acceptodds
Under review as a conference paper at ICLR 2027

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

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 particu- larly challenging for one-step causal generation, where each imperfect predic- tion immediately becomes context for subsequent frames. Existing approaches improve test-time rollouts through staged training with multiple trained mod- ules and rollout-aware regularization, yet one-step refinement quality can still fall short of what render-time deployment requires. We introduce OneFixer, a one-step autoregressive video-diffusion fixer trained in one task-specific adap- tation 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 op- timized for current-frame quality are exactly those reused as future context, fi- delity and autoregressive robustness are learned jointly, without bidirectional-to- causal conversion or teacher–student distillation. Targeting driving simulation, OneFixer further leverages cues that setting readily provides—lane geometry and dynamic-agent states—to improve geometric fidelity. Across Waymo and propri- etary driving scenes with 900-frame autoregressive rollouts, OneFixer achieves the lowest FVD, LPIPS, and DISTS among all baselines under one-step infer- ence, with strong geometric fidelity and temporal consistency matching or ex- ceeding multi-stage DMD pipelines. Under identical backbone and conditioning, it matches the final quality of a multi-stage DMD-with-Self-Forcing pipeline in under half the GPU-hours and continues to improve beyond its plateau. Project page: https://onefixer-web.vercel.app/

Primary area:
generative models
Keywords:
Autonomous Driving, 3D Gaussian Splatting, Video Diffusion, Autoregressive Generation, Single-Step Generation, Long-Horizon Consistency, World Models, Driving Simulation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_jhYxOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes OneFixer for refining 3DGS renderings in driving simulation with one-step autoregressive video diffusion. It builds on the pretrained causal OmniDreams model and uses a single task-specific training stage. Self-rollout flow matching uses the model’s own predictions as history. Perceptual supervision improves image details on the same rollout. Lane and actor conditions help preserve scene geometry. Experiments on Waymo and a private Seoul dataset report better one-step video quality and competitive temporal consistency. Closed-loop tests also report fewer collisions.

Critical strengths

S1. Shared-rollout training connects image quality with temporal consistency. It trains on the model’s own predictions and directly supervises their visual quality. This addresses both poor one-step outputs and error accumulation without an extra distillation stage.

S2. The paper includes same-backbone comparisons and component ablations. These provide useful evidence for the training design. Evaluation on 900-frame videos also supports its ability to handle sequences much longer than the training clips.

Critical weaknesses

W1. The methodological contribution appears incremental.

  1. The causal backbone and geometry conditioning come from OmniDreams, while opacity mixing follows ArtiFixer. The main new element is combining self-generated history and perceptual supervision in one training stage.
  2. The paper does not compare against Self-Forcing with the same perceptual supervision. It would clarify whether the advantage comes from the shared-rollout design or mainly from the additional loss.

W2. The experiments do not fully support the claim of faithful novel-view simulation. The main tests use a fixed 1 m cycle-render setup on Waymo and logged views on Seoul. Reliability under larger viewpoint shifts remains unclear. Table 1 also lacks a raw 3DGS baseline, making it difficult to verify improvements in vehicle and lane accuracy over the input.

Questions for authors

I noticed significant background jitter even in static scenes in the supplementary video. Is this caused by poor temporal consistency?

3 comments

Response to Reviewer jhYx\textcolor{blue}{\text{jhYx}}

W1.1: Methodological novelty

We agree that the backbone/geometry conditioning come from OmniDreams, opacity mixing from ArtiFixer, and the individual training components are established.

We rewrote the abstract and added Sec. 2.4 to clarify our contribution: the informative 3DGS flow source enables deployment-matched one-step self-rollout, allowing joint FM + LPIPS on the same generated history, with perceptual gradients through k=4k=4 positions, without multi-stage distillation. This provides simpler, low-latency training for both perceptual fidelity and temporal stability.

W1.2: Attribution of improvements

We revised Sec. 4.3 to disentangle flow-source feasibility, objective effects, and rollout optimization. We also added Self-Forcing + LPIPS, V7-J, and training diagnostics (Apps. B.1–B.2).

