Kuaishou’s Kolors Text‑to‑Image Model: Architecture, Evaluation, and Real‑World Applications

In the AICon Beijing event, Li Yan, head of Kuaishou’s Kolors large model, introduced the model’s development, emphasizing the importance of multimodal capabilities for enterprise efficiency.

Data side: High‑quality, large‑scale Chinese image‑text data, sourced from premium providers like Shutterstock, is crucial. Emphasis is placed on data volume, concept coverage, image quality, and strong image‑text relevance, as well as rigorous safety measures to prevent unsafe concept combinations.

Model side: The current mainstream frameworks are diffusion‑based (Stable Diffusion’s U‑Net) and the newer Diffusion Transformer (DiT) introduced with Sora. Kolors adopts a flexible architecture covering denoising theory (DDPM, EDM, RF), samplers (DDIM, Euler, LMS, DPM‑Solver), parameter scales (1B‑10B), and text encoders (CLIP, LLM) to support one‑stage, two‑stage, or multi‑stage generation.

Effectiveness side: Kolors has progressed through five versions, surpassing Midjourney‑V5 in the GSB (Image Generation Quality) metric and achieving a subjective score of 75.23 on the FlagEval benchmark, ranking second globally after DALL‑E 3.

Evaluation methodology: Both human (GSB) and machine metrics (CLIP similarity, FID, aesthetic scores) are used. Human evaluation provides reliable preference modeling, while machine metrics serve as early‑warning signals for model degradation.

Derived capabilities: Prompt refinement, open‑domain text rendering, and interactive visual generation enable richer user experiences, reducing the barrier from “spell‑like” prompts to natural language.

Application practices: Six major use cases are described: AI‑powered comment generation (AI 玩评), AI portrait creation, IP customization, image‑fusion, AI‑driven upscaling, and live‑stream AIGC (gifts and backgrounds). Each showcases how Kolors integrates with control modules, temporal modules, and ID‑preservation techniques.

Strategic advice for domestic visual‑generation peers: Anticipate a unified image/video generation framework within a year, launch applications in parallel with base‑model research, prioritize high‑quality data, enforce early safety governance, clarify target user groups, and explore both “old‑business + AIGC” and “AIGC‑driven new‑business” models.

The Kolors model and its code have been fully open‑sourced on GitHub, Hugging Face, and a dedicated website, quickly gaining community traction with thousands of stars and downloads, and positioning itself as a competitive open‑source alternative to leading proprietary models.