SnapFlow: One-Step Action Generation for Flow-Matching VLAs via Progressive Self-Distillation
SnapFlow is a self-distillation method that compresses multi-step action generation in flow-matching vision-language models into a single, fast forward pass. It mixes standard flow-matching supervision with consistency samples derived from the model's own two-step Euler predictions. This approach achieves significant speedups, up to 10x faster denoising, while maintaining or improving task success across various VLA architectures.