What it is
Distills diffusion and flow models into 4-8 step image/video generators by predicting multiple denoising steps per network call, improving speed and video diversity. Instead of VSD/GAN distillation or collapsing steps into one big update, it predicts several teacher trajectory steps in parallel per eval. It distills mean velocities with a simple regression loss, no JVPs, finite differences, or extra policy head, and supports variable NFE.
Key results
- SOTA with 4-8 NFE on LTX-2.3 T2V/audio, Wan 14B T2V, and Qwen-Image
- SOTA video quality at 4 NFE on Wan2.1 1.3B/14B using VBench
- Improves video diversity over DMD2, AnyFlow, and official distilled baselines
Few-step generation can match top quality without adversarial or VSD losses if the student learns parallel mean-velocity segments of the teacher trajectory.