MOSS-Image-SFT: Precise Control of Text and Layout in Images via JSON

By inputting JSON-formatted layout data into the prompt, you can structurally control text placement and object arrangement within images. This approach accurately implements complex compositions that were difficult to achieve with natural language descriptions alone, using bounding box coordinates.

1 / 3

This is a 5B parameter model fine-tuned via Supervised Fine-tuning on OpenMOSS's MOSS-Image-Base, easily integrated through the Diffusers library. It combines the FLUX.2 small decoder to maintain high-resolution output quality while improving GPU memory efficiency with bfloat16 computation.

A separate Prompt Enhancer skill automatically converts simple ideas into detailed captions or spatial layout prompts. By directly inputting the generated JSON, you can achieve sophisticated visual compositions such as text grouping and whiteboard-style layouts.

Licensed under Apache 2.0, commercial use is permitted, and native ComfyUI plugin support is currently in progress. It is highly valuable as an alternative to existing Text-to-Image models for tasks where text placement is critical, such as design mockup generation or document image creation.