acceptodds
Under review as a conference paper at ICLR 2027

Alpha as an Efficiency Signal: Visibility-Routed RGBA Image-to-Video Generation

Abstract

RGBA videos combine RGB appearance with an alpha channel, enabling animated assets to be applied across arbitrary backgrounds, which are heavily used in gaming industry. However, generating high-quality RGBA animations for games remains challenging for two reasons. First, most existing RGBA video datasets are dominated by photorealistic content, with limited coverage of game assets. Second, the traditional generate-then-matte pipelines estimate alpha only after RGB synthesis, so semi-transparent regions are often blurred by background, resulting in unstable matting outputs. More recently, many approaches begin to model RGB and alpha jointly, but existing approaches are mostly text-conditioned, and still have unresolved issues in efficiency and quality. To address these challenges, we introduce GameAlpha-2.4K, a 2.4K-clip game-style RGBA video dataset built with matte-friendly synthesis, multi-hypothesis alpha recovery, and compositing-based quality gates. Using this dataset, we train a reference-conditioned RGBA video generator that jointly produces RGB frames and alpha mattes in a single pass. To improve efficiency, we propose a visibility router that identifies transparent tokens in an early stage and bypasses their later DiT updates, while -lock guides them along the original flow-matching schedule toward self-predicted endpoints. Our model obtains lower FVD than traditional two-stage pipelines, and the visibility router skips 35% of token evaluations in the final two DiT denoising steps, providing a 1.2× backbone speedup with negligible quality degradation compared to dense inference.

Primary area:
generative models
Keywords:
RGBA video generation, diffusion transformers, flow matching, alpha matte generation, video datasets

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_Q4JBOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper achieves efficient RGBA video generation under reference image and text conditioning, aiming to extend existing video diffusion models to support transparent video synthesis. The paper first constructs GameAlpha-2.4K, a dataset containing 2,395 game-style video samples, where alpha mattes are generated using existing alpha estimation models and further filtered through multimodal model-based quality assessment. Building upon the Wan architecture, the paper introduces an RGBA fusion module and a DoRA-based adaptation module to expand the original video generation model into the RGBA generation space. To reduce redundant DiT computation caused by transparent regions, the paper further designs a transparency prediction probe that identifies regions likely to become fully transparent and skips subsequent computation accordingly, achieving approximately 1.2× inference acceleration.

Critical strengths

The proposed transparency routing strategy is simple yet effective. By reducing the number of DiT tokens involved in the final denoising stages, it achieves a 1.2× inference speedup with minimal additional complexity.

Critical weaknesses
  1. The comparison in Figure 2 only focuses on the visual difference between post-generation matting and direct RGBA generation, but it does not evaluate how the proposed method compares with existing approaches that jointly generate RGB and alpha channels. The paper lacks direct visual comparisons demonstrating the quality of semi-transparent boundaries, where subtle artifacts are often difficult to capture through quantitative metrics alone. Therefore, the authors should include comparisons with existing RGB-alpha joint generation methods to better demonstrate the visual advantages of the proposed approach.
  2. The performance gain is relatively minor, and the overall contribution appears limited.
Questions for authors

Could the authors clarify how the proposed method compares with existing approaches that jointly generate RGB and alpha channels?

2 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer Q4JB (W1/W2/Q1)

We sincerely thank the reviewer for their time and effort in reviewing our work, as well as for their constructive and valuable feedback.

W1/Q1

We thank the reviewer for pointing out the lack of a comparison with an independent direct RGBA generation method.

Based on our survey at the time of submission, we had not identified a reference-image-conditioned RGBA video generation method with both runnable inference code and model weights publicly available. Our original experiments therefore primarily compared against generate-then-matte pipelines. We would be happy to include comparisons with any publicly reproducible methods that we may have overlooked. We subsequently became aware of Wan-Alpha++, released on 2026-10-19 after our submission, and included its officially released model in our comparisons.

To further compare the two joint-generation systems on real-world human videos, we conduct experiments on VideoMatte240K. We use the official implementation, weights, and configuration of Wan-Alpha++, comparing it with our Dense and Routed variants. No additional fine-tuning or parameter selection is performed on this dataset.