Waymo @1FVD ↓LPIPS ↓FVMD ↓
V2: Self-rollout FM13.30.155180
V3: Teacher-forced FM + LPIPS6.90.123210
Self-Forcing12.70.159175
Self-Forcing + LPIPS (new)11.90.155171
V7-J: Joint, independent (new)6.050.127178
OneFixer: Joint, shared5.830.117158

LPIPS alone is insufficient: the full DMD+LPIPS schedule yields modest gains, and Fig. 6(c) shows that it remains behind OneFixer on both validation FVD and LPIPS despite higher cumulative GPU-hours.

More directly, V7-J → OneFixer isolates rollout sharing, improving LPIPS 0.127 → 0.117 and FVMD 178 → 158. The combined formulation achieves the best FVD and FVMD simultaneously among controlled variants, without trading perceptual fidelity for temporal stability.


W2: Larger novel views and geometric accuracy

We strengthen the experiments in two ways:

Raw 3DGS baseline (Table 1). We add unrefined inputs and annotation-referenced geometry evaluation. On Waymo, OneFixer improves 3D AP$_{50}$ from 2.46 to 6.37 and Lane F1 from 0.554 to 0.738; on Seoul, from 19.25 to 29.71 and 0.218 to 0.371, respectively.

Larger viewpoint shifts (new Fig. 7). We extend Waymo cycle rendering to 1, 2, and 4 m, using the same reference-view evaluation protocol. OneFixer achieves the best FVD, LPIPS, 3D AP$_{50}$, and Lane F1 at every tested displacement, including 4 m, although performance degrades with larger shifts for all methods.

These results strengthen evidence of robustness beyond 1 m; cycle rendering does not exhaustively evaluate arbitrary unseen viewpoints.


Q1: Background jitter in static scenes

We acknowledge residual appearance flickering, particularly when the input 3DGS is severely corrupted or incomplete. Importantly, this issue is not unique to OneFixer: competing fixers also exhibit temporal instability under these challenging conditions. To make this limitation and the differences between methods transparent, we add static_scenes.mp4 to the supplementary material, showing the input, ground truth, DiFix, One-Forcing, and OneFixer side by side for direct temporal inspection.

SceneMetric ↓InputGTDiFixOne-ForcingOneFixer
190778LPIPS0.28100.2080.2010.177
Flicker0.2530.2350.6960.3450.357
208886LPIPS0.22500.0900.1120.059
Flicker0.0840.1030.2240.1860.182

LPIPS: AlexNet vs. GT (stride 10). Flicker proxy: mean grayscale temporal second difference, ∣It−12(It−1+It+1)∣|I_t-\frac12(I_{t-1}+I_{t+1})|, on a 0–255 scale (lower is better). It also responds to camera/object motion.

OneFixer achieves the lowest LPIPS in both examples and substantially reduces DiFix's temporal variation. One-Forcing has slightly lower flicker in scene 190778, so we claim improved temporal consistency, not complete flicker elimination.

We have explicitly added this limitation and future-work direction to Sec. 5 (Conclusion and Limitations). In particular, reference-view anchoring, as explored in DiFix, ArtiFixer, and Seoul World Model, could further stabilize appearance when the input 3DGS is severely degraded.

Thanks for the response. Most of my concerns have been addressed. However, the novelty still seems limited, so I will maintain my rating of 3.

Thank you for your follow-up and for acknowledging that most of your concerns have been addressed.

While the individual components build on prior work, we believe the key contribution lies in their integration into a deployment-matched, single-stage training formulation that jointly improves perceptual fidelity and temporal consistency without distillation. Importantly, despite its simpler training procedure, OneFixer achieves stronger empirical results than multi-stage distillation-based methods, demonstrating that improved performance does not necessarily require increased training complexity. Beyond perceptual metrics, the improvements also translate into more reliable closed-loop driving simulation, as evidenced by reduced collision and off-road rates. We believe these findings highlight both the methodological and practical significance of our approach.

We hope the additional controlled experiments, particularly the shared-vs.-independent rollout ablation (V7-J vs. V9) and the DMD+LPIPS comparison, help clarify the distinct contribution and practical significance of our approach.

Thank you again for your constructive feedback, which has helped us strengthen both the evaluation and positioning of the paper.

