DownloadMOSS-Image-Base is an open-weight, 5B-parameter text-to-image generation model developed by OpenMOSS, designed for visual content creation, structured composition, and model customization. It accepts natural-language descriptions, tag sequences, and JSON layouts to control scene elements, text placement, and object arrangement. Its layout-to-image capabilities allow developers to specify objects, bounding boxes, and rendered text for precise spatial composition. Unlike the supervised fine-tuned MOSS-Image-SFT variant, this base checkpoint provides a foundation for further training and adaptation. It uses flow-matching Euler sampling with 50 inference steps by default and incorporates the FLUX.2 small VAE decoder. The model supports English and Chinese prompts and can generate images at configurable resolutions. It integrates with Hugging Face Diffusers and ComfyUI, supports local CUDA inference and LoRA training, and is released under the Apache 2.0 license.
Features
- 5B-parameter open-weight image generation model
- Text-to-image generation from natural-language prompts
- English and Chinese prompt support
- Tag-based prompting for detailed visual composition
- JSON layout controls for text and object positioning
- Flow-matching Euler sampling with configurable guidance
- Hugging Face Diffusers and ComfyUI integration
- LoRA fine-tuning support with Apache 2.0 licensing
Project Samples
