acceptodds
Under review as a conference paper at ICLR 2027

FlashRender: Few-Step Generative Rendering via Camera-Controlled Video MeanFlow

18 views

Abstract

We present FlashRender, a few-step generative rendering framework that retakes a source video along a target camera trajectory in seconds. We identify sampling-step-dependent camera control as a prominent manifestation of discretization error in existing multi-step generative rendering models and show that resolving this inconsistency substantially lowers denoising trajectory curvature, facilitating subsequent step distillation. To this end, we introduce Representation Transformation and Alignment (RETA), which aligns hidden source-video representations with target-video features from a frozen visual geometry model. This directly encodes the geometric transformation within the source-video stream, enabling sampling-step-consistent camera control. We then fine-tune the model with the MeanFlow objective on the lower-curvature denoising trajectory induced by RETA, allowing the model to more effectively address discretization error. Finally, we apply on-policy flow map distillation to correct self-rollout errors under fixed few-step sampling. Extensive experiments show that RETA, MeanFlow, and on-policy flow map distillation play complementary roles in few-step generative rendering. Together, they enable our approach to match multi-step baselines in video quality and geometric consistency at lower sampling cost while achieving superior camera controllability, even under out-of-distribution target camera trajectories.

Primary area:
generative models
Keywords:
camera control, video generation, representation alignment, few-step distillation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_XD9GOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents FlashRender, a framework for camera-controlled generative rendering that produces a target-view retake from an input video and a specified camera trajectory using four network evaluations. The authors identify sampling-step-dependent camera control as a challenge in accelerating existing video retaking models: reducing the sampling budget can change scene scale, object localization, and the realized camera motion. They propose that more stable geometric representations can facilitate subsequent few-step training. Built on Wan2.1-1.3B-CamCtrl, FlashRender follows a three-stage training pipeline. The first stage introduces frame-wise relative camera pose conditioning and Representation Transformation and Alignment (RETA). RETA aligns intermediate source-video representations with target-video features extracted using a frozen VGGT-based geometry encoder, encouraging the source stream to encode target-view structure. The authors associate this alignment with improved consistency across sampling budgets and reduced denoising trajectory curvature. The second stage adopts a MeanFlow objective to learn average velocities between timesteps, enabling large transitions along the denoising trajectory. The third stage performs on-policy flow map distillation using the same four-step schedule as inference, combining distribution matching and adversarial objectives to improve the quality of student-generated samples. The method is trained on MultiCamVideo and SynCamVideo and evaluated on 500 DAVIS cases comprising 50 source videos and ten camera trajectories. Comparisons include explicit and implicit multi-step methods, an existing few-step method, and baselines adapted using the proposed distillation framework. The reported results show that FlashRender largely preserves its multi-step teacher’s visual quality and geometric consistency while improving camera accuracy and reducing network evaluations from 100 to four. Additional DyCheck experiments assess generalization to out-of-distribution camera trajectories, showing strong visual and geometric scores but some loss of ground-truth fidelity relative to leading multi-step baselines. Ablations, sampling-budget comparisons, trajectory measurements, and feature visualizations examine the contributions of RETA, MeanFlow training, and on-policy distillation, supporting their complementary roles in few-step video retaking.

Critical strengths

1.A compelling quality–efficiency trade-off. FlashRender reduces the number of network evaluations from 100 to four while largely preserving its multi-step teacher’s visual quality and geometric consistency. On DAVIS, it also improves translation error from 0.0143 to 0.0122 and rotation error from 1.486 to 1.236. Comparisons with distilled ReDirector and GeoAlign variants further strengthen the evaluation by including baselines specifically adapted for few-step generation. Achieving this degree of acceleration while maintaining control over the requested camera motion is practically consequential and constitutes a primary reason for my positive assessment.

2.Task-specific geometric supervision with measurable benefits. RETA provides a well-motivated approach to cross-view conditioning by aligning intermediate source-video representations with target-view geometric features from a frozen visual geometry model. Its contribution is supported by ablations: removing RETA increases MEt3R from 0.3071 to 0.3367 and rotation error from 1.236 to 1.685. The additional sampling-budget comparisons and feature analyses provide further evidence that this supervision benefits few-step generation. Although the precise causal explanation merits further investigation, the demonstrated effectiveness of source-to-target representation alignment supports a meaningful methodological contribution and strengthens the case for acceptance.

