention-3: fast and accurate attention with asynchrony and low-precision. External Links: 2407.08608, Link Cited by: §5.1. J. Shin, Z. Li, R. Zhang, J. Zhu, J. Park, E. Shechtman, and X. Huang (2025) MotionStream: real-time video generation with interactive motion controls. arXiv preprint arXiv:2511.01266. Cited by: §2, §4.2. N. Somraj, A. Karanayil, and R. Soundararajan (2023) SimpleNeRF: regularizing sparse input neural radiance fields with simpler solutions. In SIGGRAPH Asia, External Links: Document Cited by: §2, §5.2, Table 2. Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole (2020) Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456. Cited by: §3. M. Tancik, E. Weber, E. Ng, R. Li, B. Yi, T. Wang, A. Kristoffersen, J. Austin, K. Salahi, A. Ahuja, et al. (2023) Nerfstudio: a modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, p. 1–12. Cited by: Table 1, §5.2. Z. Teed and J. Deng (2020) RAFT: recurrent all-pairs field transforms for optical flow. In ECCV, Cited by: Appendix D. H. Wan, J. Zhang, R. Zhang, J. Luo, X. Fang, L. Yang, Y. Cao, and Y. Shan (2025a) Voyager: long-range and world-consistent video diffusion for explorable 3d scene generation. ACM Transactions on Graphics. Cited by: §2. T. Wan, A. Wang, B. Ai, B. Wen, C. Mao, C. Xie, D. Chen, F. Yu, H. Zhao, J. Yang, J. Zeng, J. Wang, J. Zhang, J. Zhou, J. Wang, J. Chen, K. Zhu, K. Zhao, K. Yan, L. Huang, M. Feng, N. Zhang, P. Li, P. Wu, R. Chu, R. Feng, S. Zhang, S. Sun, T. Fang, T. Wang, T. Gui, T. Weng, T. Shen, W. Lin, W. Wang, W. Wang, W. Zhou, W. Wang, W. Shen, W. Yu, X. Shi, X. Huang, X. Xu, Y. Kou, Y. Lv, Y. Li, Y. Liu, Y. Wang, Y. Zhang, Y. Huang, Y. Li, Y. Wu, Y. Liu, Y. Pan, Y. Zheng, Y. Hong, Y. Shi, Y. Feng, Z. Jiang, Z. Han, Z. Wu, and Z. Liu (2025b) Wan: open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314. Cited by: §2, Figure 3, §4.1, Figure 7, Table 5, Table 5. G. Wang, Z. Chen, C. C. Loy, and Z. Liu (2023) Sparsenerf: distilling depth ranking for few-shot novel view synthesis. In ICCV, p. 9065–9076. Cited by: Table 10, §2. R. Wang, S. Xu, Y. Dong, Y. Deng, J. Xiang, Z. Lv, G. Sun, X. Tong, and J. Yang (2025) MoGe-2: accurate monocular geometry with metric scale and sharp details. In CVPR, Cited by: §4.1. Z. Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli (2004) Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13 (4), p. 600–612. External Links: Document Cited by: §5.2. F. Warburg, E. Weber, M. Tancik, A. Holynski, and A. Kanazawa (2023) Nerfbusters: removing ghostly artifacts from casually captured nerfs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 18120–18130. Cited by: Table 1, Table 1, §5.2, §5.2. Y. Wen, J. Kirchenbauer, J. Geiping, and T. Goldstein (2023) Tree-ring watermarks: fingerprints for diffusion images that are invisible and robust. In NeurIPS, Cited by: Appendix F. J. Z. Wu, Y. Zhang, H. Turki, X. Ren, J. Gao, M. Z. Shou, S. Fidler, Z. Gojcic, and H. Ling (2025a) DIFIX3D+: improving 3d reconstructions with single-step diffusion models. In CVPR, p. 26024–26035. Cited by: Table 11, §1, §2, §4.1, §4.1, §4.2, Table 1, Table 1, Table 1, Table 1, §5.2, §5.2, §5.2, §5.3, §5.3, Table 3, Table 3, Table 3, Table 3, §5, Figure 8. Q. Wu, J. Martinez Esturo, A. Mirzaei, N. Moenne-Loccoz, and Z. Gojcic (2025b) 3DGUT: enabling distorted cameras and secondary rays in gaussian splatting. In CVPR, Cited by: Appendix G, §5.1, §5.3, Table 3, Figure 8. R. Wu, B. Mildenhall, P. Henzler, K. Park, R. Gao, D. Watson, P. P. Srinivasan, D. Verbin, J. T. Barron, B. Poole, and A. Holynski (2024) ReconFusion: 3d reconstruction with diffusion