Kolors

Developed by Kuaishou's Kolors team, Kolors is a large scale latent diffusion text-to-image model trained on billions of text and image pairs, with genuinely bilingual understanding of Chinese and English prompts including the ability to render Chinese characters inside generated images. It accepts prompt context up to 256 tokens and scored strongly in human preference evaluations against contemporary open models. The surrounding ecosystem includes IP-Adapter-Plus for image prompts, an IP-Adapter-FaceID variant for identity preserving portraits, ControlNet conditioning on Canny edges, depth, and pose, an inpainting model, image-to-image generation, and LoRA and DreamBooth fine-tuning, with integration into Hugging Face Diffusers. The code is Apache 2.0 licensed, while the weights are free for academic research and require registration with Kuaishou for commercial use above certain scale thresholds. Teams needing accurate Chinese prompt handling are frequent adopters.