cs.CVSep 26, 2026

OneFixer: High-Quality and Consistent One-Step Autoregressive 3DGS Refinement for Driving Scenes

Authors: Boseong Jeon, Junhyeop Lee, Juhan Cha, Hayoung Kim

Organizations: 142dot · Hyundai Motor Company

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Autoregressive video diffusion is a promising render-time fixer for 3D Gaussian Splatting (3DGS) in autonomous-driving simulation, but deployment demands high visual quality and temporal consistency at low latency. This is especially hard for one-step causal generation, where each imperfect prediction immediately becomes context for subsequent frames. Existing approaches stabilize rollouts through staged training with multiple modules and rollout-aware regularization, yet one-step quality still falls short of what deployment requires. We introduce OneFixer, a one-step autoregressive video-diffusion fixer trained in a single task-specific adaptation stage. Our key idea is a deployment-matched shared rollout: the model's own one-step predictions serve as the causal context for flow matching, exposing training to deployment-time errors, while the same rollout receives direct pixel-space perceptual supervision to preserve fine detail. Because the predictions optimized for current-frame quality are exactly those reused as future context, fidelity and autoregressive robustness are learned jointly, without bidirectional-to-causal conversion or teacher-student distillation. OneFixer further exploits cues that driving simulation readily provides, lane geometry and dynamic-agent states, to improve geometric fidelity. On Waymo and proprietary driving scenes with 900-frame rollouts, OneFixer achieves the lowest FVD, LPIPS, and DISTS among all baselines at one step, with temporal consistency matching or exceeding multi-stage DMD pipelines. Under identical backbone and conditioning, it matches a multi-stage DMD-with-Self-Forcing pipeline in under half the GPU-hours and keeps improving beyond its plateau. In closed-loop simulation with a driving policy, OneFixer reduces the collision rate by a third relative to raw 3DGS rendering. Project page: https://onefixer-web.vercel.app/

Figures & tables

Figure 1: Quality–consistency–latency trade-off of 3DGS refinement. (a) Methods are positioned by video fidelity (I3D FVD, ↓\downarrow ) and temporal consistency (FVMD, ↓\downarrow ), with bubble area proportional to inference latency; lower-left and smaller are better. Dashed lines connect a method across its step-reduction or distillation path (e.g., OmniDreams @4 →\rightarrow @1). (b) Qualitative comparison on a degraded Waymo 3DGS rendering; OneFixer recovers finer foliage and a more faithful vehicle under one-step causal refinement.
Figure 2: Shared-rollout training for one-step 3DGS refinement. From opacity-mixed 3DGS inputs and scene conditioning, the deployment-time one-step causal generator rolls out its own predictions {z^01,…,z^0N}\{\hat{\mathbf{z}}_{0}^{1},\ldots,\hat{\mathbf{z}}_{0}^{N}\} exactly as at inference. The rollout serves two objectives: self-rollout flow matching supervises every position in latent space conditioned on the self-generated history (detached, sg\mathrm{sg} ), and a frame-wise pixel-space perceptual loss on one sampled prediction back-propagates through the last kk latent positions of that history to recover fine detail.
Figure 3: Complementary effects of self-rollout and perceptual supervision. Left: a Waymo example and its degraded 3DGS input. Right: one-step autoregressive rollouts and magnified regions. Diffusion Forcing exhibits appearance variation across generated frames; self-rollout flow matching improves temporal consistency, while direct pixel-space perceptual supervision further recovers fine visual details, particularly in the tree branches and foliage.
Figure 4: Qualitative comparison of 3DGS refinement methods. OneFixer better preserves realistic appearance and scene structure under challenging reconstruction errors. Its geometry-aware conditioning improves structural fidelity, while the video prior increases robustness to transient artifacts that are difficult to handle independently frame by frame, such as the popping Gaussian floater in the bottom row.
Figure 5: Closed-loop replay with different renderers. Top: policy observation and planned path (blue) where the 3DGS reconstruction degrades. Bottom: ego trajectory over the lane map. Raw 3DGS shows ghosted structure that bends the trajectory; DiFix smears the lane markings across the road and the policy leaves it; OmniDreams blurs the scene at one step. OneFixer’s faithful observation keeps the policy on route.
Figure 6: Inference efficiency, long-horizon stability, and training efficiency. (a) Per-frame enhancement latency measured on a single H200 GPU and bf16 precision for all models. @N denotes the number of denoising steps, and model parameter counts (in billions) are shown in parentheses. (b) Position-wise FVD over 900-frame autoregressive rollouts on the proprietary dataset. Self-Forcing is shown in steel blue; other colors match (a). (c) Validation FVD and FVMD vs cumulative GPU-hours, 32 Waymo scenes, 93-frame one-step rollouts, same backbone and conditioning. Baselines follow the standard fixer-distillation recipe ( de Lutio et al., 2026 ; Zhou et al., 2026 ) : teacher adaptation, causal fixer adaptation, then DMD with Self-Forcing (green) or One-Forcing (blue).
Table 2: Ablation and closed-loop evaluation. (a) SF: causal history for flow matching — ×\times ground truth, △\triangle re-denoised ground truth (Sec. 4.3 ), ∘\circ own rollout. ∘k\circ_{k} : LPIPS with backpropagation horizon kk ( k=4k{=}4 default). Map: lane and actor conditioning. Mix: opacity-mixed source; ×\times uses noise with channel-concatenated 3DGS latents. ++ : joint; →\rightarrow : 250 SF then 150 LPIPS-only steps. (b) Closed-loop replay on 60 proprietary scenes ×\times 5 policy seeds, varying only the renderer; mean and 95% interval over scene–seed runs. Collision is the at-fault rate (front or lateral); route distance is the mean deviation from the logged trajectory.
Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