Table 1: Quality comparison with Wan-Alpha++, an independent direct RGBA generation method, on VideoMatte240K.

MethodFVD ↓DINO-ID ↑Aesthetic ↑MotionSm ↑FlowDiff ↓MAD×1000 ↓
Wan-Alpha++155.20.8714.9960.9900.638124.0
Ours Dense129.70.9055.1190.9940.624105.0
Ours Routed125.70.9065.1070.9940.621105.2

As shown in the table, our Routed model achieves an FVD of 125.7, lower than Wan-Alpha++'s 155.2, while DINO-ID improves from 0.871 to 0.906. Aesthetic, Motion Smoothness, Flow Difference, and MAD scores are also better. Our system achieves competitive generation quality on VideoMatte240K while using fewer sampling steps than Wan-Alpha++.

Additionally, we provide visual comparisons of RGB and alpha outputs on VideoMatte240K via an anonymous link, showing results from our Dense and Routed variants and Wan-Alpha++. Corresponding detail regions are highlighted to facilitate visual comparison of subject appearance, hair and prop structures, and the completeness of the alpha mattes.

W2

Thank you for recognizing the simplicity and effectiveness of our method.

Our contributions include a dataset and a joint-generation approach for reference-conditioned RGBA video generation, as well as a routing method that uses predicted final visibility to reduce subsequent computation. We validate the effectiveness of these two aspects through comparisons with independent joint-generation methods and ablations with a fixed generator, respectively.

Complete generation system. We construct GameAlpha-2.4K and adopt a merged RGB–Alpha representation for reference-conditioned joint generation. TransPixeler[2] extends the generation sequence by adding Alpha tokens, and its supplementary material also discusses the resulting computational overhead; the merged representation we adopt avoids this additional sequence overhead. To evaluate the practical performance of the complete approach, we have added a comparison with the publicly released weights of Wan-Alpha++[1]. The full quantitative results and visual comparisons are provided in W1. These results provide direct evidence of our system's performance relative to existing joint-generation methods.

Additional gains from routing. Using the same generator, we further compare results with routing enabled and disabled. With the weights, inputs, and four-step sampling configuration held fixed, visibility routing reduces the mean end-to-end runtime from 125.9 seconds to 104.8 seconds, with FVD changing from 172.3 to 174.4 (Table 2 of the paper). Routing gains also vary with content. As shown in Table 2, under fixed routing parameters, the groups of samples with transparent-region proportions of 40%–60% and 80%–100% achieve end-to-end speedups of 1.12× and 1.26×, respectively, indicating that greater reductions in actual computation are possible when transparent regions occupy a larger proportion. Regarding the independent contribution of visibility as the selection criterion, the score-replacement experiment further shows (Table 2 in our response to YPyn) that, while keeping RAS's computational budget and execution mechanism unchanged, using predicted visibility scores alone substantially reduces output deviation relative to Dense; see the score-ablation table for details.

Table 2: End-to-end speedup across transparent-region proportions.

Proportion of transparent regions in the videoEnd-to-end speedup ↑
40%–60%1.12×
60%–80%1.19×
80%–100%1.26×
AuthorsOpenReview ↗

Rebuttal to Reviewer Q4JB

References

[1] Wan-Alpha++: Animating a Single RGBA Asset into Controllable Transparent Video Effects, arXiv:2610.26189, 2026.

[2] TransPixeler: Advancing Text-to-Video Generation with Transparency, CVPR 2025.

Reviewer_oSS3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces a dataset of RGBA videos focused on game assets, and introduces a method for RGBA reference-to-video generation which relies on early estimates of transparency. The key idea is to re-use existing RGBA VAE technique (Wan-Alpha), fine-tune underlying Wan I2V DiT using DoRA, and use an intermediate x0 prediction to identify tokens which will be transparent, and skip their updates to save compute. Method is validated against several two-stage pipelines, achieving competitive quality and 1.2x speedup.