Critical weaknesses

1.The central mechanistic explanation is insufficiently established. The paper links improved cross-step camera consistency to reduced denoising trajectory curvature and easier few-step distillation. However, cross-step PSNR also reflects changes in appearance and object motion, while Eq. (16) measures deviation from a constant-velocity reference. These analyses demonstrate associated improvements but do not isolate camera consistency as the mechanism responsible for improved distillability. Furthermore, matched comparisons using the same VGGT teacher to supervise source versus target representations would help distinguish the proposed transformation mechanism from the broader benefits of geometric supervision. Since the subsequent training stages substantially build on existing methods, this unresolved attribution limits confidence in the paper’s distinctive methodological contribution.

2.The comparative evaluation does not fully establish the advantage of the proposed distillation recipe. Appendix A reports an additional 20k iterations of MeanFlow training before 5k iterations of on-policy distillation, whereas the DMD2 baseline receives 5k iterations. Consequently, removing Stage 2 changes both the training objective and the training budget, making its independent contribution difficult to assess. Equal-compute comparisons or performance-versus-training-cost curves would strengthen this conclusion. Additionally, RealCam is acknowledged as a directly related few-step generative rendering approach but is neither evaluated nor accompanied by an explanation for its omission. Together, these gaps weaken the evidence that the proposed training recipe provides a clear advantage over the closest alternatives.

Questions for authors

1.Clarification of the trajectory analysis. In Eq. (16), does (X_0) denote the ground-truth target latent or the endpoint of the generated trajectory, and how is the reported quantity normalized? Were identical initial noise samples used for the comparisons across sampling budgets in Fig. 8? These details would help determine how much the measurements reflect changes in the sampling trajectory versus differences in reconstruction or appearance.

2.Targeted validation of the RETA mechanism. Holding the VGGT teacher, alignment layer, projector, camera encoding, and training budget fixed, how does aligning source representations to target-view features compare with aligning target representations to those same features? Is there evidence that specifically isolates the benefit of learning the transformation within the source stream from the general benefit of geometric supervision?

3.Clarification and targeted validation of the training comparison. Were the DMD2 and DMD2+DFD baselines matched to the proposed method in total training compute, including the additional 20k iterations of MeanFlow training? If not, are there controlled results using comparable compute, or continued flow-matching training in place of Stage 2, that establish the benefit of the MeanFlow objective independently of additional optimization?

4.Clarification of the missing direct baseline. Why was RealCam excluded from the experimental comparisons despite being identified as a directly related few-step generative rendering method? Was this due to unavailable code or checkpoints, differences in the evaluation protocol, or another limitation? Clarifying this would materially affect my assessment of the coverage of the closest competing approaches.

3 comments
AuthorsOpenReview ↗

Response to Reviewer XD9G (1/2)

We sincerely appreciate your insightful comments and efforts in reviewing our manuscript. We address each comment below.


[W1 & Q1] Trajectory analysis.

In Eq. (16), X1X_1 is the initial noise z1z_1 and X0X_0 is the ground-truth target latent. The squared L2 norm is averaged over all latent tokens, and this per-sample value is then averaged over samples from the training dataset. As the reviewer notes, Eq. (16) measures the deviation from a constant-velocity reference. This is the standard straightness measure in rectified flow and TMD, since a perfectly straight trajectory is integrated exactly by a single Euler step, and it therefore directly reflects few-step discretization error. Its absolute value depends on the latent scale, but all models in Fig. 3 and 9b are fine-tuned from the same Wan2.1 backbone and share its VAE latent space, so the curves are directly comparable. For the comparison across sampling budgets in Fig. 9a (Fig. 8 in the original submission), retakes of each input video are generated from the same initial noise, so the differences reflect the sampling budget rather than the noise.

