DownloadMOSS-Image-SFT is a 5B-parameter open-weight text-to-image generation model developed by OpenMOSS, derived from MOSS-Image-Base through supervised fine-tuning. It improves instruction following, text rendering, and spatial composition while retaining support for natural-language descriptions, tag sequences, and structured JSON layouts. Its layout-to-image capabilities enable precise control over object placement, bounding boxes, and text positioning, making it suitable for posters, illustrations, diagrams, and other visually structured content. The model supports English and Chinese prompts and uses flow-matching Euler sampling with 50 inference steps and a default guidance scale of 3. It incorporates the FLUX.2 small VAE decoder and supports customizable resolutions, negative prompts, and LoRA adaptation. MOSS-Image-SFT integrates with Hugging Face Diffusers and ComfyUI, supports local NVIDIA GPU inference, and is released under the Apache 2.0 license.
Features
- 5B-parameter supervised fine-tuned image generation model
- Text-to-image generation from natural-language descriptions
- English and Chinese prompt support
- Tag-based prompting for detailed visual composition
- Structured JSON layouts for text and object positioning
- Improved text rendering and spatial organization over the base model
- Flow-matching Euler sampling with configurable guidance
- LoRA training, Diffusers, and ComfyUI integration
Project Samples
