acceptodds
Under review as a conference paper at ICLR 2027

BRAVER: Animating RGB References into Reusable RGBA Videos

Abstract

Creators often begin with an image or video that captures the appearance or motion they want to preserve. Turning these references into reusable animated assets requires retaining the intended foreground while separating it from its original background, including through fine boundaries and translucent regions. This calls for both a faithful transparency representation and reference conditioning that distinguishes the subject from its surroundings. We introduce BRAVER (Background-Robust Animation of Visual Elements for Reuse), a matte-free framework for text-conditioned generation, image-guided animation, and aligned foreground extraction from RGB videos. BRAVER combines a pretrained video prior with a compact RGBA autoencoder. Regional alpha supervision complements reconstruction over opaque interiors, boundaries, and translucent regions. Two-stage adaptation establishes an RGBA generation prior and incorporates visual inputs through shared adaptation and task modulation. Training references are composited over varied backgrounds while their original RGBA assets remain the targets, providing supervision for separating the foreground from its surroundings. The framework requires neither user-provided mattes nor a separate output-matting stage. We also curate AlphaElements, an RGBA video collection spanning animated stickers, motion graphics, and visual effects, to support reconstruction and conditional generation. BRAVER outperforms the evaluated baselines in aesthetic quality and naturalness for text-conditioned generation, and reference consistency for image-guided animation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
RGBA video generation, transparent video generation, diffusion transformers, alpha reconstruction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ANKYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes BRAVER, a framework that generates RGB videos given text or RGB conditions. To achieve this goal, BRAVER first trains a VAE that fuses the RGB and alpha latents into a joint RGBA Latent and reconstructs RGB and alpha simultaneously. Several losses are adopted to achieve the joint learning of RGB and alpha. As for the DiT part, BRAVER uses a pre-trained WAN DiT with DoRA. Experiments show that it outperforms other baselines.

Critical strengths
  • Based on the Wan-Alpha VAE, this paper designs a set of new losses to improve reconstruction performance.
  • A unified framework using background-composited reference training successfully handles multi-task learning, such as text- and image-conditioned generation and video foreground extraction. Several compared methods are evaluated, and BRAVER gives the best results.
Critical weaknesses
  • The technical contribution of this paper is limited. The VAE architecture is from Wan-Alpha, while the proposed losses are straightforward. The DiT design (DoRA, task embedding, and conditioning) is mostly an engineering combination of existing methods.
  • The experiments are insufficient. Alpha quality is only measured with a human study; stronger quantitative evaluations are needed. Table 3 only compares 100 examples, while the proposed method marginally outperforms Wan2.1-T2V + MatAnyone, making me doubt whether the evaluations should be more thorough (maybe variations across seeds and confidence intervals should be given). The V2RGBA task, which is interesting and a claimed core task, has not been well evaluated.
Questions for authors

The techinical novelty and experiments should be enhanced, as noted in the weaknesses section.

5 comments

We thank the reviewer for the careful reading, and for recognizing the role of FADE in reconstruction quality, of background-composited reference training for multi-task learning, and of BRAVER's performance over several baselines. We respond point by point.

W1 & Q1. Limited technical contribution

Our contribution is a unified RGBA generation formulation in which the three tasks train together and improve each other. BRAVER handles text generation, RGB-image animation, and RGB-video asset extraction with one native RGBA latent model that produces foreground color and opacity together. Generation and matting therefore happen in one latent, so the asset is reusable as generated and the pipeline needs no separate output-matting model.

The three components are built for this latent. Each addresses one specific problem:

  • Reference background: it can leak into the generated asset, and DEBAR decouples the two.
  • Local alpha: errors concentrate in semi-transparent and interior regions, which FADE and opaque-core supervision target.
  • Three input types: text, image, and video conditions differ in form, and shared conditioning (visual-condition projection, task modulation, shared DoRA) maps all three into one backbone.

Joint training. We train T2RGBA-only, I2RGBA-only, and V2RGBA-only counterparts under the same budget; all three are complete. On T2RGBA, joint training is higher on Text Alignment, Aesthetic, and Naturalness (4.25/0.5969/4.87 vs. 4.04/0.5697/4.72), with slightly lower Smoothness and Flicker (0.9928/0.9919 vs. 0.9932/0.9920). On I2RGBA, it is higher on all four metrics (Aesthetic 0.6266 vs. 0.6176, Imaging 0.5784 vs. 0.5739, Subject Consistency 0.9811 vs. 0.9665, reference-subject preservation 0.8560 vs. 0.8525). On V2RGBA, all five alpha errors are lower (MAD 36.859 → 27.677). On the sticker and effect subsets, alpha MAD is 5.667 and 49.687, versus 10.685 and 63.033 for V2RGBA only.

