Data-Space Iteration: few-step flow matching generation without timestep discretization
kastnerkyle · x · 2026-10-11
A new arXiv paper by Shanchuan Lin, Yansong Peng, Fu-Yun Wang and Haoqi Fan introduces Data-Space Iteration for few-step Flow Matching generation. Existing few-step methods iterate along the probability flow and require manually tuned timestep discretizations; this framework removes discretization entirely, having a shared generator directly refine predictions in data space, with each iteration trained to produce the best sample its capacity allows. It integrates with DMD distillation with minimal changes, and on class-conditional ImageNet 256x256 outperforms standard discretization baselines while matching or beating schedule-search variants without schedule-specific training.
More from Multimodal
- Perplexity demos AI-enriched public-domain books with images, video and context — AravSrinivas · 2026-10-11
- AI-Generated Music Video: Same For You — Mr_Darkmann · 2026-10-11
- AI-Generated Short Film: Myths of Aztleau — Advanced-Power-1775 · 2026-10-11
- Kroma released: a Krea-2 fine-tune trained for 3 months locally on RTX Pro 6000 GPUs — KadahCoba · 2026-10-11
- DeemosTech extends 3D Edit to textures with prompt-based partial redo — sidahuj · 2026-10-11
- H3 video editing shows spatial and temporal shifts; FFLF model and lower res fare best in tests — Comfortable-Lab4125 · 2026-10-11