Research map / Video / Video generation
Delving into Latent Spectral Biasing of Video VAEs for Superior Diffusability
Shizhan Liu, Xinran Deng, Zhuoyi Yang, Jiayan Teng, Xiaotao Gu, Jie Tang
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 LMR and LCR (SSVAE) to 48-ch baseline VAE | MMDiT 1.3B · VBench 17x512x512 | UR: 40.9 -> 45.9 | Improved |
| Add LCR to 48-ch baseline (ablation) | VAE · MovieValid | UR: 38.1 -> 39.0 | Improved |
| Add LMR on top of LCR (ablation) | VAE · MovieValid | UR: 39.0 -> 41.4 | Improved |
| Add decoder finetune after LMR (ablation) | VAE · MovieValid | UR: 41.4 -> 41.3 | No clear effect |
| Use LCR instead of SER or VA-VAE for low-frequency bias (ablation) | 64-ch VAE · UR | UR: 34.4 -> 35.5 | Improved |
| Add LCR to 64-ch VAE (ablation) | 64-ch VAE · FVD | FVD: 553 -> 606 | Hurt |
| Over-bias LCR (omega=0.05, alpha=0.9) (ablation) | LCR · UR | UR | Hurt |
| SSVAE vs Wan 2.2 VAE at 17x256x256 | MMDiT 1.3B · VBench | FVD: 873 -> 983 | Mixed |
| SSVAE on Wan 1.3B backbone | Wan 1.3B · VBench 17x512x512 | UR: 37.7 -> 45.5 | Improved |
| Add LMR to LCR VAE (ablation) | VAE · MovieValid reconstruction | PSNR: 38.29 -> 37.84 | Hurt |
| Use LMR mask probabilities {0.7,0.1,0.1,0.1} vs no LMR (ablation) | VAE · MovieValid | UR: 39.2 -> 42.9 | Improved |
Closest papers by meaning
- CV-VAE: A Compatible Video VAE for Latent Generative Video Models (NeurIPS 2024 · similarity 0.90)
- Improving the Diffusability of Autoencoders (ICML 2025 · similarity 0.88)
- WF-VAE: Enhancing Video VAE by Wavelet-Driven Energy Flow for Latent Video Diffusion Model (CVPR 2025 · similarity 0.88)
- Improved Video VAE for Latent Video Diffusion Model (CVPR 2025 · similarity 0.88)
- Efficient Video Diffusion Models via Content-Frame Motion-Latent Decomposition (ICLR 2024 · similarity 0.88)
- Spectral Progressive Diffusion for Efficient Image and Video Generation (NeurIPS 2026 · similarity 0.87)
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.