priors. In CVPR, Cited by: §1, §2, §5.2, §5.2, Table 2. S. Wu, C. Xu, B. Huang, G. Andreas, and A. Chen (2025c) GenFusion: closing the loop between reconstruction and generation via videos. In CVPR, Cited by: Table 11, Table 8, Table 9, Appendix D, §1, §2, Figure 4, §4.1, §4.2, §5.2, §5.2, §5.2, §5.3, §5.3, Table 2, Table 3, Figure 8. T. Wu, S. Yang, R. Po, Y. Xu, Z. Liu, D. Lin, and G. Wetzstein (2025d) Video world models with long-term spatial memory. In NeurIPS, Cited by: §1. J. Wynn and D. Turmukhambetov (2023) DiffusioNeRF: regularizing neural radiance fields with denoising diffusion models. In CVPR, Cited by: §5.2, Table 2. J. Yang, M. Pavone, and Y. Wang (2023) FreeNeRF: improving few-shot neural rendering with free frequency regularization. In CVPR, Cited by: §2, §5.2, Table 2. S. Yang, W. Huang, R. Chu, Y. Xiao, Y. Zhao, X. Wang, M. Li, E. Xie, Y. Chen, Y. Lu, and S. H. Y. Chen (2025) LongLive: real-time interactive long video generation. External Links: 2509.22622 Cited by: §2, §4.2. T. Yin, M. Gharbi, R. Zhang, E. Shechtman, F. Durand, W. T. Freeman, and T. Park (2024) One-step diffusion with distribution matching distillation. In CVPR, Cited by: §4.2. T. Yin, Q. Zhang, R. Zhang, W. T. Freeman, F. Durand, E. Shechtman, and X. Huang (2025a) From slow bidirectional to fast autoregressive video diffusion models. Cited by: §2, §4.2. X. Yin, Q. Zhang, J. Chang, Y. Feng, Q. Fan, X. Yang, C. Pun, H. Zhang, and X. Cun (2025b) GSFixer: improving 3d gaussian splatting with reference-guided video diffusion priors. External Links: 2508.09667, Link Cited by: Table 8, Appendix D, Figure 4, §4.1, §5.2, Table 2. A. Yu, V. Ye, M. Tancik, and A. Kanazawa (2021) pixelNeRF: neural radiance fields from one or few images. In CVPR, Cited by: §2. W. Yu, J. Xing, L. Yuan, W. Hu, X. Li, Z. Huang, X. Gao, T. Wong, Y. Shan, and Y. Tian (2024) Viewcrafter: taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048. Cited by: §1. Z. Yu, S. Peng, M. Niemeyer, T. Sattler, and A. Geiger (2022) Monosdf: exploring monocular geometric cues for neural implicit surface reconstruction. Vol. 35, p. 25018–25032. Cited by: §2. S. Zhai, Z. Ye, J. Liu, W. Xie, J. Hu, Z. Peng, H. Xue, D. Chen, X. Wang, L. Yang, N. Wang, H. Liu, and G. Zhang (2025) StarGen: a spatiotemporal autoregression framework with video diffusion model for scalable and controllable scene generation. In CVPR, Cited by: §1. R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang (2018) The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, Cited by: §5.2. J. (. Zhou, H. Gao, V. Voleti, A. Vasishta, C. Yao, M. Boss, P. Torr, C. Rupprecht, and V. Jampani (2025) Stable virtual camera: generative view synthesis with diffusion models. arXiv preprint. Cited by: §1. K. Zhou, W. Li, Y. Wang, T. Hu, N. Jiang, X. Han, and J. Lu (2023) NeRFLix: high-quality neural view synthesis by learning a degradation-driven inter-viewpoint mixer. In CVPR, p. 12363–12374. Cited by: §2, Table 1. Z. Zhu, Z. Fan, Y. Jiang, and Z. Wang (2024) FSGS: real-time few-shot view synthesis using gaussian splatting. In ECCV, Cited by: §2, §5.2, Table 2. J. Zhuang, S. Guo, X. Cai, X. Li, Y. Liu, C. Yuan, and T. Xue (2026) FlashVSR: towards real-time diffusion-based streaming video super-resolution. In CVPR, Cited by: Appendix E. Figure 8. DL3DV results. We compare ArtiFixer3D+ to its initial 3DGUT (Wu et al., 2025b) input, two baselines that build upon bidirectional video diffusion models (top rows), and two that leverage image models (bottom rows). GenFusion (Wu et al., 2025c)’s video model generates 16 frames at a time, requiring an iterative