TL;DR:
MOSS Image is in the generator. A new open-weights text-to-image model from the OpenMOSS team.
Built for text and layout. Words that read right, and you can place them with a JSON layout.
Long prompts. Up to 4K tokens, in plain language, booru tags or JSON.
Two versions. SFT for generating. Base for fine-tuning and LoRAs.
Train your own. MOSS Image LoRA training is live in the Training Studio, on SFT or Base.
MOSS Image is live in the generator and the Training Studio. Two versions, both open weights, both ready to prompt or train.
Try MOSS Image in the generator
What MOSS Image does
MOSS Image is a 5B text-to-image model from OpenMOSS. It was made for the things image models usually fumble: text inside the image, where things sit on the canvas, and prompts too long for most models to follow.
Text that reads right. Put the words you want rendered in quotes, and use line breaks to group them.
Layout control. Describe the scene as JSON, with the text and objects where you want them, and paste it in as your prompt.
Long prompts. Up to 4K tokens. Write a paragraph, a tag list or a full layout.
Any aspect ratio. From 512px up to 2K.
SFT or Base
SFT is the fine-tuned one, and the one OpenMOSS recommends for generating. Cleaner text, tidier layouts. Start here.
Base is the raw model with no post-training. OpenMOSS recommends it for fine-tuning and LoRA training, and LoRAs trained on Base also work on SFT, mostly. Try MOSS Image Base.
Both give you steps, CFG and a negative prompt. Defaults are 25 steps at CFG 3. If your prompt is short, push CFG up a bit. English prompts work best, though Chinese and other languages are supported.
This release is text-to-image only. No editing or reference images yet.
Train a LoRA
You can train MOSS Image LoRAs on Civitai now, in the Training Studio. Pick SFT or Base as your starting point. OpenMOSS recommends Base for training, and LoRAs trained on Base also work on SFT.
Run it yourself
The weights are on the model page, in bf16 and int8, with the text encoder alongside. MOSS Image is released under the Apache License 2.0. The OpenMOSS repo has the reference pipeline and a user guide with layout JSON examples.
Show us what you make
Credit for MOSS Image goes to the OpenMOSS team. Post what you make with it to Civitai. Posters, signs, comics, menus, whatever has words in it. We want to see where you take it.
Open the generator. Write something long. Show us what MOSS Image can do.