Since pairwise PSNR also reflects changes in appearance and object motion, we additionally measure camera viewpoint consistency directly. However, TransErr and RotErr do not fully capture camera viewpoint consistency across sampling budgets. TransErr is computed after aligning the estimated and target camera trajectories with a similarity transform [1], which removes differences in scene scale. For example, if the 4-step retake renders the scene noticeably closer to the camera than the 50-step retake while moving along the same path, both obtain similar TransErr. Also, RotErr compares relative rotations between generated frames, so it ignores any rotation offset shared by all frames. For example, if the 4-step retake looks a few degrees to the left of the 50-step retake throughout the video, both obtain the same RotErr. Consequently, these metrics alone cannot determine whether retakes generated with different sampling budgets depict the same views.

We therefore adopt a homography-based evaluation, drawing on video stabilization methods [2, 3], to assess camera viewpoint consistency across sampling budgets. Specifically, for each frame index t∈Tt \in \mathcal{T} and sampling budget B∈B=4,8,12,20B \in \mathcal{B} = \\{4, 8, 12, 20\\}, we estimate a homography HtBH^B_t mapping the generated frame to the corresponding frame of the 50-step retake generated from the same initial noise. Ideally, corresponding frames depict the same view, and HtBH^B_t is the identity matrix. From each HtBH^B_t, we compute four errors that are all zero for the identity transform. These comprise scale error, translation error (the displacement of the image center), rotation error and overall error (the mean displacement of the four image corners [4]). For each metric m(⋅)m(\cdot), we average over sampling budgets and frame indices within each case 1∣B∣,∣T∣∑B∈B,,t∈Tm(HtB)\frac{1}{|\mathcal{B}|\\,|\mathcal{T}|}\sum_{B \in \mathcal{B},\\, t \in \mathcal{T}} m(H^B_t), and then average across the evaluation dataset. The results show that FlashRender-MS exhibits the smallest deviation from its 50-step reference across all four measures, supporting improved camera viewpoint consistency across sampling budgets. We have added these results to the revised manuscript (Table 5 in Appendix B). Together with Fig. 3, where RETA yields the lowest curvature among the three Stage-1 variants, these results suggest that resolving inconsistent camera control makes the denoising trajectory easier to approximate, as discussed in Appendix B.

MethodScale (|log s|)↓Translation (%)↓Rotation (°)↓Overall (%)↓
ReCamMaster0.00980.3762.0441.140
ReDirector0.00680.7691.2421.252
GeoAlign0.00400.3560.7700.615
FlashRender-MS0.00370.2410.6470.482

References

[1] Umeyama, Least-squares estimation of transformation parameters between two point patterns, IEEE TPAMI 1991
[2] Grundmann et al., Auto-directed video stabilization with robust L1 optimal camera paths, CVPR 2011
[3] Liu et al., Bundled camera paths for video stabilization, SIGGRAPH 2013
[4] DeTone et al., Deep image homography estimation, arXiv 2016

AuthorsOpenReview ↗

Response to Reviewer XD9G (2/2)

[W1 & Q2] Isolating the source-stream transformation from generic geometric supervision.

We thank the reviewer for suggesting this controlled comparison. Unlike REPA-style methods (iREPA, VideoREPA, Geometry Forcing), which align a stream with features of the same video, RETA aligns the clean source-video stream with VGGT features of the target video, which is captured from a different viewpoint. To isolate this source-to-target transformation from the general benefit of geometric supervision, we train a Stage-1 model that applies identical VGGT supervision to the target stream instead, keeping the VGGT teacher, alignment layer, projector, camera encoding, loss and training budget fixed. As shown in the table below, aligning the target stream performs worse on all metrics, isolating the benefit of learning the transformation within the source stream. In the revised manuscript, Fig. 3 also shows that RETA yields the lowest denoising trajectory curvature among the three Stage-1 variants: without RETA, with target-stream alignment, and with RETA (source-stream alignment).

We also clarify why the stream matters. Target-stream tokens are noisy, so at high noise levels they carry little geometric signal, and the source-to-target correspondence must still be re-established through attention at each step. Source-stream tokens are clean, so RETA asks the network to encode the target-view transformation where the input is noise-free and the alignment target is the same at every timestep.

