Latent Consistency Models

Latent Consistency Models cut the number of inference steps needed for high quality image generation from the usual 20 to 50 down to as few as 1 to 4. The technique distills a pretrained latent diffusion model into a consistency model, folding classifier-free guidance into the model inputs during distillation so that sampling becomes nearly a single pass. The official repository from Simian Luo and collaborators provides training and inference code, with distilled checkpoints based on Stable Diffusion v1.5 and Dreamshaper v7 generating at 768x768 resolution. The companion LCM-LoRA method packages the acceleration as a low rank adapter that attaches to existing SD 1.5, SDXL, and SSD-1B checkpoints without retraining. Support is integrated into Hugging Face Diffusers, ComfyUI, and the Stable Diffusion WebUI, making near real time generation practical on consumer GPUs. Researchers and builders of interactive image tools adopt LCMs where latency matters more than maximum fidelity. MIT license.