distillation process that leads to blurry results, especially in empty areas. Gen3C (Ren et al., 2025)’s renderings are sharper but often do not respect the source content (background in top row), have incorrect geometry (second row), and exhibit color shift (sixth row). Methods that directly take renderings as input without opacity mixing (Wu et al., 2025a; NVIDIA, 2025) fail to reconstruct empty regions. Our method can reconstruct plausible and consistent geometry even when the initial rendering is highly degraded. Please refer to our project website for comparison videos. Figure 9. Mip-NeRF 360 results. We present visualizations of Mip-NeRF’s most challenging split (3-view). Our results far exceed all prior work both quantitatively and qualitatively. Our method is able to recover the correct geometry from the reference views even in scenarios where the input rendering is completely inaccurate (table in third row). Figure 10. Nerfbusters results. As with the other datasets, our method is the only one to generate plausible visuals in unseen regions while preserving the fidelity of the original content. Supplementary Material Appendix A Opacity Mixing and Flow Matching Our opacity mixing strategy is fully compatible with the conditional flow matching (CFM) framework (Lipman et al., 2023a) as the CFM loss t,0,1‖θ(t,t,cond)−(1−0)‖2E_t,z_0,z_1 _θ(z_t,t,cond)-(z_1-z_0) ^2 is valid for any joint distribution q(0,1)q(z_0,z_1), not only 0∼(,)z_0 (0,I). In our setting, we define the source sample as: (3) 0≔zdeg+(1−z)ϵ,ϵ∼(,),z_0 _zz_deg+(1-O_z) ε, ε (0,I), where zO_z is the spatially varying, downscaled opacity map and degz_deg is the VAE-encoded degraded rendering. Let 1z_1 denote the clean target latent. We sample a global scalar t∼[0,1]t [0,1] and form the interpolant: (4) t=(1−t)0+t1,z_t=(1-t)\,z_0+t\,z_1, with target velocity t=1−0v_t=z_1-z_0. The spatial variation introduced by zO_z is encoded entirely in 0z_0 and consequently propagates to both tz_t and the target velocity tv_t, not to the scalar time variable t. No per-location timestep conditioning is required: the network receives (t,t,cond)(z_t,t,cond) with a single global t, exactly as in standard flow matching. At inference, we draw 0∼q(0)z_0 q(z_0) using the same opacity mixing procedure and integrate the learned ODE from t=0t=0 to t=1t=1 using the same global time parameterization. Appendix B Text Conditioning Dataset Δ Δ Δ Mip-NeRF 360 (3 views) +0.14 +0.003 −-0.002 Mip-NeRF 360 (6 views) +0.07 +0.002 −-0.001 Mip-NeRF 360 (9 views) +0.03 +0.003 −-0.001 DL3DV +0.02 0.000 −-0.001 Nerfbusters −-0.07 +0.001 0.000 Table 6. Text conditioning. We measure the impact of VLM-generated prompts vs. no prompt for ArtiFixer3D+. Text prompts provide a small benefit in sparse settings that diminishes with denser captures. We further quantify the contribution of text conditioning by comparing ArtiFixer3D+ results with and without VLM-generated prompts in Table 6. Text conditioning provides a minor benefit in the most sparse settings (+0.14 dB PSNR on Mip-NeRF 360 with 3 views), but this effect diminishes with denser captures. FPS ↑ Method GPUs 1 step 2 steps 3 steps 4 steps ArtiFixer (14B) 1 29.42 16.07 11.03 8.36 ArtiFixer (14B) 4 58.72 35.91 24.65 19.18 ArtiFixer (1.3B) 1 86.75 57.76 43.20 34.38 ArtiFixer (1.3B) 4 101.77 69.44 53.77 49.24 Table 7. Inference configurations. Fewer denoising steps and context parallelism across multiple GPUs further improve throughput, with the 1.3B variant reaching up to 101.77 FPS. Appendix C Denoising Steps As ArtiFixer starts from renderings instead of pure noise, it