Download
MOSS-Image is an open-source, 5-billion-parameter text-to-image generation model built on a Diffusion Transformer architecture. It creates high-quality photorealistic and anime-style images from natural language descriptions. The model specializes in accurate text rendering and structured visual compositions. Users can control the placement of text, objects, and other elements through JSON-based layouts. It supports detailed prompts of up to 4,000 tokens for complex scenes. Two model variants, Base and SFT, support general generation and improved visual organization. Developers can run inference through Python, Gradio, ComfyUI, SGLang, and vLLM-Omni. Fine-tuning workflows support custom datasets and LoRA adaptation, while configurable memory modes enable deployment on different hardware.
Features
- 5B-Parameter Diffusion Transformer Architecture
- Photorealistic and Anime-Style Image Generation
- Accurate Text Rendering Within Images
- JSON-Based Spatial Layout and Composition Control
- Long-Form Prompts up to 4,000 Tokens
- LoRA Fine-Tuning and Multiple Inference Integrations
Project Samples
