Research map / Video / Video generation
Flash-VAED: Plug-and-Play VAE Decoders for Efficient Video Generation
Lunjie Zhu, Yushi Huang, Xingtong Ge, Yufei Xue, Zhening Liu, Yumeng Zhang, Zehong Lin, Jun Zhang
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 |
|---|---|---|---|
| use independence-aware channel pruning instead of random or L1-norm pruning (ablation) | Flash-VAED · PSNR over training steps | PSNR | Improved |
| add pre-pruning retained channel enhancement (expressivity loss) | Flash-VAED channel pruning · feature reconstruction | R2: 0.87-0.93 -> exceeding 0.98 | Improved |
| replace CausalConv3D with 2D convolutions in higher-resolution layers (ablation) | Wan 2.1 VAE decoder · UCF-101 reconstruction | latency and PSNR: latency 396 -> 307 ms; PSNR 32.42 -> 32.99 (up0 vs up3) | Improved |
| reduce prune ratio from 1/2 to 1/4 (ablation) | Flash-VAED-LTX · UCF-101 reconstruction | FPS / PSNR: 625.00 -> 1167.99 FPS; PSNR 32.5421 -> 32.2400 | Mixed |
| reduce prune ratio from 1/4 to 1/8 (ablation) | Flash-VAED-LTX · UCF-101 reconstruction | FPS / PSNR: 1167.99 -> 1907.23 FPS; PSNR 32.2400 -> 31.0786 | Mixed |
| initialize projection layer with derived W instead of random (ablation) | Flash-VAED-LTX phase 3 · UCF-101 reconstruction | PSNR: 30.3061 -> 32.2400 | Improved |
| add three-phase feature distillation (vs none) (ablation) | Flash-VAED-LTX · UCF-101 reconstruction | PSNR: 30.7957 -> 32.3768 | Improved |
| distill only deep layers instead of full three-phase (ablation) | Flash-VAED-LTX · UCF-101 reconstruction | PSNR: 30.7957 -> 30.8065 vs no distillation | No clear effect |
| distill only shallow layers instead of full three-phase (ablation) | Flash-VAED-LTX · UCF-101 reconstruction | PSNR: 32.3768 -> 31.2386 | Hurt |
| swap original VAE decoder for Flash-VAED in Self Forcing-Wan 1.3B | Self Forcing-Wan 1.3B · end-to-end generation, VBench-2.0 | latency: 10.14s -> 7.41s (27%) | Improved |
| swap original VAE decoder for Flash-VAED in FastVideo-Wan 1.3B | FastVideo-Wan 1.3B · end-to-end generation, VBench-2.0 | latency: 9.66s -> 6.15s (36%) | Improved |
| use LightVAE decoder instead of Flash-VAED in generation pipelines | FastVideo-Wan 1.3B · VBench-2.0 | VBench-2.0 quality | Hurt |
Closest papers by meaning
- CV-VAE: A Compatible Video VAE for Latent Generative Video Models (NeurIPS 2024 · similarity 0.87)
- Improved Video VAE for Latent Video Diffusion Model (CVPR 2025 · similarity 0.87)
- WF-VAE: Enhancing Video VAE by Wavelet-Driven Energy Flow for Latent Video Diffusion Model (CVPR 2025 · similarity 0.87)
- MobileWan: Closing the Quality Gap for Mobile Video Diffusion (NeurIPS 2026 · similarity 0.86)
- Epsilon-VAE: Denoising as Visual Decoding (ICML 2025 · similarity 0.86)
- ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation (ICLR 2025 · similarity 0.85)
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.
