acceptodds
Under review as a conference paper at ICLR 2027

V-REG: Frame-Aligned Visual Representation Guidance for Training-Efficient Multi-View Driving Video Generation

Abstract

Multi-view driving video generation requires modeling complex temporal, cross-view, and multimodal scene structures, leading to expensive task-specific optimization. We propose V-REG, a pretrained visual representation-guided framework that accelerates the adaptation of a multi-view diffusion transformer while largely preserving its original generation pathway. V-REG introduces frame-view aligned representation entanglement and intermediate feature alignment, associating each pretrained representation with the corresponding frame and camera view. Frozen DINOv3 features serve as both conditioning signals and alignment targets during training; at inference, only one-shot condition encoding is retained, and a KV cache mechanism is employed to accelerate multi-frame autoregressive generation by reusing cached attention states from previous frames. Experiments on nuScenes show that V-REG achieves 8.98 FID and 98.75 FVD after 30K task-specific optimization steps. At the matched FID of 11.3, it requires approximately 4.5 times fewer optimization iterations than the corresponding baseline.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Driving Video Generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_LfxsOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This work focuses on multi-view driving video generation. The authors propose a method named V-REG that exploits a pretrained foundation model (DINOv3) to accelerate the optimization of a multi-view video diffusion transformer for driving scene video generation. To facilitate feature alignment between the pretrained foundation model and the diffusion transformer, the author propose frame-view aligned representation entanglement and alignment. Experiments on the nuScenes dataset show the effectiveness of the proposed method.

Critical strengths
  1. The proposed idea is technically sound and easy to follow.
Critical weaknesses
  1. The contribution of the proposed method is not significant. The idea of accelerating the optimization of diffusion transformers with supervision from foundation models has been widely explored in the literature. REPA is one of the pioneering works that proposes representation alignment to accelerate diffusion transformer training. Subsequent works, such as iREPA, VideoREPA, and GeometryForcing, have introduced various strategies for representation alignments. The proposed frame-view alignment is a simple variant of existing techniques rather than a technically new method.

[1] REPRESENTATION ALIGNMENT FOR GENERATION: TRAINING DIFFUSION TRANSFORMERS IS EASIER THAN YOU THINK. Yu et al. ICLR 2025.

[2] WHAT MATTERS FOR REPRESENTATION ALIGNMENT: GLOBAL INFORMATION OR SPATIAL STRUCTURE?. Singh et al. ICLR 2025.

[3] VideoREPA: Learning Physics for Video Generation through Relational Alignment with Foundation Models. Zhang et al. NeurIPS 2025.

[4] GEOMETRY FORCING: MARRYING VIDEO DIFFUSION AND 3D REPRESENTATION FOR CONSISTENT WORLD MODELING. Wu et al. NeurIPS 2025.

  1. Many state-of-the-art driving video generation methods are overlooked. For example, [5,6,7] achieve significantly better performance than the proposed method on the nuScenes dataset, but these methods are not compared or discussed in this work.

[5] MaskGWM: A Generalizable Driving World Model with Video Mask Reconstruction. Ni et al. CVPR 2025.

[6] UniMLVG: Unified Framework for Multi-view Long Video Generation with Comprehensive Control Capabilities for Autonomous Driving. Chen et al. ICCV 2025.

[7] MoVieDrive: Urban Scene Synthesis with Multi-Modal Multi-View Video Diffusion Transformer. Wu et al. CVPRF 2026.

  1. The statement about the existing multi-modal DiT paradigm is overly narrow. There are different paradigms to train multi-modal multi-view video diffusion transformers. For example, [7] proposes a unified multi-modal multi-view video transformer for driving video generation without auxiliary control DiT or auxiliary alignment.

  2. The claim “training-efficient” is not well supported by the experiments. Throughout the paper, the authors use the 30K step optimization as the evident to demonstrate the training efficiency. However, the proposed requires an additional foundation model, DINOv3, to process the inputs. Although this additional foundation model is frozen, this will still significantly increase the training cost. Thus, training step is not a good metric to measure training efficiency. Instead, a common practice is training the compared methods on the same machine to measure the GPU hours required for model training.