Critical strengths
  • This work targets an important problem of RGBA video generation, and proposes a single pass solution that simultaneously achieves competitive results while being more efficient than two-stage pipelines. This is consequential because producing matting is very frequently needed in practice, and existing two-stage pipelines typically lead to unnecessary overhead and struggle with ambiguous regions.
Critical weaknesses
  • The domain of game assets seems to be quite narrow. If the method is generic enough, I am not fully sure why the restriction, especially since the dataset is created with an automated pipeline. It is consequential because RGBA generation of realistic scenes would be extremely useful, and the claimed novelty of the domain in fact is a limitation of the proposed method (and more specifically its limited validation).
  • This work reuses some of the components (Wan-Alpha VAE, Wan I2V conditioning, DoRA), and the main addition is about using a x0 prediction (which is an existing parameterization) to identify transparent tokens and save compute on those. This is consequential because in this context the technical novelty of the work is unclear, and the gains of this addition are relatively modest (1.2x speedup with a slight drop in video quality).
Questions for authors
  • Why restriction to game assets? Wouldn't the same data generation pipeline and the method itself work on arbitrary domains?
  • Is the DoRA built on top of the base model, or on already distilled model? If it's on distilled, was there any special care necessary compared to full undistilled models?
  • It might be worth discussing relationship to models that allocate compute adaptively to the tokens or use varying noise levels, which provide a more generic way of allocating compute.
3 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer oSS3 (W1/Q1)

We sincerely thank the reviewer for their time and effort in reviewing our work, as well as for their thoughtful and constructive feedback.

W1/Q1

Thank you for raising the question of applicability beyond game assets.

We focus our dataset construction on game assets because game animation production requires RGBA assets that can be directly composited onto different backgrounds, while existing RGBA video datasets provide limited coverage of such content. GameAlpha was therefore constructed to address this practical need. Its synthesis and alpha-recovery pipeline is not restricted to game assets and can be adapted to other foreground content by adjusting the content sampling and quality-screening settings. For our generation method, we allocate computation based on predicted final visibility and do not introduce any game-specific routing rules.

Our experimental validation already covers photorealistic OOD data. Appendix C.7/Table 6 presents competitive results on VideoMatte240K compared with generate-then-matte pipelines, while Appendices D.4 and D.5 provide paired dense/routed generation and compositing examples on multiple backgrounds for photorealistic human subjects. Building on these results, we fixed the model and routing settings and conducted additional paired quality and latency evaluations on the same set of inputs, without any additional training. The average end-to-end latency decreased from 119.6 s to 100.7 s, yielding a 1.2× speedup.

Table 1: Quality and end-to-end latency of dense and visibility-routed inference on VideoMatte240K.

Time↓FVD ↓DINO-ID↑Aes↑Motion Sm.↑Flow-Diff↓MAD↓
Dense119.6 s129.70.9055.1190.9940.624105.0
visibility router100.7 s125.70.9065.1070.9940.621105.2

Thus, game assets are the focus of our dataset construction, but not the sole scope of our method's validation. The existing quality comparisons and the additional paired experiments jointly demonstrate that the fixed visibility router still provides practical latency gains for photorealistic human foregrounds.

AuthorsOpenReview ↗

Rebuttal to Reviewer oSS3 (W2/Q2)

W2

We thank the reviewer for the suggestion.

We adopt Wan-Alpha's RGBA representation, Wan's native I2V conditioning, and DoRA. As stated in the contributions section of the main paper, this work constructs GameAlpha-2.4K for reference-image-conditioned RGBA video generation and proposes a method that uses predicted final visibility to allocate subsequent DiT computation. We use standard x0x_0 prediction as the source of information; the new contribution is to identify ultimately transparent regions from early generation states and translate this prediction into a computation allocation policy.

This policy has a basis specific to the RGBA task: the RGB values of regions that remain transparent in the final output make no direct contribution to the composite. During generation, however, a region that is currently transparent may not remain so in the final output, since motion and foreground expansion can change the visible regions. It is therefore necessary to predict final visibility and reduce the impact of erroneous skipping through spatiotemporal safety margins, attention-context retention, state advancement, and local reactivation. The problem is to determine, before generation is complete, which regions can receive less subsequent computation while maintaining the generation process under sparse execution.

