LightVAE speeds up video VAE decoding up to 2.24x on H100 with minimal quality loss
Shiqiao Gu · hf · 2026-10-14
SenseTime and collaborators introduce LightVAE, a compression framework for video VAEs that tackles the heavy encoding/decoding cost in latent diffusion video generation.
- Exactly reparameterizes causal Conv3D layers into an aggregate spatial response plus a difference-driven temporal correction, applying temporal low-rank compression (TLR) only to the correction branch based on distinct compression sensitivities
- Adds width-preserving block pruning and probe-guided feature alignment (feedback from a frozen teacher suffix) to recover quality
- On the Wan VAE family: up to 2.24x decoding throughput of SOTA efficient decoders on H100 and 1.11x on RTX 5090, with higher DAVIS PSNR and lower LPIPS
- On MiniMax-H3: up to 2.97x throughput while retaining 99.51% of the original VBench-2.0 score
Code is open-sourced (ModelTC/LightVAE).
More from Infra
- Photonic Training of a 1B GPT-2 Without Backprop Lands in PNAS — IgorCarron · 2026-10-14
- Xinjiang's 167 GW green power, 16ms to Chongqing: the grid edge in AI race — 96Stats · 2026-10-14
- Migrating a key API from Node to Bun cut p99 latency 80% and memory 63% — DanielLockyer · 2026-10-14
- Agent traffic is a strange inference workload: long prefix caching beats raw tokens per second — rohanpaul_ai · 2026-10-14
- Orbital data centers have no off-switch: Reddit debates AI safety paradox — Zappy_Oh · 2026-10-14
- Rednote's Dots unveils MSAR: 1.58x iso-loss compute speedup over residual baseline — teortaxesTex · 2026-10-14