Table 1. T2RGBA joint vs. single-task training.

MethodText Alignment ↑Aesthetic ↑Naturalness ↑Smoothness ↑Flicker ↑
T2RGBA only4.040.56974.720.99320.9920
BRAVER (joint)4.250.59694.870.99280.9919

Table 2. I2RGBA joint vs. single-task training.

MethodAesthetic ↑Imaging ↑Subject Consistency ↑Reference Preservation ↑
I2RGBA only0.61760.57390.96650.8525
BRAVER (joint)0.62660.57840.98110.8560

Table 3. V2RGBA joint vs. single-task training (500 cases).

MethodMAD ↓MSE ↓GRAD ↓CONN ↓DTSSD ↓
V2RGBA only36.85921.70311.36314.0075.043
BRAVER (joint)27.67713.9807.33810.2183.806

The T2RGBA row of BRAVER is taken from Table 3 of the paper (100 text prompts, white-background composites); the I2RGBA row reports VBench Aesthetic / Imaging / Subject Consistency and reference-subject preservation (sample size as in §4.2 of the paper); the V2RGBA row and V2RGBA only are evaluated on the 500-case test set (250 stickers + 250 effects), all V2RGBA evaluations load the lightx2v 4-step LoRA, and alpha errors follow the paper protocol with ×10³ scaling.

Summary. These results describe one pipeline: one latent produces color and opacity, three mechanisms make that latent usable for asset production, and joint training lets the three tasks improve each other. In this sense the contribution is more than the sum of its components. The following sections support this framing with a dedicated alpha evaluation, an expanded V2RGBA evaluation, and VAE and DiT ablations.

W2. Insufficient experiments

W2.1 Quantitative alpha evaluation

BRAVER outputs foreground color and alpha in a single generation; two-stage methods such as Wan2.1-T2V + MatAnyone obtain alpha from an external matting model, so their error concentrates in the matting stage.

Evaluation protocol. Our existing metrics follow the white-background composite protocol of Wan-Alpha, which keeps the scores directly comparable with that baseline. We thank the reviewer for prompting a closer look at alpha quality, since alpha problems can be masked in this setting. We therefore add a five-dimensional alpha evaluation protocol for native RGBA video generation, checking the five problems alpha frequently exhibits: abnormal transparency, structural defects, background residue, edge defects, and temporal flicker, applied under the same protocol to T2RGBA, I2RGBA, and V2RGBA. Evaluation uses GPT-6.1-sol uniformly: the single-channel alpha is replicated into a three-channel grayscale image and placed side by side with the RGB foreground, and the pair is sent to the VLM together with the task prompt; the model scores each dimension from 1 to 5, where ≥3 is usable and 5 is defect-free.

Evaluation sets. T2RGBA and I2RGBA use 100 cases each (a baseline-comparison subset) and V2RGBA uses 500; within a task all methods use the same inputs and the same criterion. Tables report the fraction of samples scoring below 3.

Table 4. T2RGBA five-dimensional alpha defect rates (baseline-comparison set, 100 cases, %).

MethodAbnormal transparency ↓Structural defects ↓Background residue ↓Edge defects ↓Temporal flicker ↓
Wan-Alpha111121
LayerFlow76311001022
TransPixel357331
Wan2.1 + MatAnyone81812201
BRAVER (Ours)31110

Table note: values in Tables 4–6 are the fraction of samples scoring below 3 on the GPT-6.1-sol five-dimensional evaluation (scores 1–5, 5 = defect-free); all V2RGBA evaluations load the lightx2v 4-step LoRA. Table 4 uses a 100-case baseline-comparison subset, which differs from the expanded 300-case set in Table 8, so defect rates are not compared across the two tables.

Table 5. I2RGBA five-dimensional alpha defect rates (100 cases, %).

MethodAbnormal transparency ↓Structural defects ↓Background residue ↓Edge defects ↓Temporal flicker ↓
Wan2.1 + SAM2Matting24184414
Wan2.1 + MatAnyone1833841
Wan2.1 + MatAnyone 220121432
BRAVER (Ours)53312

Table 6. V2RGBA five-dimensional alpha defect rates (500 cases, %).

