LCM-LoRA: From Diffusion Models to Fast SDXL with Latent Consistency Models

LCM-LoRA is a method that dramatically accelerates image generation in large text-to-image models like SDXL by combining Latent Consistency Models (LCMs) with LoRA adapters. Instead of retraining the full model weights, only a small set of LoRA adapter weights are trained and added to the frozen pre-trained model. This enables high-quality image generation in just a few denoising steps rather than hundreds. A key advantage is that the LCM-LoRA weights can be combined with other style-tuned LoRA weights, acting as a universal acceleration module for Stable Diffusion models without requiring additional training for each style variant.