QuantityTrainValidation
Scenes (four modalities present)35080
with dynamic actors24152
Median frames per scene198198
Total aligned frame pairs69,23815,812
Appendix
Table 3: Cycle-render paired corpus statistics (front camera).
Figure 7: Robustness to geometry-conditioning errors. Evaluation uses a fixed random subset of 32 Waymo scenes. From left to right: Lane F1, lane lateral error, Cuboid F1, and I3D FVD, grouped by absolute lane displacement or planar object translation magnitude. Injected displacements are measured in meters. Boxplots summarize variation across perturbation configurations within each group. Black dashed lines indicate unconditioned OneFixer (black sdmap input), purple dotted lines indicate ArtiFixer DMD @1 on the same subset, and purple dots mark the fully unperturbed reference where shown. Lane F1 and Cuboid F1 remain above the unconditioned reference, although the largest lane displacement produces higher lateral error.
Figure 8: Peak training memory versus parallel context and perceptual gradient window. Single-GPU peak allocated memory (GiB) at 544×960544\times 960 with one-step generation, a full rollout, and the perceptual loss on the final latent position, as a function of the number of latent positions processed in parallel (rows) and the perceptual backpropagation window kk in latent steps (columns). Grey cells exceed device capacity; blank cells have kk larger than the context. The default configuration (24 positions, k=4k=4 ) is 102 GiB. Multi-GPU training with sharded parameters shows comparable per-device peaks.
Figure 9: Exposure bias in a frame-recurrent image fixer. Harmonizer in temporal mode on a Waymo sequence, conditioning on its own four previous outputs at inference. Top: trained with the official teacher-forced recipe (ground-truth prior frames as context, 50% context dropout); appearance drifts within a few dozen frames and the model begins copying its own artifacts. Bottom: the same model trained with self-generated context; the rollout remains stable. This is the variant reported as Harmonizer (temporal) in Table 1 . Frame index and elapsed time are shown in each panel.
Figure 10: Qualitative generalization on DL3DV. OneFixer readily transfers beyond driving scenes after fine-tuning on DL3DV, while DiFix and ArtiFixer are evaluated using their official DL3DV-finetuned checkpoints. Under one-step inference, OneFixer matches or exceeds the frame-level visual quality of the image-based DiFix while providing stronger temporal consistency through video diffusion and produces higher-fidelity results than one-step ArtiFixer.
 TemporalPerception
ModelWarp L1 ↓\downarrowFVMD ↓\downarrowLPIPS-Mid ↓\downarrowLPIPS-Last ↓\downarrowPSNR ↑\uparrow
DiFix @10.030667440.1400.16421.5
ArtiFixer DMD @40.017930100.1480.16422.2
ArtiFixer DMD @10.018134110.1760.19322.3
OneFixer (Ours) @10.020123260.1260.13622.6
Appendix
Table 4: Generalization to non-driving scenes on DL3DV. Denoising steps are denoted by @N. Best and second-best results are shown in bold and underlined , respectively. FVMD is reported unscaled here.
Figure 11: Appearance-conditioned 3DGS refinement. Given the same degraded 3DGS rendering, we compare prompt-conditioned generation for night, fog, and snow. General-purpose video editing models such as Cosmos3 and Qwen-Video-Edit can produce visually convincing target styles, but often substantially alter the underlying scene content and geometry inherited from the 3DGS input. In contrast, after paired prompt-conditioned training, OneFixer jointly performs 3DGS refinement and appearance transformation in a single pass, removing reconstruction artifacts while better preserving the underlying scene structure.
ModelStyleGeometry
 CLIP [-2pt]acc ↑\uparrowVLM [-2pt]acc ↑\uparrowLane [-2pt]F1 ↑\uparrowLane [-2pt] xx -err ↓\downarrow
Cosmos30.9620.9920.0980.632
Qwen-Video-Edit1.0001.0000.3330.627
OneFixer0.9180.9560.5900.219
Appendix
Table 5: Style conditioning.
Figure 12: Multi-camera 3DGS refinement. From top to bottom: degraded multi-camera 3DGS renderings, OneFixer outputs, LiDAR points from neighboring cameras projected onto the refined views, and bird’s-eye-view compositions from the ground truth and OneFixer outputs. OneFixer jointly removes reconstruction artifacts across views while preserving cross-view scene structure. The agreement of projected LiDAR points with the refined images and the resulting BEV composition provides qualitative evidence that refinement remains consistent with the shared underlying scene geometry.
Figure 13: High-resolution 3DGS refinement. Top: degraded 3DGS input and clean reference. Bottom: OneFixer outputs at 960- and 1920-pixel widths, with enlarged regions highlighting distant scene details. Scaling to higher resolution preserves the overall refinement behavior while improving the recovery of fine structures such as road signs and other small visual details. High-resolution training is enabled by shortening the self-rollout segment during training to reduce memory consumption.

