Latent Consistency Models (LCM) is the official open-source repository for the LCM research project, a distillation technique that lets diffusion models synthesize high-resolution images in just a few inference steps instead of dozens.
Standard diffusion models produce excellent images but are slow, needing many sequential denoising steps. Latent Consistency Models distill a pretrained diffusion model into a consistency model that maps noise toward the final image much more directly, enabling high-quality text-to-image and image-to-image generation in as few as one to four steps. The project’s best-known spin-off is LCM-LoRA, a small universal acceleration module that can be plugged into Stable Diffusion pipelines to speed them up without retraining the base model. The official LCM pipeline and scheduler were also integrated into the Hugging Face Diffusers library, which is how most practitioners use the technique today. The repository itself includes the model code, training scripts and a local Gradio demo for trying the models interactively. The technique influenced a wave of real-time image applications, and LCM-LoRA in particular became a standard trick for accelerating community Stable Diffusion workflows in tools built on Diffusers and ComfyUI alike. The models remain a common baseline in few-step generation research. It is MIT-licensed, has more than 4,600 stars on GitHub, and its last major updates landed in 2024, reflecting its status as a completed, widely cited research release.
Free. The Latent Consistency Models repository is open source under the MIT license; models run on your own hardware.
Image-generation developers and researchers who need near-real-time diffusion output, interactive creative tools where latency matters, and anyone studying consistency-model distillation techniques.