Alignment TokenAesthetic Quality↑Imaging Quality↑Dyn-MEt3R↑MEt3R↓TransErr↓RotErr↓
Target-stream0.51650.65190.83200.32380.01591.583
Source-stream (RETA)0.52140.66250.84910.30590.01431.486

[W2 & Q3] Training compute and the contribution of Stage 2.

The DMD2 and DMD2+DFD baselines used less total training compute than FlashRender, which additionally includes Stage 2, as shown in the table below. We therefore add two compute-matched baselines with the same total budget as FlashRender: DMD2 trained for more iterations, and, as the reviewer suggests, FlashRender with continued flow-matching training in place of Stage 2, which spends the Stage-2 budget on longer Stage-1 training. FlashRender outperforms both on all metrics, supporting the benefit of the MeanFlow objective and the training design beyond simply increasing the training budget. We have updated Table 3 in the revised manuscript to include compute-matched baselines and report training GPU-hours for all methods.

MethodStage 1Stage 2DistillationGPU-hAesthetic Quality↑Imaging Quality↑Dyn-MEt3R↑MEt3R↓TransErr↓RotErr↓
Original training budget
DMD220k–DMD2, 5k8940.51660.66290.84470.31900.01291.544
w/o Stage 220k–Stage 3, 5k9680.51320.63120.84440.32150.01281.294
Compute-matched to FlashRender
DMD220k–DMD2, 21.2k1,9600.51580.66250.84500.31770.01291.493
w/o Stage 255.2k–Stage 3, 5k1,9600.51200.65050.84750.30930.01331.299
FlashRender20k20kStage 3, 5k1,9600.51820.66540.85710.30710.01221.236

[W2 & Q4] RealCam.

RealCam is a closely related few-step rendering method. However, its code and pretrained checkpoints were unavailable. We now state this limitation in Sec. 5.1.


We hope these additions address the reviewer's concerns. All changes are highlighted in blue in the revised manuscript. We welcome any further questions.

Dear Reviewer XD9G,

As the discussion period draws to a close, we would appreciate any feedback on whether our responses have addressed your concerns. We would be happy to clarify any remaining points or answer further questions.

Reviewer_9sfmOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FlashRender generates a new-view video from a source video and a target camera trajectory using only four network evaluations. It combines target-view geometry alignment, MeanFlow training, and on-policy distillation to improve camera control under few-step sampling. Experiments show competitive visual quality and geometric consistency at substantially lower sampling cost than multi-step baselines.

Critical strengths
  • Practical acceleration. FlashRender reduces generative rendering to four network evaluations while retaining competitive visual quality, geometric consistency, and camera control on the evaluated benchmarks. This offers practical value for iterative video retaking.

  • Useful empirical validation. The paper includes comparisons with representative rendering methods, few-step adaptations of existing models, and ablations of representation alignment and training stages. These experiments support the effectiveness of the proposed recipe beyond a simple reduction in sampling steps.

Critical weaknesses
  • Limited novelty. My main concern about this paper is its novelty. The method mainly adapts existing techniques, and RETA’s distinct contribution beyond a task-specific training recipe remains insufficiently established.

  • Incomplete attribution. The ablations show effectiveness, but do not fully isolate the proposed mechanism or separate its benefits from additional training computation.

  • Unclear robustness boundaries. Severe occlusion and large newly exposed regions are not systematically evaluated, limiting support for broad content-preservation and generalization claims.

Questions for authors
  • Could you clarify the total training-compute budgets of FlashRender and the DMD2-based baselines in Table 2? Appendix A reports an additional 20k MeanFlow iterations for FlashRender. If the budgets differ substantially, one compute-matched comparison would help determine whether the improvement comes from the proposed training design or additional optimization.

  • Could you explain why RealCam, a closely related few-step rendering method cited in the paper, is excluded from the comparisons? If its implementation or checkpoint is unavailable, stating this limitation would suffice.

