LongCat-Image - An open-source image generation model launched by Meituan

What is LongCat-Image?

LongCat-Image is a high-performance image generation model open-sourced by Meituan, achieving top-tier open-source performance in text-based image generation and image editing with only 6B parameters. The model employs an innovative architecture and training strategy, supporting high-quality Chinese text rendering, covering 8105 Chinese characters, and is suitable for design scenarios such as posters and advertisements. Through multi-task learning and adversarial training, the model enhances image realism and texture details, providing a complete toolchain from pre-training to fine-tuning, helping developers explore more possibilities for visual generation with low barriers to entry.

Main functions of LongCat-Image

• Text-to-Image:Generates high-quality images based on user-input text descriptions, supports multiple styles and scenarios, and is suitable for creative design, social media content creation, and more.
• Image Editing:It offers powerful image editing capabilities, supporting style transfer, attribute editing, composition adjustment, and more. It can precisely modify image content according to user instructions, making it suitable for fields such as design, advertising, and film and television post-production.
• Chinese text rendering:It has been specially optimized to generate Chinese characters, covering 8,105 Chinese characters in the general standard Chinese character table, and supports rendering of complex strokes and rare characters. It is suitable for scenarios such as poster design, sign making, and illustration of ancient poems.
• Enhanced realism and texture detail:Through systematic data filtering and adversarial training, the generated images have higher realism and texture detail, avoiding a "plastic" texture.
• Low-barrier development and application:It provides a complete toolchain from pre-trained models to fine-tuning code, and supports advanced development features such as SFT and LoRA, making it convenient for developers to perform secondary development and customization.

The technical principles of LongCat-Image

• Architecture Design:It adopts an architecture design that integrates text generation and image editing, and achieves efficient collaborative improvement through a compact 6B parameter scale, balancing instruction compliance accuracy, image quality, and text rendering capabilities.
• Progressive learning strategy:During the pre-training phase, multi-source data and instruction rewriting strategies are used to improve the model's ability to understand diverse instructions.In the SFT stage, manually calibrated data is introduced to further improve the accuracy and generalization of instruction compliance. In the RL stage, an OCR and aesthetic dual-reward model is incorporated to optimize text accuracy and the naturalness of background blending.
• Data Engineering and Training Paradigms:By rigorously screening pre-trained data, the "plastic" texture of the generated images is avoided.In the SFT stage, data is manually screened to align with popular aesthetics, enhancing the realism and beauty of the generated images.We innovatively introduce an AIGC content detector as a reward model, using adversarial signals to guide the model to learn real-world physical textures and lighting effects.
• Chinese text generation optimizationThe program employs a course-based learning strategy, learning glyphs during the pre-training phase, covering 8105 characters from the general standard Chinese character list. The SFT phase incorporates real-world text image data to improve the generalization ability of fonts and typography layout. The RL phase further enhances text accuracy and the naturalness of background blending.

LongCat-Image project address

• GitHub repository: https://github.com/meituan-longcat/LongCat-Image
• HuggingFace model libraryhttps://huggingface.co/meituan-longcat/LongCat-Image
• Technical Papers: https://github.com/meituan-longcat/LongCat-Image/blob/main/assets/LongCat_Image_Technical_Report.pdf

Application scenarios of LongCat-Image

• Poster designIt can quickly generate high-quality posters based on creative copy, supporting text rendering and style customization to meet the needs of advertising, event promotion, and other purposes.
• Advertising material productionGenerate attractive advertising images for brands, supporting different scenarios and styles, and reducing advertising production costs.
• Film and television concept artGenerate movie posters, concept art, and scene design drawings for film and television production, and assist in scriptwriting and visual effects design.
• Teaching aidsThe model can generate images related to the teaching content, such as historical scenes and scientific experiment illustrations, to help students better understand and remember knowledge.
• Style conversion and enhancementEdit personal photos by changing styles, replacing backgrounds, and beautifying people to meet individual needs.