Reviewer_VLEqOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes OneFixer, a one-step autoregressive video-diffusion fixer that refines 3DGS renderings for driving simulation. Starting from the pre-DMD causal OmniDreams checkpoint, it uses an opacity-mixed 3D GS rendering as the flow source (following ArtiFixer) and is trained in a single stage with two objectives on one "shared rollout":

  1. Flow matching conditioned on the model's own one-step predictions as causal history (self-rollout).
  2. A pixel-space LPIPS loss on one decoded position, back-propagated through the last k=4 latent positions of that history.

Lane and agent conditioning are inherited from OmniDreams. Experiments on Waymo (198-frame rollouts) and a proprietary dataset (900-frame rollouts) report the best one-step FVD, LPIPS and DISTS against image-based , bidirectional and causal fixers. The paper also includes same-backbone comparisons with DMD-based Self-Forcing and One-Forcing pipelines (including a GPU-hours analysis), a closed-loop driving-policy evaluation, and a component ablation (V1–V9).

Critical strengths
  1. Solid empirical evaluation. The comparison covers image, bidirectional and causal fixers. The same-backbone baselines share initialization and conditioning, per-clip metrics come with 95% intervals, and the closed-loop evaluation with a driving policy goes beyond pixel metrics. Together these make the reported improvements credible as an empirical result.
  2. Simple and efficient pipeline. A single training stage, with no teacher, distillation or critic, and one-step inference make the approach practical for render-time use.
Critical weaknesses
  1. Limited technical novelty. The method, described in about one page (Eq. 3–6), combines two known ideas:

    • Training on self-generated history to reduce exposure bias is the core idea of Self-Forcing [1]. With a one-step generator, the training rollout trivially coincides with inference, which is a natural simplifi cation rather than a new insight.
    • Pixel-space perceptual losses through the VAE decoder are well established for one-step diffusion models, e.g., Difix3D+ [2], DOVE [3] and ChopGrad [4], as the paper's own Sec. 2.3 acknowledges.

    The remaining difference is that the LPIPS gradient also flows into the last k latent positions of the autoregressive history. This is a small extension, and the paper offers little analysis of why the coupling works beyond the ablation.

  2. The attribution of gains to shared-rollout training is not supported. Sec. 4.2 states that, since backbone, conditioning and initialization are fixed, the difference to the DMD baselines "is attributable to shared-rollout training". However, the Self-Forcing and One-Forcing baselines are trained without any pixel loss. Variant V3 (ground-truth history + LPIPS, no self-rollout) already outperforms both on FVD and LPI PS (6.9 / 0.123 vs. 12.7 / 0.159 and 11.4 / 0.153; Table 2a vs. Table 1). Much of the perceptual gain therefore appears to come from the well-known LPIPS loss, which could equally be added to the baselines. The pa rt attributable to shared rollout (V3 → V9) is mainly on FVMD and is much smaller than the headline comparison suggests.

Minor (presentation): Text in Fig. 1(a), 2 and 6 is too small to read at normal zoom. All tables are on p. 9, far from where they are first referenced (Table 2a is cited in Sec. 3.1 on p. 4). The abstract is den se with jargon ("deployment-matched shared rollout") and does not state the motivation and insight clearly.

References

[1] X. Huang, Z. Li, G. He, M. Zhou, E. Shechtman. Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion. NeurIPS 2025.

[2] J. Z. Wu et al. Difix3D+: Improving 3D Reconstructions with Single-Step Diffusion Models. CVPR 2025. arXiv:2503.01774.

[3] Z. Chen et al. DOVE: Efficient One-Step Diffusion Model for Real-World Video Super-Resolution. NeurIPS 2025. arXiv:2505.16239.

[4] D. Rivkin et al. ChopGrad: Pixel-wise Losses for Latent Video Diffusion via Truncated Backpropagation. arXiv:2603.17812, 2026.

Questions for authors
  1. To attribute the gains fairly, could you add the same LPIPS loss to the Self-Forcing (DMD) baseline and compare it with OneFixer under the same budget?
  2. Compared with Self-Forcing-style training on self-generated history and existing pixel-space losses for one-step models (e.g., Difix3D+, DOVE), what is the key new insight of the method, and which experiment is olates its contribution?
2 comments

Response to Reviewer VLEq\textcolor{green}{\text{VLEq}}

W1 / Q2: Methodological novelty

We rewrote the abstract and added Sec. 2.4 (Positioning of OneFixer). We acknowledge that the backbone/conditioning (OmniDreams), opacity mixing (ArtiFixer), self-generated histories, and perceptual supervision are individually established.

