- 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
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 Token | Aesthetic Quality↑ | Imaging Quality↑ | Dyn-MEt3R↑ | MEt3R↓ | TransErr↓ | RotErr↓ |
|---|---|---|---|---|---|---|
| Target-stream | 0.5165 | 0.6519 | 0.8320 | 0.3238 | 0.0159 | 1.583 |
| Source-stream (RETA) | 0.5214 | 0.6625 | 0.8491 | 0.3059 | 0.0143 | 1.486 |
[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.
| Method | Stage 1 | Stage 2 | Distillation | GPU-h | Aesthetic Quality↑ | Imaging Quality↑ | Dyn-MEt3R↑ | MEt3R↓ | TransErr↓ | RotErr↓ |
|---|---|---|---|---|---|---|---|---|---|---|
| Original training budget | ||||||||||
| DMD2 | 20k | – | DMD2, 5k | 894 | 0.5166 | 0.6629 | 0.8447 | 0.3190 | 0.0129 | 1.544 |
| w/o Stage 2 | 20k | – | Stage 3, 5k | 968 | 0.5132 | 0.6312 | 0.8444 | 0.3215 | 0.0128 | 1.294 |
| Compute-matched to FlashRender | ||||||||||
| DMD2 | 20k | – | DMD2, 21.2k | 1,960 | 0.5158 | 0.6625 | 0.8450 | 0.3177 | 0.0129 | 1.493 |
| w/o Stage 2 | 55.2k | – | Stage 3, 5k | 1,960 | 0.5120 | 0.6505 | 0.8475 | 0.3093 | 0.0133 | 1.299 |
| FlashRender | 20k | 20k | Stage 3, 5k | 1,960 | 0.5182 | 0.6654 | 0.8571 | 0.3071 | 0.0122 | 1.236 |
[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.
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), X1 is the initial noise z1 and X0 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∈T and sampling budget B∈B=4,8,12,20, we estimate a homography HtB 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 HtB is the identity matrix. From each HtB, 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(⋅), we average over sampling budgets and frame indices within each case ∣B∣,∣T∣1∑B∈B,,t∈Tm(HtB), 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.
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