The main distinction from RAS[1] lies in the basis for allocating computation. RAS selects regions to update using statistics of predicted noise and skipping history, whereas we explicitly predict the visibility of regions in the final RGBA output. To test the independent value of this criterion, we retain RAS's execution procedure, per-step token budgets, caching, and state-update mechanism, and replace only the region-selection score. FVD decreases from 218.8 to 186.8, while the Alpha and compositing deviations from Dense are reduced by 45.2% and 41.1%, respectively, with essentially unchanged latency (YPyn, Table 2). This controlled result shows that predicted final visibility provides useful information for computation allocation.

Regarding the magnitude of the gains, our method further reduces computation in a generator that already uses a merged latent representation and four-step sampling. The additional quality–latency experiments show that, for configurations saving approximately the same amount of end-to-end time, our method increases FVD by 2.1, whereas RAS at a similar latency increases FVD by 46.5 (YPyn, Table 1). This clarifies the practical value of our method: visibility-driven computation allocation achieves acceleration at a smaller FVD cost.

In addition, we evaluated the method with Wan2.1's native 50-step sampling and a four-step Wan2.2 configuration. These settings achieve denoising speedups of 1.19× and 1.17×, respectively, with FVD changes of −0.73 and +0.45 relative to their respective Dense baselines (YPyn, Table 4). These results extend the evidence of effectiveness to an additional backbone and native multi-step sampling.

In addition, we compare our method with the official Wan-Alpha++[2] model on VideoMatte240K. Our Routed variant achieves a lower FVD (125.7 vs. 155.2) and a higher DINO-ID, indicating that our complete generation system is also competitive in this evaluation on real-world human videos. Detailed experimental settings and comparison results are provided in our response to Q4JB’s W1, with quantitative results summarized in Table 1 therein.

Q2

Thank you for requesting clarification of the DoRA training setup.

We train DoRA on the original, undistilled Wan2.1-I2V-14B base model, keeping the base DiT parameters frozen and using the timestep-weighted flow-matching objective in Eq. (5) of the paper. The pretrained LightX2V distillation adapter is introduced only at inference: we first merge LightX2V into the base weights and then apply the trained DoRA parameters. Thus, we do not continue training DoRA on an already distilled model, nor do we introduce a distillation-specific training objective.

We will clarify this training and inference procedure in the main text of the revision and specify the order in which the adapters are combined.

AuthorsOpenReview ↗

Rebuttal to Reviewer oSS3 (Q3)

Q3

We thank the reviewer for the suggestion.

We agree that the relationship between our work, token-level adaptive computation, and methods that model different noise levels should be discussed more explicitly.

For token-level adaptive computation, RAS[1] selects regions for recomputation based on predictions from the preceding step and reuses cached predictions elsewhere. Our work shares the goal of allocating computation nonuniformly, while our focus is on the basis for computation allocation in RGBA generation: predicting whether a region will be visible in the final output. When the final alpha is zero, the corresponding RGB does not directly affect the composite; therefore, final visibility provides a selection signal directly tied to the task output.

To test the value of this signal relative to generic statistical scores (YPyn, Table 2), we use the same set of test inputs and retain RAS's execution procedure, per-step token budgets, K/V caching, and cached-velocity state updates, replacing only the region-selection score. With predicted visibility scores, FVD decreases from 218.8 to 186.8, while Alpha and compositing MAE relative to the paired Dense outputs decrease by 45.2% and 41.1%, respectively, with essentially unchanged end-to-end latency. This shows that, in this RGBA setting, predicted final visibility can provide useful additional information to a general token-level computation allocation mechanism.

For methods involving different noise levels, Diffusion Forcing[3] and its video extension DFoT[4] are trained to support different noise levels at different sequence positions or frames, enabling more flexible sampling schedules. Our method instead preserves the generator's original global noise schedule and reduces network computation for selected tokens based on predicted visibility; skipped tokens continue to advance along this schedule through x0x_0-lock. Ablations with fixed routing and computation budgets (YPyn, Table 3) show that holding states unchanged substantially increases deviations from Dense, whereas x0x_0-lock better preserves Dense outputs. This supports the design of continuing to advance states while reducing network computation.