Explore similar work

CardsList
Aug 18, 2026

SPVC: Structured and Panoptic Video Fixing for Cross-Dataset Driving Scene Rendering

Driving scene reconstruction and rendering, especially with 3D Gaussian Splatting, has become an important component of autonomous driving simulation. However, rendered views often degrade under extrapolated ego trajectories and scene edits, producing blurry structures, temporal flicker, and foreground-background misalignment. Existing refinement methods are commonly designed for a specific setting, such as image-level novel-view repair or object-editing correction. In this paper, we introduce SPVC, a structured and panoptic video fixing framework for cross-dataset driving scene rendering. The name summarizes four design principles. (1) Structured fixing denotes the use of explicit spatial conditions, including camera pose, 3D bounding boxes, and HD maps, to guide the repair process and reduce uncontrolled hallucination. (2) Panoptic fixing refers to correcting both background rendering artifacts, such as distorted roads, buildings, and lanes, and foreground vehicle artifacts introduced by scene editing, such as inconsistent object appearance. (3) Video fixing means that the model operates on driving sequences rather than isolated frames, allowing temporal cues to be used during artifact correction. (4) Cross-dataset fixing means that a single shared network is trained and applied across multiple driving datasets, reducing the need for dataset-specific or scene-specific fixers. Concretely, we construct paired degraded-clean training data by simulating under-constrained 3DGS rendering and foreground vehicle insertion artifacts, and train a two-stage controllable video diffusion model that first addresses video-level appearance and then refines scene layout with structured controls.