  3. Experiments need improvement. In this work, the authors mainly report results on video, depth, and semantic generation. However, in the area of driving video generation, additional experiments are usually required, such as scene controllability experiments (measured by mAP and mIoU) and data augmentation experiments (use synthetic data to train a perception model, measured by mAP and NDS), to evaluate the quality of the generated video data.

  4. Writing needs improvement, especially the Related Work section. Many state-of-the-art methods are not discussed in this section.

Questions for authors
  1. What is technically new in the proposed method? What is the key contribution of this work to the community?
  2. Please address the critical weaknesses.
4 comments

We thank the reviewer for the thoughtful feedback. We address the remaining concerns bellow.

W1. Novelty and comparison with prior alignment methods

Good question! V-REG adapts representation alignment to controllable multi-view driving video generation. It uses the same timestep–camera correspondence to inject depth and semantic conditions and to supervise intermediate DiT features. This ties the training target for each view to the geometry and semantics used to condition it.

On PandaSet, we compare multi-view adaptations of the cited methods using the same backbone, initialization, input conditions, training protocol, and KV-cache setting.

MethodParams (B)Training (h) ↓FID ↓FVD ↓
Baseline11.006.038.40152.83
+ REPA11.2211.031.14138.32
+ iREPA11.2311.035.96127.65
+ VideoREPA11.215.028.99113.96
+ Geometry Forcing11.225.529.42127.17
+ V-REG (ours)11.231.627.5689.79

Relative to VideoREPA, the strongest comparator on FID and FVD, V-REG reduces FVD by 21.2% and recorded training time by 68.0%. With the auxiliary branches fixed and caching disabled, enabling alignment also improves FID from 37.17 to 29.57 and FVD from 140.71 to 129.79 (Table 4, ID 4 → ID 6).

W2. Comparison with recent driving video generators

Thanks. MaskGWM’s public release provides partial inference code, and its repository issue discussion explains that the original OpenSora training pipeline is unavailable. MoVieDrive’s code was unavailable to us. We therefore focus our additional controlled comparison on UniMLVG, using PandaSet due to GPU memory constraints.

MethodParameters (B)Training time (h) ↓FID ↓FVD ↓
UniMLVG8.1012.030.46113.84
V-REG (ours)11.231.627.5689.79

Under the matched PandaSet evaluation, V-REG achieves lower FID (27.56 vs. 30.46) and lower FVD (89.79 vs. 113.84) with 1.6 hours of recorded training time compared with 12.0 hours for UniMLVG.

W3. Scope of the multimodal DiT paradigm

Thanks. We believe there exists a misunderstanding about the multimodal DiT paradigm. Our Fig. 1 illustrates multimodal conditioning at the architecture level. MoVieDrive [7] falls within this broader category and illustrates block-level conditioning inside a unified DiT, adding 3D occupancy embeddings to denoising features (Fig. 3 and Eq. 4 in [7]). These two paradigms share the same sprits.

W4. Training efficiency in GPU-hours

Thanks. The 30K steps in Appendix Fig. 9 specify the optimization budget. We supplement them with reported wall-clock time and GPU-hours.

MethodTraining (days)GPU-hours
Vista1826,496
DrivingWorld1218,432
Epona1416,128
MagicDrive-V24534,901
MaskGWM32,304
V-REG3432

These reuslts indicate that our approach can achive better results with lower resources.

W5. Scene controllability and data augmentation

Thanks. We conduct two additional experiments on PandaSet.

Scene controllability. We evaluate how well the generated videos follow the requested scene layout using detection mAP and segmentation mIoU.

MethodmAP ↑mIoU ↑
Baseline10.7019.80
V-REG (ours)17.3731.50

V-REG improves mAP by 6.67 points and mIoU by 11.70 points over the baseline, indicating closer agreement between the generated videos and the scene controls.

Data augmentation. We train a perception model using real data alone or real data augmented with 1,000 V-REG-generated samples. Both models are evaluated on the same real PandaSet validation set. We report detection mAP under the PandaSet protocol.

Perception training datamAP on real validation data ↑
Real data only17.37
Real data + 1,000 synthetic samples17.50 (+0.13)

Adding synthetic samples yields a modest gain of 0.13 mAP points on real validation data. We will include both experiments in the revision and describe their evaluation protocols.