7 comments
AuthorsOpenReview ↗

Response to Reviewer 9sfm (1/2)

We sincerely appreciate your insightful comments and efforts in reviewing our manuscript. We address each comment below.


[W1] Limited novelty and methodological contribution of RETA.

We adopt established few-step distillation techniques for camera-controlled video-to-video generation and find that a problem persists even when they are used together. Applying the same MeanFlow training and on-policy distillation procedure to ReDirector and GeoAlign produces four-step students with lower camera accuracy and geometric consistency than their respective multi-step teachers (Table 1). Investigating this gap, we observed that camera control itself depends on the sampling budget. Even with the input video, target trajectory, and initial noise fixed, changing the number of sampling steps can alter the realized viewpoint and object placement (Fig. 8 and Table 5).

We designed RETA to test the hypothesis that sampling-step-dependent camera control contributes to high denoising trajectory curvature, and that stabilizing camera control can reduce curvature and facilitate subsequent few-step distillation. Unlike REPA-style methods (iREPA, VideoREPA, Geometry Forcing), which align a stream with features of the same video, RETA aligns the clean source-video stream with VGGT features of the target video, which is captured from a different viewpoint. Applying identical VGGT supervision to the target stream instead performs worse on all metrics (see the table below), isolating the benefit of this source-to-target transformation. Before distillation, RETA yields the lowest denoising trajectory curvature among the three Stage-1 variants (Fig. 3), and FlashRender-MS exhibits the smallest viewpoint variation across sampling budgets (Figs. 8 and 9a and Table 5). Together, these observations support the hypothesized link between camera stability and trajectory curvature. After applying the same distillation procedure, the 4-NFE student trained with RETA remains competitive with its 100-NFE teacher in visual quality and geometric consistency while achieving higher camera accuracy (Table 1). Rather than a task-specific recipe, these results identify why off-the-shelf step distillation fails for camera-controlled video-to-video generation and resolve it at the level of the multi-step model.

Alignment TokenAesthetic Quality↑Imaging Quality↑Dyn-MEt3R↑MEt3R↓TransErr↓RotErr↓
Target-stream0.51650.65190.83200.32380.01591.583
Source-stream (RETA)0.52140.66250.84910.30590.01431.486

[W2 & Q1] Separating the gains from additional training compute.

The DMD2 and DMD2+DFD baselines used less total training compute than FlashRender, as shown in the table below. We therefore add two compute-matched baselines with the same total budget as FlashRender: DMD2 trained for more iterations, and FlashRender without MeanFlow that spends the Stage-2 budget on longer Stage-1 training. FlashRender outperforms both on all metrics, supporting the benefit of the training design beyond simply increasing the training budget. Moreover, the RETA ablations (w/o RETA and RETA → VideoREPA) are not confounded by training compute either: they keep the architecture, data, training stages and iteration counts fixed, using nearly the same GPU-hours, and only remove or replace the alignment objective. We have updated Table 3 in the revised manuscript to include compute-matched baselines and report training GPU-hours for all methods.

MethodStage 1Stage 2DistillationGPU-hAesthetic Quality↑Imaging Quality↑Dyn-MEt3R↑MEt3R↓TransErr↓RotErr↓
Original training budget
DMD220k–DMD2, 5k8940.51660.66290.84470.31900.01291.544
w/o Stage 220k–Stage 3, 5k9680.51320.63120.84440.32150.01281.294
Compute-matched to FlashRender
DMD220k–DMD2, 21.2k1,9600.51580.66250.84500.31770.01291.493
w/o Stage 255.2k–Stage 3, 5k1,9600.51200.65050.84750.30930.01331.299
FlashRender20k20kStage 3, 5k1,9600.51820.66540.85710.30710.01221.236
AuthorsOpenReview ↗

Response to Reviewer 9sfm (2/2)

[W3] Robustness to severe occlusion and large disoccluded regions.

