Data-space iteration proposed for few-step image generation

In an arXiv paper, researchers introduce data-space iteration, which starts from noise and repeatedly refines a prediction rather than iterating along a probability flow. They say it removes the need for a fixed, manually chosen timestep schedule. On class-conditional ImageNet 256x256, they report outperforming standard discretization baselines and matching or improving variants selected through schedule search, without schedule-specific training.