MethodAbnormal transparency ↓Structural defects ↓Background residue ↓Edge defects ↓Temporal flicker ↓
MatAnyone19.819.216.415.24.0
MatAnyone 217.220.020.411.07.2
SAM2Matting5.413.26.86.43.8
VideoMaMa2.07.89.02.82.6
BRAVER (Ours)1.05.26.21.42.2

On V2RGBA, BRAVER is lowest on all five metrics: 1.0% abnormal transparency, 5.2% structural defects, 6.2% background residue, 1.4% edge defects, and 2.2% temporal flicker. Edge defects drop from 15.2% for MatAnyone and 11.0% for MatAnyone 2 to 1.4%, background residue is below SAM2Matting (6.8%), and temporal flicker is below VideoMaMa (2.6%). Relative to the two-stage "generate then matte" pipeline, this gap is not marginal, and is most visible at foreground edges and in transparent regions. The pixel-level five-metric comparison with the four baselines on V2RGBA is in Table 7.

Table 7. V2RGBA five alpha metrics against baselines (500 cases).

MethodMAD ↓MSE ↓GRAD ↓CONN ↓DTSSD ↓
MatAnyone106.43690.03712.21544.4904.010
MatAnyone 297.34481.6359.72540.6053.534
SAM2Matting83.73865.37413.89634.3764.996
VideoMaMa106.72386.42411.66744.2093.843
BRAVER (Ours)27.67713.9807.33810.2183.806

The test set contains 250 stickers + 250 effects. BRAVER uses only the RGB video and text and does not need a first-frame alpha mask; VideoMaMa and SAM2Matting take the first-frame mask as input. BRAVER is lowest on MAD, MSE, GRAD, and CONN; its DTSSD is 3.806, below VideoMaMa (3.843) and SAM2Matting (4.996) and second only to MatAnyone 2 (3.534). On the 250-effect subset, the DTSSD of BRAVER is 4.175 and that of VideoMaMa is 4.098.

On I2RGBA, BRAVER obtains Aesthetic, Imaging, Subject Consistency, and reference-subject preservation of 0.6266, 0.5784, 0.9811, and 0.8560; the five-dimensional defect rates are in Table 5. The five-dimensional defect rates for T2RGBA are in Table 4.

W2.2 Sample size, margins, and seed variation

For sample size, T2RGBA is expanded from 100 to 300 cases, with all three methods evaluated on the same prompts (Table 8). BRAVER obtains Aesthetic, Imaging, Text Alignment, Naturalness, and Alpha quality of 0.5844, 0.5783, 4.26, 4.70, and 0.9913, highest in all five columns. In the 300-case alpha defect rates, BRAVER is lowest on all five: abnormal transparency 1.0%, structural defects 0.7%, background residue 1.7%, edge defects 2.3%, and temporal flicker 0.3%. V2RGBA is expanded to 500 cases (see W2.1).

For seed variation, we run three seeds on a new random set of 100 prompts and report mean ± standard deviation (Table 9). BRAVER obtains 0.5983 ± 0.0023, 0.5594 ± 0.0039, 4.17 ± 0.09, 4.67 ± 0.08, and 0.9915 ± 0.0051; the corresponding 95% t-interval half-widths are 0.0057, 0.0097, 0.22, 0.20, and 0.0127. Seed variation is small; because the intervals are wide relative to the margins, we base the main comparison on the 300-case Table 8 and use Table 9 to show seed-level variation.

Table 8. T2RGBA 300-case main evaluation and alpha defect rates.

MethodAesthetic ↑Imaging ↑Text Alignment ↑Naturalness ↑Alpha quality ↑Abnormal transparency ↓Structural defects ↓Background residue ↓Edge defects ↓Temporal flicker ↓
Wan-Alpha0.54200.51623.434.320.97874.71.05.06.32.7
Wan2.1-T2V + MatAnyone0.54970.55324.034.490.91468.712.011.320.33.3
BRAVER (Ours)0.58440.57834.264.700.99131.00.71.72.30.3

Table 9. T2RGBA 100 cases × 3 seeds, mean ± standard deviation.

MethodAesthetic ↑Imaging ↑Text Alignment ↑Naturalness ↑Alpha quality ↑
Wan-Alpha0.5430 ± 0.00190.4968 ± 0.00233.26 ± 0.074.37 ± 0.080.9781 ± 0.0052
Wan2.1-T2V + MatAnyone0.5411 ± 0.00140.5462 ± 0.00603.92 ± 0.084.34 ± 0.010.9167 ± 0.0090
BRAVER (Ours)0.5983 ± 0.00230.5594 ± 0.00394.17 ± 0.094.67 ± 0.080.9915 ± 0.0051

The alpha defect rates in Table 8 are in %, and the remaining columns are raw scores; Table 8 uses the expanded 300 prompts, which differ from the 100-case baseline-comparison set in Table 4. Table 9 reports mean ± standard deviation over three sampling seeds on a new random set of 100 prompts. Alpha quality is an alpha-quality score in [0,1], higher is better.

