Few-Step Generation via Data-Space Iteration
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.
Key points
- Existing few-step methods perform their iterative computation along the probability flow and therefore require a fixed, manually chosen timestep discretization.
- We introduce data-space iteration, a few-step generation framework that removes flow discretization altogether.
- Our formulation integrates with distribution matching distillation (DMD) with minimal changes, enabling a controlled comparison between iteration methods under matched training settings.
- 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.
Sources (1)
- [1]Few-Step Generation via Data-Space IterationarXiv (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.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 8, 2026Finsler Flow Matching: Dynamics-Aware Geodesic Interpolation for Single-Snapshot Trajectory Inference
- Oct 8, 2026One Block, Multiple Depths: Recurrent Vision Transformers with Depth-Programmed Experts
- Oct 8, 2026Normalizing Trajectory Models
- Oct 7, 2026Velocity Scaling in Flow Matching
- Oct 7, 2026MUNITE: Unified Multimodal Latent Inference for Any-to-Any Multimodal Generation
- Oct 6, 2026Consistent Distribution Matching for Data-Free Diffusion Distillation