Data-space iteration removes timestep schedules from few-step image generation

Image generators often use four to eight network evaluations to balance fidelity and latency. One-step models can lose detail, while samplers with dozens of evaluations cost more to serve. A new paper by Shanchuan Lin, Yansong Peng, Fu-Yun Wang, and Haoqi Fan proposes data-space iteration, which repeatedly refines an image prediction without following a predefined noise schedule.

On class-conditional ImageNet at 256 × 256 resolution, the authors report that the method outperforms standard timestep discretizations and matches or exceeds variants chosen through schedule search. The same trained model can also run with different iteration counts, removing the need to distill a separate checkpoint for each schedule.

Why four steps still need a schedule

Flow matching trains a neural network to predict a velocity field that transports Gaussian noise toward the data distribution. During inference, a numerical solver evaluates that field at a sequence of timesteps and advances the sample along the learned probability flow. Models including Stable Diffusion 3 and Flux use related flow-based training methods.

Few-step distillation compresses that sampling process into a small number of network evaluations. Existing methods commonly assign those evaluations to fixed points such as [1.0, 0.75, 0.5, 0.25]. Every sample receives the same sequence, even when images or spatial regions require different amounts of refinement.