Our contribution exploits the structured 3DGS flow source to jointly optimize FM and LPIPS on the same deployment-matched, one-step autoregressive rollout, propagating perceptual gradients through recent history (k=4k=4), in one critic-free adaptation stage.

MethodPixelARTraining historySingle post-training stage
DiFix3D+ [1]✓× (image)N/A✓
DOVE [2]✓× (bidirectional)N/A× (latent → pixel)
OASIS [3]✓× (bidirectional)N/A✓ (curriculum)
Resampling Forcing [4]×✓Re-denoised GT proxy✓ (from scratch)
Self-Forcing fixer, ArtiFixer [5,6]×✓Self-generated; truncated denoising gradients× (teacher → causal → DMD)
FlashVSR [7]✓✓Parallel causal× (SR → causal → DMD)
OneFixer✓✓One-step self-generated; LPIPS without denoising-chain backpropagation✓

Single-stage excludes separate objective-switching/distillation stages; a degradation curriculum counts as one. ChopGrad [8] enables efficient decoder backpropagation.


W2: Attribution of improvements

We reorganized Sec. 4.3 into dedicated paragraphs disentangling three factors, supported by controlled ablations and training diagnostics:

  1. Flow source — feasibility. The inherited opacity-mixed source makes one-step refinement practical: replacing it with noise and concatenated 3DGS (V6) increases FVD from 5.8 to 104.8. We do not claim the source itself as novel.

  2. Training objectives — fidelity. As the reviewer points out, LPIPS drives much of the perceptual improvement. Teacher-forced FM + LPIPS (V3) reaches FVD 6.9 / LPIPS 0.123, but not comparable temporal stability (FVMD 210 vs. 158).

  3. Rollout optimization — fidelity and stability. The new V7-J matches joint objectives and schedule using independent rollouts. Sharing them (V9) improves FVD (6.05 → 5.83), LPIPS (0.127 → 0.117), and FVMD (178 → 158). Sequential FM → LPIPS (V7) performs worse on FVD/FVMD (7.7/193). Shared rollouts also show lower gradient variance (7.5 vs. 8.8, App. B.1).

Thus, we do not attribute the headline gains solely to rollout sharing. Complementary objectives and deployment-matched shared optimization jointly achieve the best FVD and FVMD simultaneously among controlled ablations.


W2 / Q1: Does LPIPS close the DMD gap?

We add Self-Forcing + LPIPS to the full Stage-3 DMD run, using the same VGG supervision, causal Wan decoder, ChopGrad [8], and k=4k=4 AR gradient horizon (App. C.2). We retain the entire DMD schedule, giving this baseline more training compute than OneFixer.

Waymo @1FVD ↓LPIPS ↓FVMD ↓
Self-Forcing12.70.159175
Self-Forcing + LPIPS (new)11.90.155171
OneFixer5.830.117158

Adding LPIPS alone does not close the gap (Seoul FVD: 27.2 → 26.8, vs. 7.13 for OneFixer).

More decisively, we updated Fig. 6(c) to directly compare validation FVD and LPIPS against cumulative GPU-hours. DMD+LPIPS improves perceptual quality during Stage 3 but remains behind OneFixer on both metrics despite substantially greater training compute.

We also add gradient-dynamics and alignment analysis (Fig. 9, App. B.2). The LPIPS-to-DMD gradient-norm ratio changes by 7–10× as the critic adapts. DMD gradients remain nearly orthogonal to LPIPS, whereas endpoint FM is more LPIPS-aligned (cos⁡≈0.05\cos\approx0.05–0.070.07). These observations are consistent with DMD2 [9], which distinguishes regression and distribution matching, and Adaptive Video Distillation [10], which motivates careful balancing of regression within DMD training. They suggest an optimization explanation, not causal proof.

OneFixer avoids critic-dependent tuning by jointly optimizing FM and LPIPS in a single-network, single-stage procedure.


Minor: Presentation

We have redrawn Figs. 1(a), 2, and 6 with larger text and clearer labels, and repositioned Tables 1 and 2 from p. 9 to p. 7, closer to their first references..


