Tsinghua and Shengshu's APEX replaces fake-score nets with condition shifting for faster streaming video generation
Tsinghua · hf · 2026-11-11
Researchers from Tsinghua, Shengshu Technology, and Westlake University propose APEX, extending Distribution Matching Distillation (DMD) to streaming video generation.
- Uses condition shifting to estimate the generated distribution's velocity field: the original condition selects generation while a shifted condition performs fake estimation with the same parameters, eliminating the separate fake-score network
- Reduces training memory and computation vs DMD and no longer needs to track the evolving generator
- Retains S2-Avatar's frozen reference teacher, Self-Replay Forcing, and perceptual regularization
- Evaluated on 180-second videos with a 70-case human study and 320-case StreamAV-Bench comparison: APEX beats S2-Avatar on instruction following (3.856 vs 3.831 on the Progressive track) and visual aesthetics (0.6419 vs 0.6326)
More from Multimodal
- AI-generated holographic cyberpunk shop scene impresses with detail — Zard-Oz · 2026-11-11
- Seedream 5.5 edits hold character identity across 30 frames in user test — aziz4ai · 2026-11-11
- Alibaba's PhysOPSD uses executable code as privileged context to make video generation physically plausible — AlibabaTongyiLab · 2026-11-11
- AI short film 'A Single Fry' shows cinematic seagull scene — gps2019 · 2026-11-11
- Designer: fable 5.2 plus paper is an unstoppable combo for extending existing visual languages — evilrabbit_ · 2026-11-11
- ComfyUI v0.45.0 Ships With Frontend Package Bump to 1.58.5 — Fcking_Chuck · 2026-11-11