Research map / Video / Video generation
Video Generation with Stable Transparency via Shiftable RGB-A Distribution Learner
Haotian Dong, Wenjing Wang, Chen Li, Jing Lyu, Di Lin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add rendering operations (input pre-rendering and reconstruction losses) to VAE training (ablation) | RGB-A VAE · RGB-A video reconstruction, alpha channel | PSNR: 39.98 -> 41.22 | Improved |
| add transparency-aware bidirectional diffusion loss (TABD) on top of rendering (ablation) | RGB-A VAE · RGB-A video reconstruction, alpha channel | PSNR: 41.22 -> 42.22 | Improved |
| add TABD on top of rendering (ablation) | RGB-A VAE · RGB-A video reconstruction, RGB channels | PSNR: 40.88 -> 41.47 | Improved |
| add TABD to VAE training (ablation) | RGB-A video generation · qualitative generation | holes in opaque regions | Improved |
| add transparency-guided mean sampler (MS) to DiT training (ablation) | RGB-A video generation · qualitative generation | unwanted background artifacts | Improved |
| set mean-shift strength mu too small (ablation) | mean sampler · qualitative | transparency control | No clear effect |
| set mean-shift strength mu to 0.5 (10x default 0.05) (ablation) | mean sampler · qualitative | color bias | Hurt |
| train DiT with DoRA instead of LoRA (ablation) | Wan2.1-T2V-14B · RGB-A video generation | semantic alignment and quality | Improved |
| use shiftable RGB-A distribution learner over TransPixeler | RGB-A video generation · inference speed | generation time: 32 minutes for 49 frames -> 128 seconds for 81 frames | Improved |
| generate then matte (Wan + MatAnyone) instead of joint RGB-A generation | video generation + matting · transparency quality | qualitative | Hurt |
Closest papers by meaning
- TransPixeler: Advancing Text-to-Video Generation with Transparency (CVPR 2025 · similarity 0.88)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.84)
- Spectral Progressive Diffusion for Efficient Image and Video Generation (NeurIPS 2026 · similarity 0.84)
- Improving the Diffusability of Autoencoders (ICML 2025 · similarity 0.84)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.83)
- Decouple Content and Motion for Conditional Image-to-Video Generation (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
