SDXL in 4 steps with Latent Consistency LoRAs

Hugging Face Blog · 2 years ago

Researchers released Latent Consistency LoRAs, small adapter modules that reduce SDXL image generation from 25-50 steps to 4-8 steps by applying distillation techniques without retraining the full model. On an M1 Mac, generating a 1024x1024 image takes approximately 6 seconds with LCM LoRA compared to 64 seconds with standard SDXL, and on a 4090 GPU it takes under 1 second versus 3.4 seconds. The method makes image generation faster and cheaper, enabling real-time applications and broader access to generative tools across different hardware including CPUs and older GPUs.