Meituan releases LongCat-Image image generation model, achieving top-tier editing capabilities in the open-source state-of-the-art.

Meituan has released and open-sourced the LongCat-Image image generation model, which, with 6B parameters, approximates larger models in terms of text-based image generation and image editing capabilities. Key technical highlights include: employing a homogeneous architecture and progressive learning to achieve state-of-the-art (SOTA) controllability in image editing; significantly improving the coverage and accuracy of Chinese text generation through course learning and character-level encoding; and enhancing image realism through data filtering and adversarial training.