Few-Step Generation via Data-Space Iteration

Key points

• Existing few-step methods perform their iterative computation along the probability flow and therefore require a fixed, manually chosen timestep discretization.1
• We introduce data-space iteration, a few-step generation framework that removes flow discretization altogether.1
• Our formulation integrates with distribution matching distillation (DMD) with minimal changes, enabling a controlled comparison between iteration methods under matched training settings.1
• On class-conditional ImageNet 256x256, data-space iteration outperforms standard discretization baselines and matches or improves upon variants selected through schedule search, without requiring schedule-specific training.1

Sources (2)

• [1]Few-Step Generation via Data-Space Iteration
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 8, 03:03 PM

  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.

  Existing few-step methods perform their iterative computation along the probability flow and therefore require a fixed, manually chosen timestep discretization.

• [2]Few-Step Generation via Data-Space Iteration
  Hugging Face Daily Papers · Oct 8, 12:00 AM · same content

Extractive summary: sentences quoted from the sources.