Gen Li, Shu Han, Yun Xi Qiao +5

Institute for AI Industry Research (AIR), Tsinghua University, Beijing, China. · Zhejiang University, Hangzhou, China. · University of Wisconsin–Madison, Madison, WI, USA. +4

Autonomous Driving SimulationVideo Restoration3D Gaussian SplattingVideo Diffusion ModelsTemporal Consistency in Video Generation
AskPDF
May 10, 2026

ConFixGS: Learning to Fix Feedforward 3D Gaussian Splatting with Confidence-Aware Diffusion Priors in Driving Scenes

Feedforward 3D Gaussian Splatting (3DGS) often struggles in trajectory-based sparse-view driving scenes. Existing Gaussian repair methods mainly target optimization-based 3DGS, while diffusion-based repair is typically restricted to iterative refinement near observed viewpoints, leaving feedforward 3DGS repair underexplored. We propose ConFixGS, a plug-and-play method that learns to fix feedforward 3DGS with confidence-aware diffusion priors. Starting from a pretrained feedforward model, ConFixGS generates diffusion-enhanced local pseudo-targets and validates them through reprojection-based cross-checking against support views. The resulting dense confidence maps guide refinement, enhancing reliable details while suppressing hallucinated or inconsistent evidence. On Waymo, nuScenes, and KITTI, ConFixGS improves challenging novel view synthesis, with PSNR gains of up to 3.68 dB and FID reduced by nearly half. Our results highlight confidence-aware fusion of generative priors and support-view consistency as a key principle for robust feedforward 3D driving scene reconstruction.

Rui Song, Tianhui Cai, Markus Gross +5

University of California, Los Angeles · Technical University of Munich · University of Cambridge

3D Gaussian SplattingDiffusion GuidanceSparse-View 3D ReconstructionFeed-Forward 3D Reconstruction
AskPDF
Jun 15, 2026

RealityBridge: Bridging Editable 3D Gaussian Splatting Driving Simulations and Real-World Videos

Long-tail hazardous scenarios are essential for safety-oriented autonomous driving, yet they are difficult to collect and reproduce at scale. Editable 3D Gaussian Splatting (3DGS) simulation offers a promising alternative by reconstructing real driving scenes and supporting controllable scene editing. However, edited 3DGS-rendered videos still suffer from a significant Sim-to-Real gap, including rendering artifacts, degraded foreground assets, inconsistent illumination, and temporal flickering. Existing restoration and video generation methods are insufficient for this task, as they often fail to jointly repair 3DGS-specific artifacts, improve visual realism, and ensure temporal consistency. To fill this gap, we propose RealityBridge, a structure-preserving and asset-aware Sim-to-Real framework for edited 3DGS driving videos. RealityBridge uses multimodal controls, including rendered videos, foreground masks, edge maps, and semantic masks, together with a lightweight GateNet for adaptive condition allocation across backbone layers. We further construct targeted training data and introduce autoregressive long-video training with reward-guided post-training to improve restoration quality, temporal stability, and hallucination suppression. Extensive experiments on internal and public driving datasets show that RealityBridge outperforms existing methods in artifact removal, illumination harmonization, and long-sequence temporal consistency.

Zhenhua Wu, Yun Pang, Mingkun Chang +4

1Sun Yat-sen University · 2Shanghai Innovation Institute · 3Shenzhen Loop Area Institute

3D Gaussian SplattingSim-to-Real TransferAutonomous Driving SimulationVideo RestorationTemporal Consistency in Video Generation
AskPDF