Research map / Video / Video generation
VC-VAE: Leveraging Video Codecs for Training-Efficient and High-Fidelity Video VAE
Xinxu Ge, Shang Chai, Litong Gong, Zitong Yu, Xin Liu, Tiezheng Ge
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 TDC to causal conv (ablation) | Causal Conv Video VAE · Kinetics-600 val | PSNR: 33.29 -> 33.64 | Improved |
| add TDC to group causal conv (ablation) | GCConv Video VAE · Kinetics-600 val | PSNR: 33.49 -> 33.79 | Improved |
| add TDC at higher temporal compression (ablation) | GCConv Video VAE 8x8x8 · Kinetics-600 val | PSNR: 31.11 -> 32.01 | Improved |
| use in-sequence instead of separate first-frame processing (ablation) | VC-VAE · Kinetics-600 val | PSNR: 32.15 -> 33.34 | Improved |
| initialize from Flux-VAE instead of SD3.5-VAE (no TDC) (ablation) | Video VAE w/o TDC · Kinetics-600 val | PSNR: 33.13 -> 33.49 | Improved |
| add TDC to SD3.5-VAE-initialized model (ablation) | Video VAE SD3.5-VAE init · Kinetics-600 val | PSNR: 33.13 -> 33.74 | Improved |
| add TDC to Flux-VAE-initialized model (ablation) | Video VAE Flux-VAE init · Kinetics-600 val | PSNR: 33.49 -> 33.79 | Improved |
| use VC-VAE over WFVAE at 8x8x8 | WFVAE · WebVid-10M test (8x8x8) | FVD: 44.13 -> 36.63 | Improved |
| use VC-VAE over Hunyuan-VAE at 4x8x8 (reconstruction PSNR) | Hunyuan-VAE · WebVid-10M test (4x8x8) | PSNR: 35.84 -> 36.04 | Improved |
| use VC-VAE over Hunyuan-VAE at 4x8x8 (FVD) | Hunyuan-VAE · WebVid-10M test (4x8x8) | FVD: 36.23 -> 36.80 | Hurt |
| use VC-VAE latents for Latte-XL generation instead of IVVAE | IVVAE · SkyTimelapse | gFVD: 118.6 -> 112.8 | Improved |
| use VC-VAE over Hunyuan-VAE on text reconstruction | Hunyuan-VAE · TokBench-Video 256x | T-ACC mean: 61.15 -> 66.24 | Improved |
Closest papers by meaning
- Improved Video VAE for Latent Video Diffusion Model (CVPR 2025 · similarity 0.92)
- CV-VAE: A Compatible Video VAE for Latent Generative Video Models (NeurIPS 2024 · similarity 0.91)
- WF-VAE: Enhancing Video VAE by Wavelet-Driven Energy Flow for Latent Video Diffusion Model (CVPR 2025 · similarity 0.91)
- VidTwin: Video VAE with Decoupled Structure and Dynamics (CVPR 2025 · similarity 0.88)
- VRVVC: Variable-Rate NeRF-Based Volumetric Video Compression (AAAI 2025 · similarity 0.87)
- VideoMAC: Video Masked Autoencoders Meet ConvNets (CVPR 2024 · 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.