Thus, our contribution centers on using final visibility in RGBA to guide computation allocation. We will expand the discussion of these relationships and the scope of applicability in the revision.

References

[1] Region-Adaptive Sampling for Diffusion Transformers, arXiv:2502.10389, 2025.

[2] Wan-Alpha++: Animating a Single RGBA Asset into Controllable Transparent Video Effects, arXiv:2610.26189, 2026.

[3] Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion, NeurIPS 2024.

[4] History-Guided Video Diffusion, ICML 2025.

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

The paper studies reference-conditioned RGBA video generation for compositable game-asset animation. It introduces GameAlpha-2.4K, adapts a merged RGB–alpha latent representation to Wan2.1-I2V, and proposes visibility-based routing to reduce computation during four-step inference. Tokens predicted to be transparent bypass later DiT updates, with x₀-lock, cached attention context, and local reactivation supporting sparse execution. Experiments compare against generate-then-matte systems and report a 1.20× end-to-end speedup over the dense generator with a small degradation in FVD.

Critical strengths

A useful application with substantial implementation effort. Generating compositable animated game assets is practically relevant. The work brings together dataset construction, RGBA generation, and inference acceleration into a coherent pipeline. The staged data screening and attention to temporal consistency reflect considerable engineering effort.

An interesting routing signal supported by several evaluations. Using predicted transparency to allocate computation is intuitive and task-relevant. The paper includes matched dense-versus-routed comparisons, multi-seed results, reactivation ablations, cross-domain evaluation, and a blinded human study. These provide useful evidence that routing can reduce runtime while largely retaining the dense model’s quality.

Critical weaknesses
  1. The methodological advance appears limited. Joint RGB–alpha generation and the merged latent representation build substantially on prior work, while native I2V conditioning and DoRA adaptation use established components. Visibility routing is the most distinctive contribution, but its advantage over existing regional sampling and caching approaches is not sufficiently established. The experiments use a RAS-style proxy rather than a faithful implementation, and do not isolate the benefit of x₀-lock against simpler state-update alternatives.
  2. The broader significance of the acceleration remains uncertain. The reported 1.20× speedup is useful, but is demonstrated only on one backbone with a four-evaluation schedule. Cross-domain evaluation is encouraging, yet does not establish transfer across architectures or sampling configurations. Routing also introduces a small FVD degradation relative to dense inference. A broader quality–latency comparison would help establish when the proposed trade-off is preferable to existing acceleration methods.
  3. Generation comparisons do not cover the closest methodological alternatives. The main comparisons evaluate generate-then-matte pipelines, without an independent direct RGBA generator. These results support the proposed system over the selected two-stage baselines, but leave its relative performance against existing joint-generation approaches unclear.
  4. Dataset quality and availability require further clarification. The planned public release is welcome, particularly because the dataset is a central contribution. However, release scope and redistribution permissions remain to be finalized. More importantly, the construction pipeline estimates alpha from generated RGB videos without clearly explaining whether background-independent foreground colors are recovered. This matters for semi-transparent regions and reliable compositing onto new backgrounds.
  5. Some evaluation and reproducibility details could be strengthened. MAD is sensitive to differences in generated motion, while checkerboard-based metrics and ratings do not directly establish arbitrary-background compositing quality. Region-specific evaluation of soft transparency and boundaries would be informative. The timing hardware and identities or versions of the generation and screening models should also be specified.
Questions for authors
  1. Under matched routing masks and computation budgets, how does x₀-lock compare with cached-velocity updates or other simple state-advancement alternatives?

  2. How does the method compare with a faithful regional sampling or caching baseline? Is there evidence that the quality–latency benefit transfers beyond Wan2.1 and four-step sampling?

  3. Can an existing direct RGBA generator be adapted for a comparable evaluation? If this is infeasible, which specific limitations prevent a meaningful comparison?

