Kolors: Effective Training for Realistic Text-to-Image Synthesis Diffusion Models (arXiv-2024)

Kolorsis a large-scale text-to-image generation model based on latent diffusion developed by the Kolors team at Kwai. Trained on billions of text-image pairs, Kolors demonstrates significant advantages in visual quality, complex semantic accuracy, and rendering of Chinese and English texts over both open-source and proprietary models. Additionally, Kolors supports Chinese and English inputs and excels in understanding and generating Chinese-specific content.

Evaluation

We have curated a comprehensive text-to-image evaluation dataset called KolorsPrompts for comparing Kolors with other state-of-the-art open-source and proprietary text-to-image models. KolorsPrompts consists of over 1,000 prompts across 14 categories and 12 evaluation dimensions. The evaluation process combines manual and automated assessments. In relevant benchmark evaluations, Kolors demonstrates highly competitive performance, reaching industry-leading standards

Official introduction https://kwai-kolors.github.io/post/post-2/