Latent Consistency Models: real-time image generation in 4 steps
In one sentence Tsinghua University publishes LCM: distillation of a diffusion model reducing sampling from 50 steps to 4 with minimal quality loss. LCM-LoRA makes any SD model 10x faster. First technique enabling real-time generation on consumer hardware.
A standard diffusion model generating an image takes 20-50 denoising "steps", each a full neural network inference. On a consumer GPU that's 5-30 seconds. Not exactly "real-time".
Latent Consistency Models solve this with a distillation technique: a trained model is "taught" to skip intermediate steps and land directly on a good-quality image in 4 steps. Like teaching someone to do mental math instead of going through the calculator step by step.
The result is dramatic: image generation in 1-2 seconds on home GPUs, with quality comparable to the standard model. LCM-LoRA applies this technique to any existing SD model via a few-MB file. For the first time, generating images in real time as you type a prompt becomes possible on ordinary hardware.
Companies
Tsinghua University
Tools
—
Tags
Sources