Q1. Technical novelty and contribution to the community

Thanks. As detailed in W1, V-REG couples view-specific depth and semantic conditioning with view-specific representation alignment while preserving the main denoising path. This gives practitioners a way to integrate pretrained visual priors into synchronized multi-camera generation, with measured gains in training efficiency and generation quality on PandaSet.

AuthorsOpenReview ↗

We welcome any further discussion or feedback from the reviewer Lfxs.

Dear Reviewer Lfxs:

Hoping this message finds you well. Your comments and the provided insights are very valuable to our work. We will attach these analysis and additional experiments to our final version. As the rebuttal-discussion period is nearing its end, could you please review our response to see if it addresses your concerns? Your timely feedback will be extremely valuable to us. Could you read and let us know if there are more questions? We would be very grateful! Your decision is of utmost importance to us, and we earnestly hope that you will consider the contribution of this research to the field of controllable multi-view driving video generation. Thank you very much !

Best,

The Authors

Thanks for the response. I appreciate the authors' effort. I encourage the authors to carefully revise the manuscript to address the critical weaknesses pointed out by the reviewers.

  1. Novelty: My concern is that the proposed alignment method is primarily an extension of existing techniques from general tasks to autonomous driving tasks (with multiple cameras).

  2. Performance: State-of-the-art methods have achieved significantly better performance on the nuScenes dataset, and the authors' response does not address this concern. For example, MaskGWM reports an FVD of 65.4 (see Table 1 of the MaskGWM paper), UniMLVG reports an FVD of 36.1 (see Table 1 of the UniMLVG paper), MoVieDrive reports an FVD of 46.8 (see Table 1 of the MoVieDrive paper), while the proposed method only achieves an FVD of 98.75.

  3. Scene controllability: I encourage the authors to report scene controllability results on nuScenes, which can be directly compared with state-of-the-art methods.

  4. Data augmentation: The improvement achieved by adding the synthetic data generated by the proposed method is marginal (only +0.13). This raises concerns about the practical value of the proposed approach.

AuthorsOpenReview ↗

Response for additional concerns

We thank the reviewer for the prompt follow-up. We appreciate the clear guidance on the remaining concerns. We address each point below, focusing on methodological novelty, comparisons on nuScenes, scene controllability, and the practical value of synthetic data augmentation.

Novelty. Thank you for clarifying this concern. We respectfully believe that V-REG’s contribution goes beyond applying an existing alignment loss to an autonomous driving dataset. V-REG couples depth- and semantic-based conditioning with supervision of intermediate DiT features, using the same timestep and camera view for both. This correspondence matters in synchronized driving video: cameras at the same instant observe different objects, road regions, and occlusions, so an alignment target from another view would not describe the content being generated. To the best of our knowledge, V-REG is the first to couple view-specific control signals and view-specific representation alignment in this way for synchronized multi-view driving video generation. The resulting design achieves strong generation quality while substantially reducing training time, as discussed in our response to W3 for Reviewer fary. We would welcome a reference to any earlier work using this same coupling and would clarify the distinction in the manuscript accordingly.

Performance Thanks. We fully understand the concern that strong performance on a large-scale benchmark such as nuScenes is essential for assessing the effectiveness of V-REG. The original 98.75 FVD was measured at a 30K-step checkpoint before convergence. After continuing training on 6 H100 GPUs until convergence (50K steps), V-REG achieves 7.57 FID and 43.38 FVD. We have updated this result in Tab 1&2&3.

The following table summarizes the principal differences between the compared methods:

ConfigurationMaskGWM-mviewUniMLVGMoVieDriveV-REG (ours)
Multiple training datasets✓✓××
Multiple training stages✓✓××
Training GPUs32 × A800N/RN/R6 × H100
Generation conditionsR/A,R/T/C/B/M,R/T/C/M/O,R/C/Traj.
Multiple output modalities××✓✓
FID ↓8.905.80—7.57
FVD ↓65.4036.1046.8043.38

R: reference frames; A: action controls; T: text; C: camera parameters; B: 3D boxes; M: HD maps, road maps, or object-box maps; O: 3D occupancy; Traj.: ego trajectories; N/R: not reported.