References

[1] Wu et al., DiFix3D+ (2025); [2] Chen et al., DOVE (2025); [3] Guo et al., OASIS (2025); [4] Guo et al., Self-Resampling (2026); [5] Huang et al., Self-Forcing (2025); [6] de Lutio et al., ArtiFixer (2026); [7] Zhuang et al., FlashVSR (2025); [8] Rivkin et al., ChopGrad (2026); [9] Yin et al., DMD2 (2024); [10] You et al., Adaptive Video Distillation (2026).

Dear Reviewer VLEq, Thank you again for your thoughtful and valuable review! As the discussion period approaches its end, we would appreciate your feedback on our response. To respond your concerns, we performed new experiments and reorganize paper's developments for better attribution.

Could you let us know whether this addresses your concern? We would be happy to clarify.

Again, thank you for your time

Reviewer_ojzsOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses one-step autoregressive refinement of degraded 3DGS renderings. It jointly trains flow matching and pixel-space perceptual supervision using deployment-matched self-generated rollouts, with truncated gradients through recent history.

Critical strengths

1,The shared-rollout formulation is well motivated and directly addresses the interaction between current-frame fidelity and future-context quality.

2,Extensive baselines, same-backbone comparisons, and component ablations provide substantial empirical support.

3,Long rollouts, conditioning perturbations, memory analysis, and closed-loop evaluation make the study practically relevant.

Critical weaknesses

1,The sequential baseline changes both optimization scheduling and exposure to each objective. It does not distinguish shared-rollout coupling from joint optimization alone.

2, The reported per-frame latency is amortized over blocks producing eight RGB frames. This does not establish the delay before a policy receives a newly conditioned observation.

3,Geometry evaluation relies on indirect proxies. Cuboid F1 measures projected-mask overlap rather than 3D box accuracy, while lane metrics compare detector outputs. These may overlook safety-relevant geometric errors.

Questions for authors

1,Could the authors compare shared-rollout training against joint training with separate rollouts, matching objective evaluations and compute?

2,Explain how eight-frame generation interacts with policy feedback and updated camera poses during closed-loop simulation.

3,Please provide detector-based geometry metrics with annotation-referenced measurements, particularly actor position, missed/hallucinated vehicles, and lane displacement. A long-range tracking or identity-consistency analysis would also strengthen the 900-frame consistency claim.

2 comments

Response to Reviewer ojzs\textcolor{red}{\text{ojzs}}

W1 / Q1: Joint optimization vs. shared rollout

We add V7-J, jointly optimizing FM and LPIPS on independent autoregressive rollouts, with matched training examples, objectives, schedule, and perceptual horizon (k=4k=4) (Table 2a, Sec. 4.3).

Waymo ablationFVD ↓LPIPS ↓FVMD ↓
V7: Sequential FM → LPIPS7.700.119193
V7-J: Joint, independent (new)6.050.127178
V9: Joint, shared (OneFixer)5.830.117158

Joint optimization explains much of the FVD improvement; sharing the rollout further improves FVD, LPIPS, and FVMD. Shared rollouts also reduce median gradient variance (7.5 vs. 8.8, App. B.1), suggesting an optimization benefit without establishing causality.


W2 / Q2: Closed-loop latency and feedback

We clarify the protocol in Sec. 4.2 and Table 2b. Video models generate four frames per inference block, with VLA replanning every four simulation frames. MPC executes the current plan between replans, and subsequent observations use control-realized camera poses.

To isolate the effect of rendering quality on policy behavior, the simulator advances synchronously, so inference latency affects throughput, not closed-loop scores. With the shared FlashDreams acceleration stack, OneFixer generates one VAE block (four RGB frames) in approximately 62 ms (H200, bf16)—faster in wall-clock time than the native DiFix implementation's 212 ms for a single RGB frame. These timings use different inference stacks; Fig. 6(a) separately reports unaccelerated per-frame latency for a comparable benchmark.

We also add DiFix with per-frame replanning (RP=1):

RendererReplan intervalCollision ↓
DiFix4 frames13.0%
DiFix (new)1 frame11.7%
OneFixer4 frames6.9%

Despite more frequent feedback, DiFix does not close the gap to OneFixer.


