OpenMOSS releases MOSS-Image, an open text-to-image model with layout control
MOSS-Image is a new open-weight text-to-image family from OpenMOSS, with Base and SFT checkpoints under Apache 2.0. Besides plain prompts, it accepts tag lists and JSON layouts that place text and objects with bounding boxes.
What it does
MOSS-Image-Base is the base text-to-image model, and MOSS-Image-SFT is fine-tuned from it on curated data. You can prompt in natural language, with tag sequences, or with a JSON layout that sets the caption, the rendered text and the position of each element, which suits posters, whiteboards and diagrams. Prompts can be up to 4K tokens long.
It runs through Diffusers, a Gradio app with GPU-resident, component and sequential memory modes, and a native ComfyUI plugin you can install from ComfyUI Manager.
Results and limits
The team's own comparison shows SFT grouping text more clearly than Base on a whiteboard layout; there are no independent benchmarks yet. Comfy-Org published repackaged ComfyUI files with INT8 builds on the same day, and a community GGUF quant with nodes is labelled experimental by its author.
Worth knowing
The recommended settings are 50 steps, CFG 3 and shift 5, with negative-branch layer skipping at layer 22, at 1 to 2 megapixels. It needs an NVIDIA GPU with CUDA and PyTorch 2.9 or later; if memory runs short, the Gradio app's sequential mode or a smaller image size helps. The Base model is meant for LoRA training and fine-tuning. WanGP added MOSS Image in v18.17, with a layout editor.
Try it yourself
In ComfyUI, load the example moss-image.json workflow and prompt: "A cinema poster for [film title]: the title '[TITLE]' in large letters at the top, [main character] in the center, [setting] behind, small credits text at the bottom." Keep 1024×1024, 50 steps, CFG 3.