On the margin over Wan2.1-T2V + MatAnyone, BRAVER is better on all five main metrics over 300 cases, and the gap is larger in alpha defect rates (edge defects 2.3% vs. 20.3%, background residue 1.7% vs. 11.3%); Table 9 gives the variation across three seeds. The DiT ablation in Table 10 isolates teacher regularization and T2RGBA pre-adaptation on the quality metrics, and Table 11 gives the corresponding five alpha metrics; the joint model is best on both sets of metrics. Table 10. DiT training ablation.

VariantTeacherT2RGBA pre-adapted baseT2RGBA Aesthetic ↑T2RGBA Imaging ↑I2RGBA Ref. Preservation ↑I2RGBA Aesthetic ↑I2RGBA Imaging ↑Text Alignment ↑Naturalness ↑
w/o teachernoyes0.58640.53770.85250.61450.56474.204.66
w/o T2RGBA pre-adaptationnono0.58980.57020.85550.62150.56844.144.78
BRAVER (Ours)yesyes0.59690.59170.85600.62660.57844.254.87
Reviewer_k8T5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a unified framework for generating reusable RGBA video assets from text, images, or videos.

The method adapts a pretrained Wan ckpt to a compact RGBA latent via a tailored autoencoder trained with region wise alpha supervision and uses a two-stage adaptation scheme plus background-augmented reference conditioning to disentangle subject appearance from reference backgrounds.

The authors also curate a large RGBA video dataset, and report state-of-the-art aesthetic/naturalness scores in text-to-RGBA, improved VAE reconstruction, and stronger subject consistency in image-conditioned animation, with human preference favoring their alpha cleanliness over several baselines.

Critical strengths
  1. Introduces a practical, unified architecture that handles T2RGBA, I2RGBA, and V2RGBA in a single model with task modulation, leveraging a pretrained video prior with parameter efficient adaptation.
  2. A single-stage method that robustly supports text, image, and video inputs addresses a practical and important production use case not fully solved by two-stage pipelines.
  3. A large scale dataset(if released) that would be a valuable resource to advance transparent video generation and evaluation.
Critical weaknesses
  1. The novelty of this paper lies more in careful engineering than new theory. This paper explicitly adopts the RGBA autoencoder architecture of Wan-Alpha, while its main additions include opaque-core supervision and Fine-grained Alpha Detail Enhancement (FADE), visual-condition projection, task modulation and shared DoRA adaptation, and Decoupling Background from Asset References (DEBAR).
  2. The paper emphasizes reusable RGBA assets, but the author does not evaluate foreground RGB fidelity or compositing quality under different backgrounds. This lack of evaluation makes it difficult to fully assess whether the generated assets preserve their intended appearance when composited onto different scenes.
  3. The V2RGBA evaluation relies on 50 sticker and 50 effect clips composited over synthetic backgrounds, which may not fully reflect the challenges of extracting foreground assets from real-world videos.
Questions for authors
  1. To better assess whether the generated RGBA assets can preserve their appearance when reused across different scenes, could the authors provide quantitative evaluations of foreground RGB fidelity and compositing quality under different backgrounds?
  2. To better understand the generalization ability of BRAVER beyond synthetically composited videos, could the authors provide evaluation on real-world RGB videos with complex backgrounds and foreground structures?
3 comments

We thank the reviewer for recognizing the unified architecture, native RGBA generation, and practical value for reusable asset production. We also appreciate the questions about foreground color, cross-background reuse, and real-video generalization.

W1. Engineering versus technical contribution

Our contribution is a unified RGBA generation formulation in which the three tasks train together and improve each other. BRAVER handles text, RGB-image, and RGB-video references with one native RGBA latent model that produces foreground color and opacity together. Generation and matting therefore happen in one latent, so the asset is reusable as generated and the pipeline needs no separate output-matting model.

The three components are built for this latent. Each addresses one specific problem:

  • Reference background: it can leak into the generated asset, and DEBAR decouples the two.
  • Local alpha: errors concentrate in semi-transparent and interior regions, which FADE and opaque-core supervision target.
  • Three input types: text, image, and video conditions differ in form, and shared conditioning (visual-condition projection, task modulation, shared DoRA) maps all three into one backbone.

