DuoMatching distills image priors into few-step video generation, lifting VBench from 80.67 to 83.53
量子位 · wechat · 2026-10-22
Researchers from CUHK MMLab, ByteDance, and HKUST propose DuoMatching, which trains few-step video generators with both joint video distribution matching and frame-level marginal matching against an image teacher, injecting visual/semantic priors without extra inference cost.
- LatentBridge: a lightweight module maps temporally compressed video latents to single-frame latents the image teacher can consume (motion score ablation: 76.94 → 93.61)
- LVS: samples frames across temporally segmented bins so limited supervision covers diverse content states
- Both modules are training-only; inference stays unchanged
Results: in the per-frame 2-step setting, VBench total rises from 80.67 to 83.53 (+2.86 over CausalForcing++) while motion dynamics are preserved. Human eval (23 participants, 40 pairs each) shows preference rates of 91.30% vs CausVid, 80.87% vs CausalForcing++, 94.35% vs One-Forcing, and 83.91% vs RewardForcing.
Paper, code, and weights are public.
More from Multimodal
- Meshy offers Hermes users a month of Pro with API access for $1 — Teknium · 2026-10-23
- Customuse launches Smart Mesh AI with geometry-aware adaptive character clothing tool — SimplyAnnisa · 2026-10-23
- AI-generated movie poster 'The Chase' crams a wolf, a plane and muscle cars into one chase — Aggravating-Sir-115 · 2026-10-23
- SteerTTS-edge: an 11.84MB steerable TTS with unchanged WER running 19x realtime on CPU — Working_Hat5120 · 2026-10-23
- ElevenLabs' Chimera pairs Eleven v4 voice with Seedance video to make character images speak — AIwithGhotai · 2026-10-23
- Open-Source Tennis Models Plus Claude Turn Video Into a World-Class Coach — tweetsatpreet · 2026-10-23