- 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
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, %).
| Method | Abnormal transparency ↓ | Structural defects ↓ | Background residue ↓ | Edge defects ↓ | Temporal flicker ↓ |
|---|---|---|---|---|---|
| Wan-Alpha | 11 | 1 | 1 | 2 | 1 |
| LayerFlow | 76 | 31 | 100 | 10 | 22 |
| TransPixel | 3 | 5 | 7 | 33 | 1 |
| Wan2.1 + MatAnyone | 8 | 18 | 12 | 20 | 1 |
| BRAVER (Ours) | 3 | 1 | 1 | 1 | 0 |
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, %).
| Method | Abnormal transparency ↓ | Structural defects ↓ | Background residue ↓ | Edge defects ↓ | Temporal flicker ↓ |
|---|---|---|---|---|---|
| Wan2.1 + SAM2Matting | 24 | 18 | 4 | 4 | 14 |
| Wan2.1 + MatAnyone | 18 | 33 | 8 | 4 | 1 |
| Wan2.1 + MatAnyone 2 | 20 | 12 | 14 | 3 | 2 |
| BRAVER (Ours) | 5 | 3 | 3 | 1 | 2 |
Table 6. V2RGBA five-dimensional alpha defect rates (500 cases, %).
| Method | Abnormal transparency ↓ | Structural defects ↓ | Background residue ↓ | Edge defects ↓ | Temporal flicker ↓ |
|---|---|---|---|---|---|
| MatAnyone | 19.8 | 19.2 | 16.4 | 15.2 | 4.0 |
| MatAnyone 2 | 17.2 | 20.0 | 20.4 | 11.0 | 7.2 |
| SAM2Matting | 5.4 | 13.2 | 6.8 | 6.4 | 3.8 |
| VideoMaMa | 2.0 | 7.8 | 9.0 | 2.8 | 2.6 |
| BRAVER (Ours) | 1.0 | 5.2 | 6.2 | 1.4 | 2.2 |
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).
| Method | MAD ↓ | MSE ↓ | GRAD ↓ | CONN ↓ | DTSSD ↓ |
|---|---|---|---|---|---|
| MatAnyone | 106.436 | 90.037 | 12.215 | 44.490 | 4.010 |
| MatAnyone 2 | 97.344 | 81.635 | 9.725 | 40.605 | 3.534 |
| SAM2Matting | 83.738 | 65.374 | 13.896 | 34.376 | 4.996 |
| VideoMaMa | 106.723 | 86.424 | 11.667 | 44.209 | 3.843 |
| BRAVER (Ours) | 27.677 | 13.980 | 7.338 | 10.218 | 3.806 |
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.
| Method | Aesthetic ↑ | Imaging ↑ | Text Alignment ↑ | Naturalness ↑ | Alpha quality ↑ | Abnormal transparency ↓ | Structural defects ↓ | Background residue ↓ | Edge defects ↓ | Temporal flicker ↓ |
|---|---|---|---|---|---|---|---|---|---|---|
| Wan-Alpha | 0.5420 | 0.5162 | 3.43 | 4.32 | 0.9787 | 4.7 | 1.0 | 5.0 | 6.3 | 2.7 |
| Wan2.1-T2V + MatAnyone | 0.5497 | 0.5532 | 4.03 | 4.49 | 0.9146 | 8.7 | 12.0 | 11.3 | 20.3 | 3.3 |
| BRAVER (Ours) | 0.5844 | 0.5783 | 4.26 | 4.70 | 0.9913 | 1.0 | 0.7 | 1.7 | 2.3 | 0.3 |
Table 9. T2RGBA 100 cases × 3 seeds, mean ± standard deviation.
| Method | Aesthetic ↑ | Imaging ↑ | Text Alignment ↑ | Naturalness ↑ | Alpha quality ↑ |
|---|---|---|---|---|---|
| Wan-Alpha | 0.5430 ± 0.0019 | 0.4968 ± 0.0023 | 3.26 ± 0.07 | 4.37 ± 0.08 | 0.9781 ± 0.0052 |
| Wan2.1-T2V + MatAnyone | 0.5411 ± 0.0014 | 0.5462 ± 0.0060 | 3.92 ± 0.08 | 4.34 ± 0.01 | 0.9167 ± 0.0090 |
| BRAVER (Ours) | 0.5983 ± 0.0023 | 0.5594 ± 0.0039 | 4.17 ± 0.09 | 4.67 ± 0.08 | 0.9915 ± 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.
| Variant | Teacher | T2RGBA pre-adapted base | T2RGBA Aesthetic ↑ | T2RGBA Imaging ↑ | I2RGBA Ref. Preservation ↑ | I2RGBA Aesthetic ↑ | I2RGBA Imaging ↑ | Text Alignment ↑ | Naturalness ↑ |
|---|---|---|---|---|---|---|---|---|---|
| w/o teacher | no | yes | 0.5864 | 0.5377 | 0.8525 | 0.6145 | 0.5647 | 4.20 | 4.66 |
| w/o T2RGBA pre-adaptation | no | no | 0.5898 | 0.5702 | 0.8555 | 0.6215 | 0.5684 | 4.14 | 4.78 |
| BRAVER (Ours) | yes | yes | 0.5969 | 0.5917 | 0.8560 | 0.6266 | 0.5784 | 4.25 | 4.87 |
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:
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.
Table 2. I2RGBA joint vs. single-task training.
Table 3. V2RGBA joint vs. single-task training (500 cases).
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.