Continuous-Time Distribution Matching: A new SOTA method for step distillation.
New method reduces diffusion sampling steps by 10x with higher quality.
Deep Dive
Researchers propose Continuous-Time Distribution Matching (CDM) for diffusion models. According to the paper, CDM matches distributions in continuous time and works across architectures, enabling fast inference.
Key Points
- CDM reduces diffusion sampling to 1–2 steps while maintaining FID ≤ 2 on CIFAR-10.
- Uses continuous-time distribution matching via reverse KL divergence, avoiding discretization errors.
- Works with latent diffusion and pixel-space models; achieves 4-step text-to-image generation comparable to 50-step DDIM.
Why It Matters
Faster, higher-quality generation from diffusion models unlocks real-time image/video synthesis with lower compute cost.