is able to generate plausible visuals in fewer than four steps in most cases. Reducing the number of denoising steps significantly improves throughput, with context parallelism across multiple GPUs providing further gains (Table 7). However, sharpness and temporal consistency suffer somewhat in empty areas (Fig. 11). This is largely mitigated when revisiting previously explored areas in our ArtiFixer3D and ArtiFixer3D+ variants, as the 3D distillation process provides strong conditioning for subsequent generations. Figure 11. Denoising steps. We vary the number of denoising steps when beginning from the initial degraded rendering. ArtiFixer can render plausible content in as few as 1 step, although sharpness and temporal consistency suffer somewhat in empty areas. PSNR ↑ SSIM ↑ LPIPS ↓ Method 3-view 6-view 9-view 3-view 6-view 9-view 3-view 6-view 9-view GenFusion (Wu et al., 2025c) 15.29 17.16 18.36 0.369 0.447 0.496 0.585 0.500 0.465 GSFixer (Yin et al., 2025b) 15.61 17.27 18.63 0.370 0.426 0.481 0.559 0.478 0.420 CAT3D (Gao* et al., 2024) 16.62 17.72 18.67 0.377 0.425 0.460 0.515 0.482 0.460 ArtiFixer3D+ (1.3B) 16.60 18.04 19.44 0.414 0.466 0.513 0.486 0.435 0.394 ArtiFixer3D+ (14B) 17.51 18.95 20.16 0.444 0.498 0.537 0.441 0.396 0.359 Table 8. Impact of model scale on Mip-NeRF 360. Our 1.3B variant matches CAT3D within 0.02 dB on the 3-view split and exceeds other video model baselines despite using fewer parameters. Method PSNR ↑ SSIM ↑ LPIPS ↓ FID ↓ GenFusion (Wu et al., 2025c) 17.03 0.624 0.392 132.91 Gen3C (Ren et al., 2025) 15.50 0.491 0.476 68.36 ArtiFixer3D+ (1.3B) 19.04 0.635 0.352 22.3 ArtiFixer3D+ (14B) 20.15 0.662 0.307 13.91 Table 9. Impact of model scale on novel content generation (DL3DV). Even with a 1.3B backbone, ArtiFixer3D+ outperforms the other video model baselines by a wide margin. Method PSNR ↑ SSIM ↑ LPIPS ↓ 3DGS (Kerbl et al., 2023) 9.57 0.108 0.779 SparseNeRF (Wang et al., 2023) 9.23 0.191 0.632 DNGaussian (Li et al., 2024) 10.23 0.156 0.643 ReconX (Liu et al., 2026) 14.28 0.394 0.564 ArtiFixer3D+ 14.75 0.464 0.463 Table 10. Tanks and Temples (2-view). ArtiFixer3D+ outperforms all baselines. Method MASt3R ↓ RAFT ↓ Fixer (NVIDIA, 2025) 0.1288 0.1236 Difix3D+ (Wu et al., 2025a) 0.0974 0.0959 GenFusion (Wu et al., 2025c) 0.0817 0.0786 Gen3C (Ren et al., 2025) 0.0766 0.0757 ArtiFixer 0.0749 0.0749 ArtiFixer3D+ 0.0697 0.0697 ArtiFixer3D 0.0646 0.0647 Table 11. Multi-view consistency. We measure multi-view consistency via MEt3R (Asim et al., 2025) with MASt3R and RAFT backbones. All ArtiFixer variants outperform baselines, with ArtiFixer3D achieving the best results due to its explicit multi-view-consistent 3D representation. Appendix D Additional Experiments Model scale. To disentangle the contribution of our method from backbone capacity, we train the full pipeline with Wan 2.1 T2V-1.3B and report ArtiFixer3D+ results in Tables 8 and 9. For reference, GenFusion (Wu et al., 2025c) uses a 1.4B-parameter backbone, GSFixer (Yin et al., 2025b) 5B, and Gen3C (Ren et al., 2025) 7B. Our 1.3B variant matches CAT3D (Gao* et al., 2024) within 0.02 dB on the 3-view Mip-NeRF 360 split and exceeds all other baselines. Tanks and Temples. To further evaluate generalization, we report results on the Tanks and Temples dataset (Knapitsch et al., 2017) using the 2-view setting from ReconX (Liu et al., 2026) in Table 10. Multi-view consistency. We evaluate multi-view consistency using MEt3R (Asim et al., 2025) with MASt3R (Leroy et al., 2024) depth-based reprojection and RAFT (Teed and Deng, 2020) optical flow-based warping backbones in