Few-Step Generation via Data-Space Iteration
Data-space iteration enables few-step image generation without tuning a flow timestep schedule.
The paper introduces data-space iteration for few-step generative sampling, removing the need for a manually chosen probability-flow timestep discretization. A shared generator starts from noise and iteratively refines its prediction in data space, with each step trained to produce the best sample its capacity allows. The method plugs into distribution matching distillation with minimal changes. On class-conditional ImageNet 256x256 it outperforms standard discretization baselines and matches or improves schedule-searched variants without schedule-specific training.
- Data-space iteration refines samples without a fixed flow timestep schedule.
- It integrates with distribution matching distillation with minimal changes.
- On ImageNet 256x256 it matches or beats schedule-searched flow baselines.
From the original article
1 min read at huggingface.co
Flow matching has emerged as a scalable paradigm for training high-quality generative models, but sampling from the learned probability flow requires many network evaluations. Distillation can reduce this cost to one or a few evaluations; however, one-step generation often sacrifices quality, making few-step generation the practical operating regime.
By Shanchuan Lin, Yansong Peng, Fu-Yun Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.