Joint training. We train the three single-task models and the joint model under the same budget (Tables 1–3). The joint model is higher on Text Alignment, Aesthetic, and Naturalness for T2RGBA (4.25/0.5969/4.87 vs. 4.04/0.5697/4.72), higher on all four I2RGBA metrics, and lower on all five V2RGBA alpha errors (MAD 36.859 → 27.677). The only exceptions are Smoothness and Flicker on T2RGBA (0.9928/0.9919 vs. 0.9932/0.9920), a gap of 0.0004 and 0.0001. On the sticker and effect subsets, the V2RGBA gap remains: alpha MAD is 5.667 and 49.687, versus 10.685 and 63.033 for V2RGBA only.

Table 1. T2RGBA joint vs. single-task training.

MethodText Alignment ↑Aesthetic ↑Naturalness ↑Smoothness ↑Flicker ↑
T2RGBA only4.040.56974.720.99320.9920
BRAVER (joint)4.250.59694.870.99280.9919

T2RGBA values use the paper's 100-prompt evaluation set.

Table 2. I2RGBA joint vs. single-task training.

MethodAesthetic ↑Imaging ↑Subject Consistency ↑Reference Preservation ↑
I2RGBA only0.61760.57390.96650.8525
BRAVER (joint)0.62660.57840.98110.8560

Table 3. V2RGBA joint vs. single-task training (500 cases).

MethodMAD ↓MSE ↓GRAD ↓CONN ↓DTSSD ↓
V2RGBA only36.85921.70311.36314.0075.043
BRAVER (joint)27.67713.9807.33810.2183.806

Summary. These results describe one pipeline: one latent produces color and opacity, three mechanisms make that latent usable for asset production, and joint training lets the three tasks improve each other. In this sense the contribution is more than the sum of its components.

W2 / Q1. Foreground RGB and multi-background compositing

W2.1 Foreground RGB fidelity (premultiplied foreground)

We check whether the same generated RGBA asset preserves its foreground appearance when reused in different scenes. On the 500-case V2RGBA test set, we evaluate premultiplied-foreground fidelity and report foreground decontamination (Table 4). All V2RGBA evaluations load the lightx2v 4-step LoRA.

Table 4. V2RGBA foreground fidelity and decontamination (500 cases).

MethodPremultiplied FG PSNR ↑Premultiplied FG SSIM ↑Premultiplied FG LPIPS ↓Premultiplied FG MAE ↓White-bg recompositing MAE ↓Colored-bg recompositing MAE ↓
MatAnyone30.830.86850.1523104.43118.87104.28
MatAnyone 232.750.87710.133189.12108.7092.30
SAM2Matting30.210.86820.165491.7598.3088.07
VideoMaMa29.530.85110.1733118.03122.31111.36
BRAVER (Ours)32.860.92170.069455.8263.8457.24

The region is the GT semi-transparent band 0.01≤α⋆≤0.990.01\le\alpha^\star\le0.99 dilated by 3 px; masks are built from GT only and shared by all methods. Alpha-only baselines are decontaminated with the known source background, while BRAVER uses its native foreground. MAE is RGB ×10310^3.

BRAVER obtains the highest premultiplied-foreground PSNR (32.86) and SSIM (0.9217), the lowest LPIPS (0.0694), and the lowest values on all three decontamination metrics.

W2.2 Compositing quality across backgrounds

We composite each output over black, white, colored, and natural backgrounds and report composite PSNR per background together with background leakage inside the GT-transparent region and halo inside the GT boundary band, in RGB MAE ×10^3 computed per background and then macro-averaged over the four backgrounds; masks are built from GT only and the backgrounds are identical for all methods. Results are in Table 5; foreground decontamination is in Table 4.

Table 5. V2RGBA multi-background compositing, background leakage, and halo (500 cases).

MethodPSNR black ↑PSNR white ↑PSNR colored ↑PSNR natural ↑Background leakage ↓Halo ↓
MatAnyone30.1629.2329.3040.5617.8291.71
MatAnyone 231.7330.8830.9743.6118.3180.45
SAM2Matting29.4029.4628.8534.878.3085.35
VideoMaMa28.9828.5528.3438.9416.88100.74
BRAVER (Ours)31.8331.7331.0142.032.8654.11

BRAVER is highest on the black, white, and colored backgrounds and lowest on background leakage (2.86) and halo (54.11); MatAnyone 2 is highest on the natural background. The compositing columns use each method's raw output, so alpha-only baselines use F^=I\hat F=I and their natural-background PSNR is high.

