VA-TDM: Velocity Adversarial Distillation Sets SOTA for Few-Step Diffusion Models
Yihong Luo · hf · 2026-11-05
Researchers from HKUST and CUHK-Shenzhen propose VA-TDM, a unified framework combining velocity adversarial distillation with trajectory-aware supervision for few-step diffusion and flow generation.
- Replaces conventional distillation objectives with an adversarial loss in velocity space, enabling high-quality generation in very few steps
- Qualitative comparisons show VA-TDM at low NFE matches or exceeds MinMAX-H3 at 100 NFE
- Authors report state-of-the-art results for few-step generation across benchmarks
More from Multimodal
- Ant Open-Sources Ming-3.0 Audio Family: TTS with Voice Cloning & Emotion Control, MIT Licensed — Promptmethus · 2026-11-05
- AI-generated music video 'Soulless' shows a cyborg singer in a cyberpunk city — BotLordBotflix · 2026-11-05
- AI-generated anime cooking clip of spring rolls shows off video model animation skills — Plenty-Reindeer-3251 · 2026-11-05
- Dev uses new image model to redo game art, says OpenAI has no image-gen rival — CtrlAltDwayne · 2026-11-05
- AI-generated anime music video 'Jealousy' showcases cinematic AI animation — Few_Beautiful2589 · 2026-11-05
- 31.7MB Open-Weights TTS Model Woof Powers an Entire Browser Game, Fully On-Device via WebGPU — brandon_xyzw · 2026-11-05