SDXL and Stable Diffusion Fast Inference with Latent Consistency LoRAs

Hugging Face has introduced Latent Consistency LoRAs (LCM LoRAs), a method that enables Stable Diffusion and SDXL models to generate high-quality images in just 4 to 8 steps, compared to the typical 25 to 50 steps. This advancement significantly reduces inference time, allowing for near real-time image generation on high-end GPUs and drastically improving accessibility on lower-end hardware.

LCM LoRA Method Overview

LCM LoRAs utilize Low-Rank Adaptation (LoRA) to implement latent consistency distillation without requiring the full model to be distilled separately. Instead of a costly and data-intensive full-model distillation process, a small number of adapter layers are trained. These resulting LoRAs can be applied to any fine-tuned version of the base model, providing the benefits of latent consistency distillation across various model versions.

The implementation process involves three primary steps:

  1. Selecting a teacher model (e.g., SDXL base or a fine-tuned version).
  2. Training an LCM LoRA using Parameter-Efficient Fine-Tuning (PEFT).
  3. Applying the LoRA with any SDXL diffusion model and the LCMScheduler for high-quality, low-step inference.

Performance Benchmarks and Hardware Impact

LCM LoRAs reduce image generation time by an order of magnitude across various hardware configurations. For a 1024x1024 image, the speed improvements are as follows:

Hardware SDXL LoRA LCM (4 steps) SDXL standard (25 steps)
Mac, M1 Max 6.5s 64s
2080 Ti 4.7s 10.2s
3090 1.4s 7s
4090 0.7s 3.4s
T4 (Google Colab Free) 8.4s 26.5s
A100 (80 GB) 1.2s 3.8s
Intel i9-10980XE CPU 29s 219s

On an NVIDIA RTX 4090, response times are under one second, enabling the use of SDXL in applications requiring real-time interaction. On M1 Macs, generation time drops from approximately one minute to roughly 6 seconds.

Quality, Guidance, and Model Compatibility

Inference Quality vs. Step Count

While 1-step generation produces only approximate shapes without texture, quality improves rapidly between 4 and 6 steps. In comparison, standard SDXL pipelines typically produce unusable images until approximately 20 steps, with peak detail requiring 50 steps.

Guidance Scale and Negative Prompts

For maximum speed, a guidance_scale of 1 is recommended, which effectively disables guidance. To utilize negative prompts, a guidance scale between 1 and 2 can be used; however, values larger than 2 are generally ineffective.

Compatibility with Fine-tuned Models

LCM LoRAs are compatible with any fine-tuned SDXL or Stable Diffusion model. For example, the technique can be applied to collage-diffusion (a Dreambooth fine-tune of Stable Diffusion v1.5) by loading the corresponding LCM LoRA for SD v1.5, enabling 4-step inference for specialized styles.

Integration and Advanced Workflows

Diffusers Integration

LCM is fully integrated into the diffusers library, providing access to:

  • Native mps support for Apple Silicon.
  • Performance optimizations including torch.compile() and flash attention.
  • Memory-saving strategies such as model offload for low-RAM environments.
  • Support for ControlNet and image-to-image workflows.

Combining LoRAs

Using the diffusers and PEFT integration, users can combine LCM LoRAs with regular SDXL LoRAs. This allows specialized style or subject LoRAs (such as CiroN2022/toy_face) to benefit from the 4-step inference speed of the LCM process by setting multiple adapters and their respective weights.

Available Models and Resources

Hugging Face has released several LCM LoRAs and full models:

  • LCM LoRAs: Available for SDXL 1.0 base, Stable Diffusion v1.5, and Segmind's SSD-1B (a distilled SDXL model).
  • Full Models: Full fine-tuned consistency models derived from SDXL 1.0 base and SSD-1B.

Training and fine-tuning scripts are now available within the diffusers repository for both Stable Diffusion 1.5 and SDXL, allowing the community to perform either full-model distillation or the more accessible LCM LoRA training.

Sources