Description
The hard part of AI posters is the text: crooked titles, small print melting into mush, Chinese characters missing strokes, and ten generations with not one you can use. MOSS-Image from OpenMOSS goes straight at that, a 5B-parameter text-to-image model built to get the lettering right and the layout where you put it.
What sets it apart is layout control: besides plain language, the prompt can be a JSON description where every object and every line of text gets its own bounding box, so you decide where the headline, the chart or the product shot goes instead of hoping. Prompts can run to 4K tokens, enough for a handwritten sticky note or a whole whiteboard of text.
It ships as Base and SFT checkpoints under Apache 2.0, with code and weights public. Use Base if you plan to fine-tune and SFT if you just want pictures, and adapter training is reported to fit in 6GB of VRAM.
Dense text rendering: Writes English and Chinese into the image with automatic line breaks, and you can add \n in the prompt to control them.
JSON layout input: Give each object and each line of text a position to fix the composition. This is still experimental.
Several prompt styles: Natural language, tag sequences and JSON layouts, up to 4K tokens long.
Resolution and ratio: From 512×512 up to 2K at any aspect ratio, with 1 to 2 megapixels recommended.
Anime and photoreal: One model covers both illustration styles and photographic looks.
Easy to fine-tune: The Base checkpoint is released for LoRA training, and the community already has trainers, distilled and FP16 builds.
Ecosystem: Day-0 support in ComfyUI, SGLang and vLLM-Omni, plus a Gradio UI and a Python API.
What sets it apart is layout control: besides plain language, the prompt can be a JSON description where every object and every line of text gets its own bounding box, so you decide where the headline, the chart or the product shot goes instead of hoping. Prompts can run to 4K tokens, enough for a handwritten sticky note or a whole whiteboard of text.
It ships as Base and SFT checkpoints under Apache 2.0, with code and weights public. Use Base if you plan to fine-tune and SFT if you just want pictures, and adapter training is reported to fit in 6GB of VRAM.
Features
Dense text rendering: Writes English and Chinese into the image with automatic line breaks, and you can add \n in the prompt to control them.
JSON layout input: Give each object and each line of text a position to fix the composition. This is still experimental.
Several prompt styles: Natural language, tag sequences and JSON layouts, up to 4K tokens long.
Resolution and ratio: From 512×512 up to 2K at any aspect ratio, with 1 to 2 megapixels recommended.
Anime and photoreal: One model covers both illustration styles and photographic looks.
Easy to fine-tune: The Base checkpoint is released for LoRA training, and the community already has trainers, distilled and FP16 builds.
Ecosystem: Day-0 support in ComfyUI, SGLang and vLLM-Omni, plus a Gradio UI and a Python API.
