- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies how to extend a pretrained RGB video generator to produce transparent RGBA videos while retaining its RGB generation capability. It adapts the causal video VAE at its input and output boundaries, preserves the existing 16-channel latent interface, and uses lightweight DoRA tuning for downstream generation. The paper evaluates reconstruction, text-to-video generation, image-to-video generation, and RGB compatibility. On its RGBA reconstruction benchmark, UniRGBA reports a 4.16 dB PSNR improvement over a retrained Wan-Alpha baseline.
Critical strengths
-
The method extends the causal VAE at its input and output boundaries lets the method keep the DiT’s 16-channel latent interface. This gives the design a coherent reason for avoiding a separate alpha-token stream.
-
Selectively updating roughly 5% of the VAE parameters limits how much of the pretrained representation must be relearned.
Critical weaknesses
-
RGBA representation learning, compositing losses, and latent constraints have precedents in AlphaVAE[1]; transparent-video generation already has TransPixeler[2] and Wan-Alpha[3]. The paper's contribution is narrowly defined as adapting a causal video VAE while retaining the pretrained DiT interface.
-
Retaining the 16-channel latent width plausibly avoids alpha-token overhead, but the experiments do not establish the resulting differences in sequence length, memory, speed, or training cost under matched conditions. This leaves an important part of the method’s motivation supported by design reasoning rather than measurement.
[1] AlphaVAE: Unified End-to-End RGBA Image Reconstruction and Generation with Alpha-Aware Representation Learning [2] TransPixeler: Advancing Text-to-Video Generation with Transparency [3] Video Generation with Stable Transparency via Shiftable RGB-A Distribution Learner
Questions for authors
- At matched resolution and video length, how do token count, peak memory, inference time, and training cost compare with TransPixeler and Wan-Alpha?
- How much of the reconstruction gain comes from the architecture rather than additional training? Please compare the proposed channel expansion and selective unfreezing with alternative adaptations under the same data and update budget.
- Does Stage 2 help because of the GAN loss or because training continues?
- How representative is the internal RGBA benchmark? Please give a breakdown of transparent objects, diffuse effects, difficult boundaries, clip lengths, and source types, and specify which evaluation assets can be released.