Kolors

Kolors is Kuaishou’s open-weight text-to-image model, released in July 2024. It is a latent diffusion model with about 2.6 billion parameters, built on a U-Net architecture, and it uses ChatGLM as its text encoder instead of a CLIP-style encoder, which gives it a notable edge at reading and rendering Chinese-language prompts and Chinese text within images. It generates at a native 1024x1024 resolution and supports multiple aspect ratios. Kuaishou trained it on billions of text-image pairs and released the weights on GitHub and Hugging Face, free for research use, with a separate registration process required for commercial deployment. Alongside models like Flux and Stable Diffusion 3, Kolors is one of the few widely-used open image models that handles non-Latin scripts well.