Paper summary: LCM-LoRA: A Universal Stable-Diffusion Acceleration Module

Problem

Diffusion models need many denoising steps per image. Latent Consistency Models (LCMs) cut this to a handful by distilling a pretrained latent diffusion model, and the report notes the distillation needs only about 32 A100 GPU hours. But each LCM was tied to the model it was distilled from.

Method

Two extensions:

1. Do the LCM distillation with LoRA (low-rank adapters) instead of full fine-tuning. This brings it to larger models (SD-V1.5, SSD-1B, SDXL) with much less memory.
2. Observe that the resulting LoRA parameters are a universal acceleration module, LCM-LoRA. It can be added to other fine-tuned Stable Diffusion models or combined with style LoRAs, without retraining.

The authors compare it to numerical solvers like DDIM and DPM-Solver. LCM-LoRA acts as a learned solver for the probability-flow ODE.

Key results

The report shows few-step, high-quality generation across SD variants. It claims strong generalisation when plugging LCM-LoRA into fine-tuned models and LoRAs that weren’t seen during distillation.

Limitations

It’s a short technical report without a dedicated limitations section, so the strength of the “universal” claim rests on the variants and examples shown in the report.