W2.3 Five-dimensional alpha defect rates

Beyond foreground fidelity and compositing error, we directly check alpha itself with a five-dimensional defect-rate evaluation that measures abnormal transparency, structural defects, background residue, edge defects, and temporal flicker. Evaluation uses GPT-6.1-sol uniformly: the single-channel alpha is replicated into a three-channel grayscale image and placed side by side with the RGB foreground, and the pair is sent to the VLM together with the task prompt; the model scores each dimension from 1 to 5, where 5 is defect-free and a score below 3 counts as a defect. V2RGBA uses 500 cases, all methods use the same inputs and the same criterion, and the table reports the fraction of samples scoring below 3.

Table 6. V2RGBA five-dimensional alpha defect rates (500 cases, %).

MethodAbnormal transparency ↓Structural defects ↓Background residue ↓Edge defects ↓Temporal flicker ↓
MatAnyone19.819.216.415.24.0
MatAnyone 217.220.020.411.07.2
SAM2Matting5.413.26.86.43.8
VideoMaMa2.07.89.02.82.6
BRAVER (Ours)1.05.26.21.42.2

Table note: values are the fraction of samples scoring below 3 on the GPT-6.1-sol five-dimensional evaluation (scores 1–5, 5 = defect-free); all V2RGBA evaluations load the lightx2v 4-step LoRA.

On V2RGBA, BRAVER is lowest on all five metrics: 1.0% abnormal transparency, 5.2% structural defects, 6.2% background residue, 1.4% edge defects, and 2.2% temporal flicker.

W3 / Q2. Generalization to real videos

W3.1 Representativeness of the synthetic-background evaluation

The V2RGBA evaluation is expanded to 500 clips (250 stickers + 250 effects), but synthetic backgrounds still do not cover the complex backgrounds and foreground structures of real captures. The existing V2RGBA results are used only for comparison under that protocol; the qualitative real-video comparison is in W3.2.

W3.2 Qualitative comparison on real videos

Real-world captures have no per-pixel RGBA ground truth, so reference-based metrics such as alpha MAD and dtSSD are not defined; we therefore provide a qualitative comparison rather than a numerical evaluation. The anonymous repository contains eight showcases comparing BRAVER with the two-stage baseline that segments the first frame with SAM3 and then mattes with SAM2Matting on real videos taken from film and VFX footage, covering semi-transparent materials such as bubbles and woven fabric, together with extreme motion and deformation. All methods receive the same input video, and the clips can be inspected frame by frame for foreground integrity, semi-transparent regions, and background residue: https://anonymous.4open.science/r/braver-rebuttal-demo/index.html

We have added more qualitative results in the main paper (marked blue), and failure cases will be added to the supplementary material.

Thank you again for your questions about how well RGBA assets preserve their foreground appearance when reused across scenes. We have added further evaluations and visual comparisons to the revised paper. We also invite you to view the results in motion on our anonymous video page.

If any concerns remain, we’d be glad to discuss them before the deadline. If the additional evidence has resolved your concerns, we would appreciate it if you would consider raising your score.

Reviewer_J75wOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets the creation of reusable animated assets: RGBA videos in which a foreground is cleanly separated from its original background, with temporally coherent alpha through fine boundaries and translucent regions. The authors propose BRAVER, a single framework that unifies text-conditioned generation, image-guided animation, and aligned video foreground extraction, requiring no user-provided input matte and no separate output-matting stage. The key contributions are a region-wise alpha-supervised RGBA autoencoder, a shared condition projection with temporal masks and task-modulated parameter-efficient adaptation of a frozen video diffusion prior, and DEBAR, a training strategy that composites assets over diverse backgrounds while keeping the original RGBA asset as the target. Experiments report consistent gains in reconstruction fidelity, generation quality, and human preference over the evaluated baselines.