MaskGWM uses OpenDV-2K pretraining and three training stages, while UniMLVG uses multiple datasets (OpenDV-YouTube, nuScenes, Waymo, Argoverse 2) and stages; its ablation shows that additional datasets and final-stage training substantially reduce FVD. The methods also differ in generation conditions and evaluation protocols. Consequently, their published FVD values provide useful context but are not a fair comparison under matched data, conditioning, compute, and evaluation settings. We detail these differences in Appendix C (Tab.5). Crucially, V-REG's converged 43.38 FVD directly addresses the reviewer's concern about the earlier 30K-step result.

Scene controllability. We thank the reviewer for this constructive suggestion. We have evaluated scene controllability on the nuScenes validation set using a pretrained BEVFormer. Following established practice, we report mAP for 3D object detection and mIoU for bird’s-eye-view (BEV) segmentation.

MethodmAP ↑mIoU ↑
MagicDrive9.7015.60
MagicDrive-V211.5017.40
MoVieDrive22.7035.80
V-REG (ours)22.7435.91
V-REG achieves 22.74 mAP and 35.91 mIoU, exceeding MoVieDrive’s reported 22.70 mAP and 35.80 mIoU on both metrics. These results provide the requested nuScenes controllability comparison. We have attached the results in Appendix D(Tab. 6).

Data augmentation. We thank the reviewer for raising the question of practical value. We respectfully disagree that a +0.13%t gain, by itself, establishes that the benefit of synthetic data is marginal. The gain must be interpreted in the context of the downstream model’s training schedule and the amount of synthetic data added. The +0.13% gain on PandaSet comes from a preliminary experiment using only 1,000 generated samples, rather than an optimized augmentation study. Its size should be interpreted in light of the downstream training schedule: MagicDrive (Appendix F, Tab.6) reports gains as small as +0.25 mAP and +0.03 NDS under its longer 2× schedule. We have clarified the scope of our result in Appendix E(Tab. 7).

We hope that these clarifications will further help the reviewer to solve the remaining concerns. To facilitate a direct assessment of the generated results, we also provide an anonymous project page (https://anonymous.4open.science/w/VREG-4C74/).

Reviewer_faryOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes V-REG, a representation-guided framework for improving the training efficiency of multi-view driving video diffusion models. The method leverages frozen DINOv3 features extracted from depth and semantic maps, injecting frame-view-specific representations into a pretrained DiT backbone and aligning its intermediate features with the corresponding pretrained representations. The framework also incorporates a KV cache mechanism for long-horizon autoregressive generation.

Critical strengths

Training efficiency is an important challenge for large-scale multi-view driving video generation. The paper addresses this problem with a relatively lightweight representation guidance strategy without substantially modifying the original diffusion backbone. The simplicity of the design and the attempt to improve convergence rather than relying on increasingly complicated generation architectures is acceptable.

The method achieves competitive RGB video generation performance while simultaneously generating depth and semantic information. The additional depth and occupancy evaluations help demonstrate the potential usefulness of the generated representations beyond visual appearance.

Critical weaknesses
  1. The paper states that the model jointly predicts future RGB, depth, and semantic videos. However, Sec. 3.4 also states that DINOv3 encodes available depth and semantic conditions during inference, providing frame-view-specific features for the diffusion backbone.

It is unclear where these conditions come from, particularly for future frames. Are they obtained from ground-truth? The use of ground-truth depth and semantic information, particularly during inference, is unclear. It is important to clarify how the corresponding conditioning features are obtained. This directly affects the task formulation and the fairness of comparisons with other methods.

  1. KV caching contributes substantially to the FVD improvement (Table 4). This makes it difficult to isolate the benefits of the proposed representation alignment. Since KV caching is not the main methodological contribution, its impact should be clearly distinguished from that of the representation guidance mechanism.

  2. Limited methodological novelty. The proposed feature alignment is highly similar to existing methods such as REPA (Yu et al., ICLR 2025), CREPA (Hwang et al., arXiv 2025), and Align4Gen (Lee et al., arXiv 2025). Related representation guidance has also been explored in REG (Wu et al., NeurIPS 2025). In particular, CREPA and Align4Gen have already investigated frame-wise representation alignment for video diffusion models. The main contribution of V-REG appears to be extending existing techniques to multi-view generation, rather than introducing a fundamentally new method. The paper does not sufficiently demonstrate why view-specific alignment provides additional benefits beyond existing frame-wise alignment approaches.