W3 / Q3: Annotation-referenced geometric evaluation

We add 3D AP$_{50},center−distancemAP,annotation−basedLaneF1,andHOTA, center-distance mAP, annotation-based Lane F1, and HOTA_d$ (Table 1, App. C.1), replacing detector-agreement proxies in the main comparison. Actor/lane metrics use annotations; HOTA$_d$ measures detection and identity association without tracking resets.

All methods use identical frozen evaluators over 80 Waymo scenes (198 frames) and 32 Seoul scenes (900 frames). We also include raw 3DGS.

Seoul (900 frames)3D AP$_{50}$ ↑mAP$_{\mathrm{dist}}$ ↑Lane F1 ↑HOTA$_d$ ↑
Raw 3DGS (new)19.2525.500.21826.56
Best competing baseline25.4629.310.35129.63
OneFixer29.7132.010.37131.19

OneFixer also achieves the highest point estimates on all four metrics on Waymo, strengthening evidence of geometric fidelity and temporal identity preservation.

Dear Reviewer ojzs, Thank you again for your thoughtful review! As the discussion period approaches its end, we would appreciate your feedback on our response. We clarified the closed loop evaluation, and added annotation referenced geometric evaluations. Also, we added experiments for clarify the attribution of shared rollout. Regards.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Overall Author Response

We thank reviewers ojzs\textcolor{red}{\text{ojzs}}, VLEq\textcolor{green}{\text{VLEq}}, and jhYx\textcolor{blue}{\text{jhYx}} for their constructive feedback. We have carefully refined the manuscript to address their concerns (**revision locations are indicated in parentheses below; highlighted in blue\textcolor{blue}{\text{highlighted in blue}} in the manuscript):

  1. Contribution and positioning [VLEq\textcolor{green}{\text{VLEq}}, jhYx\textcolor{blue}{\text{jhYx}}] (Abstract, Sec. 2.4). We simplified the abstract to clarify our central contribution and added a dedicated "Positioning of OneFixer" section to distinguish our formulation from prior work. The key distinction is jointly optimizing FM and LPIPS on deployment-matched, one-step self-rollouts, enabling single-stage adaptation without distillation.

  2. Attribution [ojzs\textcolor{red}{\text{ojzs}}, VLEq\textcolor{green}{\text{VLEq}}, jhYx\textcolor{blue}{\text{jhYx}}] (Sec. 4.3, App. B.1). We reorganized the analysis into dedicated paragraphs addressing flow-source feasibility, complementary objectives and causal history, and objective scheduling and rollout sharing. The new controlled V7-J ablation isolates sharing under joint FM+LPIPS, improving LPIPS (0.127 → 0.117) and FVMD (178 → 158), with lower gradient variance.

  3. DMD + perceptual supervision [VLEq\textcolor{green}{\text{VLEq}}, jhYx\textcolor{blue}{\text{jhYx}}] (Table 1, Figs. 6(c), 9, App. B.2). We add Self-Forcing + LPIPS under the full DMD schedule. The updated GPU-hours/performance comparison shows that OneFixer achieves lower validation FVD and LPIPS with less compute. Gradient-dynamics analysis further clarifies the optimization challenges of DMD+LPIPS in connection with prior work.

  4. Geometric fidelity and novel-view robustness [ojzs\textcolor{red}{\text{ojzs}}, jhYx\textcolor{blue}{\text{jhYx}}] (Table 1, Fig. 7, App. C.1). We add raw 3DGS, annotation-referenced actor/lane metrics, HOTA_d (adapted from HOTA; Luiten et al., 2021), and 1–4 m viewpoint-shift evaluations. OneFixer achieves the best FVD, LPIPS, 3D AP$_{50}$, and Lane F1 across the tested displacements.

  5. Closed-loop evaluation [ojzs\textcolor{red}{\text{ojzs}}] (Sec. 4.2, Table 2b). We clarify the four-frame replanning protocol, synchronous simulation, and accelerated inference (~62 ms per four-frame block). To account for DiFix's frame-independent refinement capability, we also evaluate DiFix with per-frame replanning, which still yields more collisions than OneFixer (11.7% vs. 6.9%).

(We provide detailed responses below.)

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.