Table 11. All ArtiFixer variants outperform baselines, with ArtiFixer3D achieving the best consistency due to its explicit 3D representation. Appendix E Limitations While ArtiFixer reaches interactive rates, it remains significantly slower than direct rendering from neural scene representations. Decoding in temporal chunks also introduces latency that may be undesirable for applications such as embodied AI. Additionally, the ArtiFixer and ArtiFixer3D+ variants are limited to 720p by the backbone video model, whereas ArtiFixer3D renders at the native resolution of the underlying 3D representation. As with other video diffusion models, our method can occasionally blur fine details and text, and may introduce subtle color shifts when the rendering condition is absent or highly degraded. Promising directions for future work include further reducing denoising steps, enabling single-frame decoding while maintaining temporal coherence, and applying video super-resolution (Zhuang et al., 2026) to close the resolution gap. Appendix F Societal Impact ArtiFixer synthesizes photorealistic scene content and can plausibly inpaint unobserved regions, raising concerns about potential misuse for generating deceptive visual media. Appropriate safeguards such as watermarking generated content (Wen et al., 2023) should be considered for deployment. From an environmental perspective, training our 14B-parameter model requires approximately 15k GPU-hours on H100 hardware. Our truncated training schedule achieves near-full quality at roughly 25% of this cost, and our 1.3B-parameter variant further reduces training compute while remaining competitive with prior work. Appendix G Sparse Reconstruction Camera Sampling. We describe our camera sampling strategy in Algorithm 1. Given a set of camera poses P, we define the pairwise distance between two poses as d=‖i−j‖F+‖i−j‖2d=||R_i-R_j||_F+||t_i-t_j||_2. We initialize the clustering process by identifying the pair (P1,P2)(P_1,P_2) that maximizes this distance and using them as seeds for groups G1G_1 and G2G_2. The remaining cameras are assigned to the group of their nearest seed. Finally, to evaluate varying levels of sparsity, we apply farthest point sampling within each group to select subsets of size K=2,⋯,12K=\2,·s,12\. Input: Camera poses P, Selection count K, Distance function d Output: Selected subsets 1⊂G1S_1⊂ G_1 and 2⊂G2S_2⊂ G_2 /* 1. Find global farthest camera pair */ (P1,P2)←argmaxPi,Pj∈d(Pi,Pj)(P_1,P_2)← *argmax_P_i,P_j d(P_i,P_j); /* 2. Cluster: Assign cameras to nearest seed camera */ G1←P∈∣D(P,P1)≤D(P,P2)G_1←\P D(P,P_1)≤ D(P,P_2)\; G2←∖G1G_2 G_1; /* 3. Select Top-K points in EACH group */ foreach i∈1,2i∈\1,2\ do i←PiS_i←\P_i\ ; // Start with the seed camera while |i|<K|S_i|<K and |i|<|Gi||S_i|<|G_i| do /* Find pose maximizing distance to current selection */ Pnext←argmaxP∈Gi∖i(mins∈iD(P,s))P_next← *argmax_P∈ G_i _i ( _s _iD(P,s) ); i←i∪PnextS_i _i∪\P_next\; end while end foreach return 1,2S_1,S_2 ALGORITHM 1 CameraSampling Reconstruction. We generate the initial reconstructions we pass to the ArtiFixer model using the official 3DGUT implementation (Wu et al., 2025b) with MCMC (Kheradmand et al., 2024) sampling (reconstructions used during training are prepared offline). We run each reconstruction for 10,000 iterations, taking slightly less than 10 minutes per reconstruction. Captioning. We generate captions for each DL3DV scene from Qwen3-VL-30B-A3B-Instruct (Bai et al., 2025) on different frame subsets to encourage prompt diversity. Similar to (Hong et al., 2025), we suppress descriptions of ego-camera