  4. Are the dataset RGB targets recovered foreground colors or the original generated composites? Please clarify any color decontamination and the alpha convention used.

  5. What artifacts are intended for public release, and have permissions for redistributing the synthetic data been confirmed? Please also specify the timing hardware and dataset-generation models.

4 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer YPyn (W1/Q1)

We sincerely thank the reviewer for their time and effort in reviewing our work, as well as for their constructive and valuable feedback.

W1/Q1

We thank the reviewer for the suggestion.

Although we adopt an established RGBA representation, native I2V conditioning, and DoRA adaptation from prior work, as summarized at the end of Section 1 of the main paper, our core contributions are:

  1. A task and dataset for game assets. We study reference-image-conditioned RGBA video generation and construct GameAlpha-2.4K.
  2. Allocating computation based on predicted final visibility. We predict the final transparent regions from early denoising states and reduce subsequent DiT computation in these regions, while maintaining the generation process through attention-context preservation, state advancement, and local reactivation.
  3. Validating generation quality and practical efficiency. Compared with generate-then-matte baselines using the same backbone, our joint generation model achieves lower FVD; compared with its own full-computation version, visibility routing achieves a 1.20× speedup while maintaining quality metrics close to those of full computation.

Together, these three contributions support high-quality, low-cost reference-conditioned RGBA video generation.

Comparison with a Faithful RAS Implementation

To supplement the comparison with a faithful RAS implementation, we adapt the authors' publicly available code to Wan, retaining the original scoring, dynamic reselection, prediction reuse, and K/V update mechanisms. The additional evaluation uses the same model, inputs, and four-step sampling setup as the paper. RAS parameters are selected on the training set and held fixed during evaluation. The table below compares the RAS settings with our method.

Table 1. Generation quality and runtime of our method and RAS adapted to Wan based on the authors' implementation. RAS (20% / 35%) denotes the proportion of tokens skipped at each sparse denoising step.

MethodFVD ↓End-to-End Time (s) ↓
Dense172.3125.9
RAS (20%)275.1112.0
RAS (35%)218.8105.5
Ours174.4104.8

With this comparison based on the authors' RAS implementation, at similar end-to-end runtimes (104.8 seconds for our method and 105.5 seconds for RAS), our method achieves an FVD of 174.4, lower than RAS's 218.8 and close to Dense's 172.3. This indicates that, even when RAS's original acceleration mechanisms are retained, our method better preserves generation quality within approximately 105 seconds of generation time. This additional comparison also supports the advantage observed previously.

We additionally provide a visual comparison of Alpha details Link. The three columns show the full-computation Dense baseline, RAS (35%), and our method, respectively, using the same frames and regions from the same samples. RAS (35%) exhibits weakened small droplets, fragmented halo boundaries, and hammer-handle deformation and trailing artifacts. Our method better preserves these fine structures, semitransparent transitions, and object contours, producing results more consistent with Dense. These examples further demonstrate our method's ability to preserve Alpha structures and details while reducing computation.

RAS allocates regional computation based on statistics of the predicted noise, but such scores do not directly determine whether a region will be visible in the final RGBA output. For RGBA generation, RGB values in the final transparent regions do not directly affect the composited result. Therefore, predicting final visibility can provide a more direct basis for computation allocation. To verify this, we retain the RAS computation procedure and per-step token budgets, replacing only the region-selection score. The predicted visibility score reduces FVD from 218.8 to 186.8 and Alpha MAE relative to Dense from 0.0127 to 0.0070, with essentially unchanged runtime. This indicates that, for this task, explicitly using final visibility can allocate computation more effectively than a generic statistical score.

Table 2: Alpha and compositing deviations are measured as MAE relative to the paired Dense outputs.

Selection ScoreFVD ↓Alpha Deviation ↓Compositing Deviation ↓End-to-End Time (s) ↓
RAS218.80.01270.024105.5
RAS + Our Visibility Score186.80.00700.014105.4
AuthorsOpenReview ↗

Rebuttal to Reviewer YPyn (W1/Q1/W2/Q2/W3/Q3)

State-Update Experiment: x₀-lock and Simple Alternatives

To verify the benefit of x₀-lock over simple state-update approaches, we hold the routing mask, caching mechanism, and computation budget fixed, changing only how the states of skipped tokens are updated. We compare x₀-lock with keeping the states unchanged. The results are shown below.

