OpenMOSS releases open-source MOSS-Image-Base text-to-image model with JSON layout control
NERDDISCO · x · 2026-10-27
The OpenMOSS team has released MOSS-Image-Base, an open-source (Apache 2.0) text-to-image base model on Hugging Face, supporting English and Chinese.
It accepts natural-language prompts, tag sequences, and JSON layouts for controlling text placement and object arrangement — the demo shows multi-line handwritten Chinese text on a sticky note. An SFT variant produces clearer text grouping and more organized layouts.
Quick start uses a custom Diffusers pipeline (MOSSImagePipeline) with Python 3.10+, PyTorch 2.9+ and an NVIDIA GPU; transformer and text encoder run in bfloat16, with full example code and ComfyUI support provided.