We agree that this boundary deserves explicit evaluation. Since RETA supervises the source-to-target transformation, its benefits may depend on how much of the target view is supported by the input video. We therefore evaluate them under increasing disocclusion. For each of the 500 DAVIS cases, we compute the new-content ratio, i.e., the fraction of target-view pixels unobserved in any input frame, by lifting the input frames to 3D with metric depth maps estimated by UniDepth and projecting them into the target cameras. Since this ratio depends only on the input video and the target trajectory, all methods are evaluated on the same cases. The table below groups the cases into three bins and compares all 4-NFE methods.

The results show that FlashRender achieves the best geometric consistency and camera accuracy across all three groups. Even in the highest-ratio group, FlashRender substantially improves MEt3R, which measures consistency with the input video, over the other few-step methods. These results suggest that the source-content preservation advantage of FlashRender persists even when substantial unseen content must be synthesized. We add this analysis to Table 2 and Sec. 5.2 of the revised manuscript.

MethodAesthetic Quality↑Imaging Quality↑Dyn-MEt3R↑MEt3R↓TransErr↓RotErr↓
New-content ratio < 20% (342 cases)
NeoVerse0.50140.66200.75310.34670.03703.717
ReDirector-FS0.51790.63510.84680.32640.01421.213
GeoAlign-FS0.52130.64720.85460.29330.01451.280
FlashRender0.52850.66570.86370.28500.01060.942
New-content ratio 20–40% (121 cases)
NeoVerse0.49720.69060.75230.40090.02702.704
ReDirector-FS0.47690.66050.81670.40000.02122.705
GeoAlign-FS0.48350.65460.83460.36020.01511.696
FlashRender0.49830.66470.84970.35700.01331.624
New-content ratio ≥ 40% (37 cases)
NeoVerse0.51240.70170.74380.37460.03354.033
ReDirector-FS0.46630.66420.79480.39440.04113.539
GeoAlign-FS0.46620.63340.81550.38220.03033.273
FlashRender0.48770.66620.82040.34810.02322.681

[Q2] Comparison with RealCam.

RealCam is a closely related few-step rendering method. However, its code and pretrained checkpoints were unavailable. We now state this limitation in Sec. 5.1.


We hope these additions address the reviewer's concerns. All changes are highlighted in blue in the revised manuscript. We welcome any further questions.

Dear Reviewer 9sfm,

As the discussion period draws to a close, we would appreciate any feedback on whether our responses have addressed your concerns. We would be happy to clarify any remaining points or answer further questions.

Reviewer_L6yBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents FlashRender, a four-step framework for camera-controlled video retaking. It combines RETA, which aligns source representations with target-view VGGT features, with MeanFlow training and on-policy distillation. Experiments demonstrate competitive visual quality and geometric consistency with multi-step baselines using 25×25\times fewer network evaluations, alongside improved camera accuracy on DAVIS.

Critical strengths

1. Practical speed-quality trade-off. FlashRender reduces sampling from 100 to 4 network evaluations while retaining competitive visual quality and geometric consistency and improving camera accuracy on DAVIS.

2. Effective geometry-guided conditioning. RETA guides source representations toward target-view geometry, with standalone ablations demonstrating benefits before the acceleration stages.

3. Substantial experimental validation. Alternative alignment and distillation comparisons, alongside ablations across stages and sampling budgets, support the training recipe.

Critical weaknesses

1. The distinctive methodological contribution is insufficiently isolated. Much of the pipeline builds on appropriately cited prior work: AnyFlow combines MeanFlow with on-policy flow-map distillation, while Geometry Forcing aligns diffusion representations with VGGT features using cosine similarity. The main distinction lies in source-to-target alignment and its interaction with camera conditioning. Although standalone RETA ablations demonstrate its benefit, a matched comparison applying the same VGGT supervision to target-stream tokens is missing. This leaves unclear how much the gains specifically depend on the proposed source-stream transformation.

2. The proposed mechanism remains partially supported. RETA improves sampling consistency and reduces trajectory curvature, but the results do not establish that stabilizing camera control causes this reduction. Cross-budget PSNR (Fig. 8a) is not camera-specific, and joint source-target attention does not itself enforce timestep-independent source features. Direct camera-trajectory comparisons across sampling budgets with matched initial noise would strengthen this interpretation; alternatively, the causal explanation should be stated more cautiously.