Table 3. Deviations of different state-update approaches from Dense outputs. All metrics are MAE; lower values indicate outputs closer to Dense.

State-Update MethodAlpha MAE ↓White-Background Compositing MAE ↓Black-Background Compositing MAE ↓Checkerboard Compositing MAE ↓
x₀-lock0.0041150.0099460.008530.009273
Keeping States Unchanged0.1735410.1076470.0805420.073885

Under the same routing and computation budget, x₀-lock better preserves the Dense outputs than keeping the states unchanged, indicating that states still need to advance after network computation is skipped.

W2/Q2

Thank you for suggesting broader evaluations of transferability and quality–latency trade-offs.

To examine the scope of the acceleration benefits and the quality–latency trade-off relative to region-sampling methods, we conduct the following two sets of experiments.

First, do the acceleration benefits extend to an additional backbone and different sampling configurations?

We add the native 50-step UniPC configuration of Wan2.1-I2V-14B (with CFG enabled) and the native four-step configuration of Wan2.2-TI2V-5B-Turbo. Each configuration is compared against its own full-computation version (Dense) in a paired evaluation. Inference is performed on the same test set, with results as follows:

Table 4: Results on an additional backbone and native sampling configurations.

ConfigurationDenoising Time, dense→routedSpeedupFVD, dense→routed
Wan2.1, native 50-step sampling2216.9→1862.81.19 ×181.17→180.45
Wan2.2 Turbo, four-step sampling4.34→3.721.17×266.02→266.47

Under these two additional configurations, visibility routing achieves denoising speedups of 1.19× and 1.17×, respectively, with small FVD changes relative to their corresponding Dense versions. This indicates that the acceleration benefits are not limited to the original four-step Wan2.1 setting, but also extend to an additional backbone and different sampling configurations, reducing denoising time with small changes in FVD.

Second, after accounting for quality loss, when is our method's trade-off more advantageous?

Under the original four-step Wan2.1 setting, we additionally compare three spatial protection ranges of our method with RAS configurations.

Table 5: Quality–latency comparison with RAS.

MethodFVDTime (s)
Dense172.3125.9
Ours r=1177.698.8
Ours r=2 (this paper)174.4104.8
Ours r=3176.3114.9
RAS (20%)275.1112.0
RAS (35%)218.8105.5

The default configuration achieves a favorable balance between generation quality and inference efficiency. Compared with Dense, it reduces runtime by 21.1 s while increasing FVD by only 2.1; compared with RAS 35%, which has a similar runtime, it achieves substantially lower FVD (174.4 vs. 218.8). Therefore, the default configuration is suitable for scenarios that seek to retain generation quality close to Dense while obtaining practical acceleration, balancing quality and efficiency among the compared configurations.

W3/Q3

We thank the reviewer for pointing out the lack of a comparison with an independent direct generation method.

At the time of submission, to the best of our knowledge, no direct RGBA I2V method had publicly released both runnable inference code and model weights, so our original experiments primarily compared against generate-then-matte pipelines. If we have overlooked any publicly reproducible method, we welcome the reviewer's suggestions and are willing to add the corresponding comparison. To supplement this comparison, we include Wan-Alpha++(Dong et al., 2026), released on 2026-10-19 after our submission, using its official weights as an independent RGB–Alpha joint generation baseline. Detailed experimental settings and comparison results are provided in our response to Q4JB’s W1, with quantitative results summarized in Table 1 therein.

AuthorsOpenReview ↗

Rebuttal to Reviewer YPyn (W4/Q4)

W4/Q4

We thank the reviewer for pointing out that the foreground colors and alpha representation require further clarification.