movement to avoid entanglement with camera ray conditioning. We use the prompt below: You are a video captioning specialist whose goal is to generate high-quality English prompts by referring to the details of the user’s input videos. Your task is to carefully analyze the content, context, and actions within the video, and produce a complete, expressive, and natural-sounding caption that accurately conveys the scene. The caption should preserve the original intent and meaning of the video while enhancing its clarity and descriptive richness. Strictly adhere to the formatting of the examples provided. Task Requirements: 1. You need to describe the main subject of the video in detail, including their appearance, actions, expressions, and the surrounding environment. 2. You should never describe any details about the camera movement or camera angles. 3. Your output should convey natural movement attributes, incorporating natural actions related to the described subject category, using simple and direct verbs as much as possible. 4. You should reference the detailed information in the video, such as character actions, clothing, backgrounds, and emphasize the details in the photo. 5. Control the output prompt to around 80-100 words. 6. No matter what language the user inputs, you must always output in English. Example of the English prompt: 1. A Japanese fresh film-style photo of a young East Asian girl with double braids sitting by the boat. The girl wears a white square collar puff sleeve dress, decorated with pleats and buttons. She has fair skin, delicate features, and slightly melancholic eyes, staring directly at the camera. Her hair falls naturally, with bangs covering part of her forehead. She rests her hands on the boat, appearing natural and relaxed. The background features a blurred outdoor scene, with hints of blue sky, mountains, and some dry plants. The photo has a vintage film texture. A medium shot of a seated portrait. 2. An anime illustration in vibrant thick painting style of a white girl with cat ears holding a folder, showing a slightly dissatisfied expression. She has long dark purple hair and red eyes, wearing a dark gray skirt and a light gray top with a white waist tie and a name tag in bold Chinese characters. The background has a light yellow indoor tone, with faint outlines of some furniture visible. A pink halo hovers above her head, in a smooth Japanese cel-shading style. A close-up shot from a slightly elevated perspective. 3. CG game concept digital art featuring a huge crocodile with its mouth wide open, with trees and thorns growing on its back. The crocodile’s skin is rough and grayish-white, resembling stone or wood texture. Its back is lush with trees, shrubs, and thorny protrusions. With its mouth agape, the crocodile reveals a pink tongue and sharp teeth. The background features a dusk sky with some distant trees, giving the overall scene a dark and cold atmosphere. A close-up from a low angle. 4. In the style of an American drama promotional poster, Walter White sits in a metal folding chair wearing a yellow protective suit, with the words ”Breaking Bad” written in sans-serif English above him, surrounded by piles of dollar bills and blue plastic storage boxes. He wears glasses, staring forward, dressed in a yellow jumpsuit, with his hands resting on his knees, exuding a calm and confident demeanor. The background shows an abandoned, dim factory with light filtering through the windows. There’s a noticeable grainy texture. A medium shot with a straight-on close-up of the character. Directly output the English text.