Minor weakness

  • Unequal training compute complicates attribution. FlashRender adds 20k MeanFlow iterations, each requiring multiple forward passes, before distillation. Compute-matched comparisons with DMD2 would help separate methodological gains from additional training.
Questions for authors
  1. How does RETA compare with identical VGGT supervision applied to target-stream tokens?
  2. Can you measure camera-trajectory variation across sampling budgets using matched initial noise?
  3. How sensitive is FlashRender to errors in ViPE's source-camera estimates?
4 comments
AuthorsOpenReview ↗

Response to Reviewer L6yB (1/2)

We sincerely appreciate your insightful comments and efforts in reviewing our manuscript. We address each comment below.


[W1 & Q1] RETA vs. identical VGGT supervision applied to target-stream tokens.

We agree that this is an important ablation study that isolates the benefit of source-stream transformation. We conducted the matched comparison in Stage-1. Under identical VGGT supervision, the below table shows that source-stream alignment outperforms target-stream alignment across all evaluation metrics. These results demonstrate that the choice of alignment stream matters and support the effectiveness of transforming source representations toward the target view. We will also include the Stage-3 results in the final manuscript, as completing these experiments takes approximately two weeks.

Alignment TokenAesthetic Quality↑Imaging Quality↑Dyn-MEt3R↑MEt3R↓TransErr↓RotErr↓
Target-stream0.51650.65190.83200.32380.01591.583
Source-stream (RETA)0.52140.66250.84910.30590.01431.486

We have also revised Fig. 3 to compare denoising trajectory curvature across three Stage-1 models trained without RETA, with target-stream alignment, and with RETA (source-stream alignment). We note that RETA yields the lowest curvature, providing further evidence that source-stream transformation helps straighten the denoising trajectory.

[W2 & Q2] Camera trajectories across sampling budgets with matched initial noise.

Thank you for this concrete suggestion. We first note that TransErr and RotErr do not fully capture camera viewpoint consistency across sampling budgets. TransErr is computed after aligning the estimated and target camera trajectories with a similarity transform [1], which removes differences in scene scale. For example, if the 4-step retake renders the scene noticeably closer to the camera than the 50-step retake while moving along the same path, both obtain similar TransErr. Also, RotErr compares relative rotations between generated frames, so it ignores any rotation offset shared by all frames. For example, if the 4-step retake looks a few degrees to the left of the 50-step retake throughout the video, both obtain the same RotErr. Consequently, these metrics alone cannot determine whether retakes generated with different sampling budgets depict the same views.

We therefore adopt a homography-based evaluation, drawing on video stabilization methods [2, 3], to assess camera viewpoint consistency across sampling budgets. Specifically, for each frame index t∈Tt \in \mathcal{T} and sampling budget B∈B=4,8,12,20B \in \mathcal{B} = \\{4, 8, 12, 20\\}, we estimate a homography HtBH^B_t mapping the generated frame to the corresponding frame of the 50-step retake generated from the same initial noise. Ideally, corresponding frames depict the same view, and HtBH^B_t is the identity matrix. From each HtBH^B_t, we compute four errors that are all zero for the identity transform. These comprise scale error, translation error (the displacement of the image center), rotation error and overall error (the mean displacement of the four image corners [4]). For each metric m(⋅)m(\cdot), we average over sampling budgets and frame indices within each case 1∣B∣,∣T∣∑B∈B,,t∈Tm(HtB)\frac{1}{|\mathcal{B}|\\,|\mathcal{T}|}\sum_{B \in \mathcal{B},\\, t \in \mathcal{T}} m(H^B_t), and then average across the evaluation dataset. The results show that FlashRender-MS exhibits the smallest deviation from its 50-step reference across all four measures, supporting improved camera viewpoint consistency across sampling budgets. We have added these results to the revised manuscript (Table 5 in Appendix B).