Critical strengths
  1. Important and practical problem. BRAVER addresses reusable RGBA video creation from text, RGB images, and RGB videos without requiring user-provided mattes. A single model supports T2RGBA, I2RGBA, and V2RGBA, avoiding separate generation and matting pipelines.
  2. Well-motivated alpha supervision. FADE explicitly targets boundaries and translucent regions, leading to consistent improvements in VAE reconstruction.
  3. Effective background augmentation. DEBAR directly addresses foreground–background ambiguity. Inverting the known compositing process yields exact supervision for foreground–background separation.
Critical weaknesses
  1. The framework primarily combines existing RGBA VAEs, Wan2.1, DoRA, and compositing augmentation, with innovation largely at the system level.
  2. Metrics are misaligned with reusability. All generation scores use white-background composites, which hide dirty alpha. VBench and the VLM judge evaluate the appearance of these composites rather than whether the resulting assets remain clean when reused in new scenes. The paper should evaluate compositing over diverse backgrounds and explicitly measure background leakage, halo artifacts, and foreground color decontamination.
  3. The paper’s main claim, reliable local opacity, is never actually measured for T2RGBA or I2RGBA. There’s no alpha MAD/IoU, no boundary band accuracy, no gradient domain alpha error, and no alpha specific temporal consistency metric. The only alpha error metric (Table 5, V2V ablation) is a full-frame mean, insensitive to exactly the local failures (interior leakage, residue in empty regions) that Figure 5 highlights.
  4. The ablations are insufficient to attribute the improvements to the proposed components. The FADE components are only ablated as a whole: the three region terms in Eq. 3 are never separated, and it is unspecified whether “w/o FADE” also removes the opaque-core term. There is also no ablation of two-stage training or the first-stage teacher regularization.
Questions for authors

Please address the concerns raised in the Weaknesses section. I am willing to reconsider and increase my score.

9 comments

We thank the reviewer for the careful and constructive review, and for recognizing the importance of the problem BRAVER addresses, the practical value of unifying three input types, the motivation for FADE, and the exact supervision enabled by DEBAR. We respond point by point.

W1. System-level contribution

Our contribution is a unified RGBA generation formulation in which the three tasks train together and improve each other. Text, RGB-image, and RGB-video conditions are handled by one native RGBA latent model that produces foreground color and alpha together. Generation and matting therefore happen in one latent, so the asset is reusable as generated and the pipeline needs no separate output-matting model.

The three components are built for this latent. Each addresses one specific problem:

  • Reference background: it can leak into the generated asset, and DEBAR decouples the two.
  • Local alpha: errors concentrate in semi-transparent and interior regions, which FADE and opaque-core supervision target.
  • Three input types: text, image, and video conditions differ in form, and shared conditioning (visual-condition projection, task modulation, shared DoRA) maps all three into one backbone.

Joint training. We train the three single-task models and the joint model under the same budget (Tables 1–3). The joint model is higher on Text Alignment, Aesthetic, and Naturalness for T2RGBA (4.25/0.5969/4.87 vs. 4.04/0.5697/4.72), higher on all four I2RGBA metrics, and lower on all five V2RGBA alpha errors (MAD 36.859 → 27.677). The only exceptions are Smoothness and Flicker on T2RGBA (0.9928/0.9919 vs. 0.9932/0.9920), a gap of 0.0004 and 0.0001. On the sticker and effect subsets, the V2RGBA gap remains: alpha MAD is 5.667 and 49.687, versus 10.685 and 63.033 for V2RGBA only.

Table 1. T2RGBA joint vs. single-task training.

MethodText Alignment ↑Aesthetic ↑Naturalness ↑Smoothness ↑Flicker ↑
T2RGBA only4.040.56974.720.99320.9920
BRAVER (joint)4.250.59694.870.99280.9919

T2RGBA values use the paper's 100-prompt evaluation set.

Table 2. I2RGBA joint vs. single-task training.

MethodAesthetic ↑Imaging ↑Subject Consistency ↑Reference Preservation ↑
I2RGBA only0.61760.57390.96650.8525
BRAVER (joint)0.62660.57840.98110.8560

Table 3. V2RGBA joint vs. single-task training (500 cases).

MethodMAD ↓MSE ↓GRAD ↓CONN ↓DTSSD ↓
V2RGBA only36.85921.70311.36314.0075.043
BRAVER (joint)27.67713.9807.33810.2183.806

Summary. These results describe one pipeline: one latent produces color and alpha, three mechanisms make that latent usable for asset production, and joint training lets the three tasks improve each other. In this sense the contribution is more than the sum of its components.

W2. Metrics misaligned with reusability

W2.1 White-background composites hide dirty alpha

We adopt white-background composites because they follow the evaluation protocol of Wan-Alpha, which keeps our generation scores directly comparable with that baseline. We thank the reviewer for pointing out that this protocol can hide dirty alpha. We therefore add a five-dimensional alpha evaluation: abnormal transparency, structural defects, background residue, edge defects, and temporal flicker.

Protocol. Evaluation uses GPT-6.1-sol uniformly. The single-channel alpha is replicated into a three-channel grayscale image and placed side by side with the RGB foreground; the pair is sent to the VLM with the task prompt. The model scores each dimension from 1 to 5, where 5 is defect-free and below 3 counts as a defect.

