Kuaishou Kolors

Über Kuaishou (Kolors Team)

Gegründet 2011 · Beijing, China

Kuaishou Technology (快手) is a Chinese short-video platform founded in 2011 in Beijing by Su Hua and Cheng Yixiao, listed on the Hong Kong Stock Exchange since 2021. Its AI research division, often referred to as the Kuaishou KwaiVGI or Kling AI team, has produced several generative-media foundation models including Kling (video) and Kolors (image). Kolors was open-sourced in July 2024 with weights released on Hugging Face and GitHub (Kwai-Kolors/Kolors). The model is positioned as the best open Chinese-English bilingual text-to-image model, particularly strong on Chinese-language prompts, calligraphy and culturally specific content.

Architektur

Latent diffusion U-Net with ChatGLM3-6B text encoder (bilingual Chinese-English)

Kolors is a latent diffusion text-to-image model released open-source by Kuaishou in July 2024. The architecture follows the Stable Diffusion XL family — a latent-space U-Net with cross-attention conditioning on text embeddings — but replaces the standard CLIP text encoder with the much larger ChatGLM3-6B, a bilingual Chinese-English instruction-tuned LLM developed by Tsinghua KEG and Zhipu AI. This gives Kolors strong understanding of Chinese-language prompts, idioms, calligraphy and culturally specific content (food, festivals, Chinese architecture, traditional clothing) that Western models like SDXL and FLUX often struggle with. The U-Net has ~2.6B parameters and is trained at 1024x1024 resolution with the standard noise-prediction objective. Kuaishou also released a Kolors-Inpainting variant and a Kolors-ControlNet collection (Canny/Depth). The model is open-source under the Apache 2.0 license, making it the leading commercial-use open bilingual model.

Parameter

~2.6B U-Net + 6B ChatGLM3 text encoder

Kontext

256 Token

Funktionen

• Best-in-class understanding of Chinese-language prompts
• Strong English performance comparable to SDXL
• 1024x1024 native resolution
• Open weights under Apache 2.0
• Excellent rendering of Chinese characters and calligraphy
• Culturally accurate Chinese-food, festival, architecture and clothing imagery
• Companion models: Kolors-Inpainting, Kolors-ControlNet (Canny/Depth/Pose)
• Best for: Chinese-market apps, bilingual creative tools, culturally specific content, self-hosted Chinese services.

Training & Lizenz

Trained on a large mixed Chinese-English image-text corpus including high-quality Chinese-captioned data. Exact composition is not disclosed.

Lizenz: Apache 2.0 — open weights, full commercial use permitted including redistribution and derivatives.

Sicherheitstests: Internal safety filters; the model adheres to Chinese content regulations including filters for sensitive political content. Operators outside China should apply their own moderation.