LCMs

Latent Consistency Models (LCMs) distill diffusion models like Stable Diffusion, enabling high-quality image generation in 2–4 steps for near real-time performance.

LCMs fundamentally accelerate text-to-image generation: they solve the probability flow ODE (PF-ODE) in latent space, bypassing the slow, iterative sampling of traditional Latent Diffusion Models (LDMs). This is a massive efficiency win. We’re seeing distillation from a pre-trained Stable Diffusion model in just 4,000 training steps (about 32 A100 GPU hours) to achieve high-resolution, 768x768 images. The result: a task that once required 50+ steps now executes in 2 to 4 steps, making real-time AI applications viable for the first time.

https://huggingface.co/latent-consistency