Our data pipeline first generates an RGB video C0C_0 on a fixed background color B0B_0, then estimates alpha and recovers the foreground colors using the known background color:

F^=C0−(1−α^)B0α^,α^>0.\hat{F}=\frac{C_0-(1-\hat{\alpha})B_0}{\hat{\alpha}},\qquad \hat{\alpha}>0.

Here, α^\hat\alpha is the estimated alpha, and F^\hat F is the recovered foreground-color estimate. The recovered F^\hat F is paired with the original alpha to form the RGBA data.

We are completing the work required for the public release of the full GameAlpha-2.4K dataset. We will retain the original RGB, alpha, recovered foregrounds, and processing documentation to support reproduction of the original manuscript and subsequent use. Further release details are provided in W5/Q5.

AuthorsOpenReview ↗

Rebuttal to Reviewer YPyn (W5/Q5)

W5/Q5

We thank the reviewer for raising these concerns.

MAD is affected by differences in generated motion and should therefore be interpreted alongside other evaluation metrics, rather than as a standalone measure of alpha estimation accuracy.

To further evaluate soft transparency, boundary artifacts, and visual quality after compositing, we conducted two additional blinded evaluations with a dedicated data annotation team, using the same 20 cases as in the original human evaluation and covering our method and three generate-then-matte baselines.

  1. Soft transparency and boundary assessment. Ten annotators independently evaluated the checkerboard composites for all cases, focusing on residual backgrounds, bright or dark edge artifacts, preservation of soft transparency, and temporal boundary stability.

  2. Blinded evaluation across multiple backgrounds. For each method, we composited the same original RGB and alpha outputs onto white, black, and blue backgrounds. The three versions differed only in the compositing background, with no regeneration or additional post-processing. Ten annotators were divided into groups to evaluate these composites.

Both evaluations concealed method identities. For each video, annotators answered three yes/no questions: whether soft transparency was naturally preserved, whether the boundaries flickered over time, and whether the result was usable overall.

As shown in Tables 6 and 7, our method outperforms the compared baselines on all three criteria. These results provide human evaluation evidence addressing the concerns in W5 regarding soft transparency, boundary stability, and compositing quality beyond checkerboard backgrounds.

Table 6. Blinded human evaluation of soft transparency preservation and boundary quality.

MethodNatural preservation of soft transparency ↑Boundary flicker ↓Overall usability ↑
Wan2.169.533.061.5
CogVideoX1.536.074.511.5
HunyuanVideo37.562.514.0
Ours79.022.075.5

Table 7. Aggregated results of blinded human evaluation on black, white, and blue backgrounds.

MethodNatural preservation of soft transparency ↑Boundary flicker ↓Overall usability ↑
Wan2.166.6728.8956.67
CogVideoX1.535.5666.118.89
HunyuanVideo32.7857.2212.22
Ours80.0018.8970.00

Due to institutional information-disclosure requirements, we are currently unable to disclose the specific hardware model or the complete identities of some generation and screening models, but we can provide verifiable experimental configurations and processing procedures. Inference uses a single device, a batch size of 1, and BF16 precision. The device's theoretical peak dense BF16 Tensor Core throughput is 148 TFLOPS per device, excluding gains from structured sparsity. Dense and routed runs are measured in pairs on the same physical device. End-to-end timing includes conditioning encoding, denoising, RGB/Alpha decoding, video encoding and disk writing, and all routing overhead. For data construction, we can specify the responsibilities and screening criteria of each module: the text-conditioned image generation module produces 1920×19201920\times1920 reference images; the image-to-video module generates animations based on the reference images and action descriptions; and the multimodal screening module checks subject completeness, consistency of attributes and actions, background interference, transparency boundaries, and temporal stability in stages. When a reference image fails the checks, the feedback is converted into prompt revisions, with at most three attempts. Finally, a predefined 10% subset is independently inspected by human annotators.

Due to institutional approval requirements for public release, we are not yet able to make the complete dataset, code, and model weights publicly available, but we are pursuing release approval for all three, including a review of the permission to redistribute the synthetic data. During the approval process, we can still provide the data construction and screening procedures, experimental configurations, data statistics, anonymized examples, and weights to support the reviewers in verifying the experimental process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.