  3. Figure 7 are difficult to identify without highlighted regions or zoomed-in comparisons.

Questions for authors

How are depth and semantic conditions obtained during inference? Is any future ground-truth information used during generation or evaluation?

How much of the reported improvement comes from KV caching rather than representation alignment? Are the comparisons conducted under consistent caching settings?

1 comment

We thank the reviewer for recognizing the importance of training efficiency in multi-view driving video generation and the value of jointly generating RGB, depth, and semantic outputs. We detail the correction and address the remaining concerns below.

W1 & Q1. Inference conditions and use of future ground truth

Thanks for the question. We initialize both auxiliary conditioning streams with the initial RGB frame and then update them autoregressively using the model’s own predictions.

  • First frame. The initial RGB frame, which is also provided to the baselines, directly replaces the depth and semantic maps as the input to DINOv3 for both conditioning streams.
  • Subsequent frames. At each step, DINOv3 encodes the predicted depth and semantic maps from the immediately preceding frame. The resulting features condition generation of the next frame.

W2 & Q2. Contributions of KV caching and representation alignment

Good question ! Thank you for raising this point. KV caching contributes substantially to the FVD improvement. To distinguish its effect from that of representation alignment, we add a cache-only baseline (ID 5) to the revised Table 4.

IDConfigurationKV cacheFID ↓FVD ↓
1Baseline×38.40152.83
2Depth branch with alignment×35.51135.99
3Semantic branch with alignment×36.95147.83
4Depth and semantic branches without alignment×37.17140.71
5Baseline + KV cache✓35.86113.01
6Depth and semantic branches with alignment×29.57129.79
7Full V-REG✓27.5689.79

Representation alignment improves both metrics with caching disabled. Relative to the baseline, depth-guided alignment (ID 2) and semantic-guided alignment (ID 3) each improve FID and FVD, and combining both (ID 6) yields the best results among the uncached variants. The direct test of alignment is ID 4 vs. ID 6. With both auxiliary branches held fixed and caching disabled, enabling alignment reduces FID from 37.17 to 29.57 and FVD from 140.71 to 129.79.

KV caching yields a larger improvement in FVD, and the two components are complementary. Adding caching alone (ID 1 → ID 5) reduces FID by 2.54 and FVD by 39.82. Adding caching to the configuration that already uses depth and semantic alignment (ID 6 → ID 7) further reduces FID from 29.57 to 27.56 and FVD from 129.79 to 89.79. These comparisons show that alignment improves frame-level image quality as measured by FID and also improves FVD, while the larger FVD gain from caching is consistent with its role in preserving spatiotemporal consistency. The full model achieves the best results on both metrics.

W3. Methodological novelty and the role of camera-view correspondence

Thank you for highlighting the connections to REPA, CREPA, Align4Gen, and REG. V-REG builds on these advances in representation alignment. Our methodological contribution is to adapt representation alignment specifically for controllable multi-view driving video generation. We use depth- and semantic-derived representations to guide generation for individual camera views and to supervise the corresponding intermediate DiT features during training. Both conditioning and alignment are indexed by timestep and camera view, linking the control signal for each view to the features that generate that view.

This design addresses a specific requirement of synchronized driving video. At the same timestep, different cameras can observe distinct objects, road regions, and occlusion patterns. Matching each DiT feature to the depth and semantic representations from its own camera provides supervision tied to the geometry and content that the camera actually observes. Coupling view-specific conditioning with view-specific representation alignment is the central design contribution of V-REG.

We compare V-REG with naive CREPA and Align4Gen on PandaSet.

MethodParameters (B)Training time (h) ↓FID ↓FVD ↓
Baseline11.006.038.40152.83
+ CREPA11.215.031.55129.17
+ Align4Gen11.225.532.02126.78
+ V-REG (ours)11.231.627.5689.79

V-REG achieves the lowest FID and FVD and the shortest reported training time. Table 4 (ID 4 → ID 6) also shows that alignment alone improves both metrics when the auxiliary branches are held fixed and caching is disabled. These experiments evaluate the complete design and the incremental benefit of alignment, respectively.