Results (Tables 4–6). Tables 4, 5, and 6 report these five scores for T2RGBA, I2RGBA, and V2RGBA, respectively, comparing BRAVER with the baselines on the same cases; for each dimension we report the percentage of samples scoring below 3. BRAVER's defect rates are 3% / 1% / 1% / 1% / 0% on 100 T2RGBA cases, 5% / 3% / 3% / 1% / 2% on 100 I2RGBA cases, and 1.0% / 5.2% / 6.2% / 1.4% / 2.2% on 500 V2RGBA cases, in the order above. These rates remain low across tasks and dimensions. Against the baselines, BRAVER is the lowest or tied-lowest on all five dimensions for T2RGBA and on four of five for I2RGBA; the only exception is temporal flicker on I2RGBA (2% vs. 1% for Wan2.1 + MatAnyone), which reflects accurate alpha and consistent results.

Table 4. T2RGBA five-dimensional alpha defect rates (100 cases, %).

MethodAbnormal transparency ↓Structural defects ↓Background residue ↓Edge defects ↓Temporal flicker ↓
Wan-Alpha111121
LayerFlow76311001022
TransPixel357331
Wan2.1 + MatAnyone81812201
BRAVER (Ours)31110

Table 5. I2RGBA five-dimensional alpha defect rates (100 cases, %).

MethodAbnormal transparency ↓Structural defects ↓Background residue ↓Edge defects ↓Temporal flicker ↓
Wan2.1 + SAM2Matting24184414
Wan2.1 + MatAnyone1833841
Wan2.1 + MatAnyone 220121432
BRAVER (Ours)53312

Table 6. V2RGBA five-dimensional alpha defect rates (500 cases, %).

MethodAbnormal transparency ↓Structural defects ↓Background residue ↓Edge defects ↓Temporal flicker ↓
MatAnyone19.819.216.415.24.0
MatAnyone 217.220.020.411.07.2
SAM2Matting5.413.26.86.43.8
VideoMaMa2.07.89.02.82.6
BRAVER (Ours)1.05.26.21.42.2

W2.2 Multi-background compositing, background leakage, and foreground decontamination

We report cross-background compositing and foreground decontamination; all regions are defined by the GT alpha and shared by all methods.

Compositing. Background leakage is the RGB MAE between the composite and the background inside the GT-transparent region (α⋆<0.01\alpha^\star<0.01). Halo is the RGB MAE against the GT composite inside the GT boundary band (0.01≤α⋆≤0.990.01\le\alpha^\star\le0.99, dilated by 3 px). Composite PSNR is computed on the full frame for black, white, colored, and natural backgrounds.

Foreground decontamination. Three errors are computed inside the GT semi-transparent band (0.01≤α⋆≤0.990.01\le\alpha^\star\le0.99, dilated by 3 px): premultiplied foreground MAE, white-background recompositing MAE, and colored-background recompositing MAE. BRAVER uses its native foreground color; alpha-only baselines are decontaminated with the known source background. Errors are in RGB MAE ×10³, averaged per frame, then across frames, and finally over the four backgrounds.

Compositing (Table 7). BRAVER obtains the highest composite PSNR on black, white, and colored backgrounds, and the lowest background leakage (2.86) and halo (54.11).

Decontamination (Table 7). Its three decontamination errors are 55.82, 63.84, and 57.24, also the lowest; switching from black to white and colored backgrounds increases the error by only 14.4% and 2.5%.

Alpha side (Table 6). BRAVER is lowest on all five defect rates: 1.0% abnormal transparency, 5.2% structural defects, 6.2% background residue, 1.4% edge defects, and 2.2% temporal flicker.

Table 7. Multi-background compositing, background leakage, halo, and foreground decontamination (V2RGBA, 500 cases).

MethodPSNR black ↑PSNR white ↑PSNR colored ↑PSNR natural ↑Background leakage ↓Halo ↓Premultiplied FG MAE ↓White-bg recompositing MAE ↓Colored-bg recompositing MAE ↓
MatAnyone30.1629.2329.3040.5617.8291.71104.43118.87104.28
MatAnyone 231.7330.8830.9743.6118.3180.4589.12108.7092.30
SAM2Matting29.4029.4628.8534.878.3085.3591.7598.3088.07
VideoMaMa28.9828.5528.3438.9416.88100.74118.03122.31111.36
BRAVER (Ours)31.8331.7331.0142.032.8654.1155.8263.8457.24

The test set contains 500 cases (250 stickers + 250 effects). Compositing columns use each method's raw output (alpha-only baselines take F^=I\hat F=I, so their natural-background PSNR is high).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.