MethodScale (|log s|)↓Translation (%)↓Rotation (°)↓Overall (%)↓
ReCamMaster0.00980.3762.0441.140
ReDirector0.00680.7691.2421.252
GeoAlign0.00400.3560.7700.615
FlashRender-MS0.00370.2410.6470.482

References

[1] Umeyama, Least-squares estimation of transformation parameters between two point patterns, IEEE TPAMI 1991
[2] Grundmann et al., Auto-directed video stabilization with robust L1 optimal camera paths, CVPR 2011
[3] Liu et al., Bundled camera paths for video stabilization, SIGGRAPH 2013
[4] DeTone et al., Deep image homography estimation, arXiv 2016

AuthorsOpenReview ↗

Response to Reviewer L6yB (2/2)

[W3 (minor)] Training compute.

Stage-2 is essential for on-policy flow map distillation in Stage-3, as reliable flow map rollouts require a model already trained with rr-conditioning. Therefore, applying Stage-3 alone yields worse performance than DMD2 on most metrics except camera accuracy, despite requiring more training compute. In contrast, FlashRender combines Stage-2 and Stage-3 and outperforms both compute-matched baselines, one using more DMD2 iterations and the other using more Stage-1 iterations. We have also updated Table 3 with these compute-matched baselines.

MethodStage 1Stage 2DistillationGPU-hAesthetic Quality↑Imaging Quality↑Dyn-MEt3R↑MEt3R↓TransErr↓RotErr↓
Original training budget
DMD220k–DMD2, 5k8940.51660.66290.84470.31900.01291.544
w/o Stage 220k–Stage 3, 5k9680.51320.63120.84440.32150.01281.294
Compute-matched to FlashRender
DMD220k–DMD2, 21.2k1,9600.51580.66250.84500.31770.01291.493
w/o Stage 255.2k–Stage 3, 5k1,9600.51200.65050.84750.30930.01331.299
FlashRender20k20kStage 3, 5k1,9600.51820.66540.85710.30710.01221.236

[Q3] Sensitivity to errors in ViPE source poses.

Since the ViPE pose estimates already contain some error, we evaluate sensitivity to additional noise in the DAVIS dataset with substantial source camera motion. We select the eight scenes whose ViPE source trajectories exhibit at least 5m of translation and 5° of rotation and generate retakes for all 10 target trajectories in each scene. We perturb the estimated source poses with increasing levels of rotational or translational noise in separate experiments. For each source pose, we apply either a rotation through the specified angle about a random axis or a translation in a random direction by the specified fraction of the extent of the source trajectory.

Added noiseAesthetic Quality↑Imaging Quality↑Dyn-MEt3R↑MEt3R↓TransErr↓RotErr↓
None0.53540.65940.84390.35330.01331.582
Rotation 10°0.53520.65690.84240.36610.01511.760
Rotation 20°0.53610.65700.81890.37160.01762.303
Rotation 45°0.52490.65150.82020.39420.02022.193
Translation 10%0.53520.65810.84110.36170.01381.614
Translation 20%0.53500.65770.80550.36240.01591.748
Translation 50%0.52450.65240.79830.40940.01721.905

As source pose errors increase, visual quality remains largely stable, while geometric consistency and camera accuracy gradually decline. We hypothesize that conditioning on the target camera trajectory partially offsets the effects of source pose errors. As a result, FlashRender remains robust even under substantial source pose errors, with visual quality largely preserved. Nevertheless, geometric consistency and camera accuracy generally improve with more accurate source poses.


We hope these additions address the reviewer's concerns. All changes are highlighted in blue in the revised manuscript. We welcome any further questions.

Reviewer_L6yBOpenReview ↗

Response to the rebuttal

Thanks to the authors for the detailed response. The matched source-versus-target alignment comparison and compute-matched baselines substantially address my concerns about RETA’s contribution and the effect of additional training. The new viewpoint-consistency analysis also strengthens the empirical motivation.

The proposed causal link between camera stability and reduced trajectory curvature remains only partially established, but the additional evidence supports the method’s effectiveness. Given the practical acceleration and demonstrated visual quality and camera-control performance, I continue to support weak acceptance (3). I have no further experimental requests.

We are glad that our responses helped address your concerns. Thank you for your thoughtful reassessment and continued support.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.