W4. Readability of the qualitative comparisons in Fig. 7

Thanks. We have revised Fig. 7 (p. 9 of the revised manuscript) by marking corresponding regions across all methods and adding crops at the same magnification. The updated figure is outlined in red.

Reviewer_N3FQOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes V-REG, which injects frame-view-specific DINOv3 representations into an 11B-parameter multimodal diffusion transformer and aligns intermediate DiT features with these representations during training. The controlled experiments suggest faster iteration-level convergence and improved RGB, depth, and occupancy generation.

Critical strengths

S1. Important practical motivation. Reducing the adaptation cost of large driving-video generators is valuable. The paper also appropriately distinguishes iteration count from fully normalized compute in Table 1.

S2. Lightweight integration. Section 3.4 preserves the main denoising path and removes the alignment loss and projection at inference. Figure 5(c) reports modest training overhead: iteration time increases from 12.80 s to 13.62 s, memory from 43.60 GB to 46.00 GB, and parameters from 11.22B to 11.23B.

Critical weaknesses

W1. To verify that the method is generally effective, please also compare it against methods in Table 1 that use different DiT models, and check whether it likewise accelerates training convergence. Prior methods typically use smaller DiT models, so the effect of model size needs to be ruled out.

Questions for authors

Q1. Figure 6 shows temporal consistency, but the sequences last only 4 seconds. This span is too short to establish the effectiveness of the spatiotemporal cache mechanism. Please add quantitative results and visualizations over longer horizons, for example 20 seconds.

2 comments

We sincerely thank the reviewer for recognizing the practical value of reducing the adaptation cost of large driving-video generators and for highlighting V-REG’s modest training overhead. We address both comments below.

W1. Convergence across DiT backbones and model sizes

Thank you for this suggestion. Table 1 reports our main evaluation on the large-scale nuScenes dataset. Given the constraints on running the full set of cross-backbone experiments on nuScenes, we conduct this additional study on PandaSet, a smaller, established autonomous-driving benchmark. We integrat V-REG into DiT, fastDiT, SiT, MaskDiT, and SD-DiT and compare each variant with its corresponding baseline on PandaSet.

Baseline → V-REG applies to all different backbones.

BackboneBackbone params (M)Training time (h) ↓FID ↓FVD ↓
DiT6756.0 → 1.638.40 → 27.56152.83 → 89.79
fastDiT6754.5 → 3.061.50 → 42.21229.76 → 173.01
SiT6759.0 → 7.077.94 → 55.91353.34 → 199.33
MaskDiT7309.0 → 7.565.72 → 40.51297.20 → 165.00
SD-DiT7406.9 → 3.063.74 → 37.92277.27 → 194.06

Across all five backbones on PandaSet, V-REG achieves lower FID and FVD at the reported checkpoints with 16.7%–73.3% less wall-clock training time. Each paired comparison uses the same backbone, showing that scales well for different backbones. We have included these results in the revision.

Q1. Quantitative and visual evaluation over 20-second rollouts

Thank you for raising this point. We report FVD at 1, 5, 10, 15, and 20 seconds for category-aligned and frame-aligned REG. We also extend Fig. 6 (p. 8 of the revised manuscript) to show a 20-second autoregressive rollout, illustrating how vehicle trajectories evolve across successive generation windows. The updated panel is outlined in red.

MethodFVD₁ ↓FVD₅ ↓FVD₁₀ ↓FVD₁₅ ↓FVD₂₀ ↓
Category-aligned REG80.8115.8122.8126.9130.8
Frame-aligned REG73.298.8104.4104.8105.2

Frame-aligned REG achieves lower FVD at every evaluated horizon. At 20 seconds, it reduces FVD from 130.8 to 105.2, a 19.6% relative reduction. From 10 to 20 seconds, FVD increases by 0.8% with frame-aligned REG and by 8.0% with category-aligned REG. These results show that the video quality gains from frame-level alignment persist through the 20-second rollout.

We welcome any further discussion if needed.

Thank the authors for the detailed response. My concerns have been addressed with the revised claim and additional experiments. I will keep my previous score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.