Open 5B model · no GPU needed
MOSS Image AI Generator Online
Describe the image and the exact words it should carry. MOSS Image prints them spelled right, in English or Chinese, on posters, menus, covers and infographics, in 1K or 2K.
Made with MOSS Image here
- 6 free credits when you sign up · no card needed
- English and Chinese text
- No watermark, commercial use
The model
What is MOSS Image?
MOSS Image is an open text-to-image model from the OpenMOSS team (Fudan University, Shanghai Innovation Institute and MOSI), released on October 16, 2026. It is a 5B model in two versions, Base and SFT, under the Apache 2.0 license, so you can build on it and use it commercially.
Its strength is text and layout. MOSS Image prints headlines, signs and labels in English and Chinese with every character right, takes prompts up to 4K tokens, and reads a JSON layout that puts each headline and object in its own box. In the technical report it places objects with 0.88 mean IoU on 7Bench, ahead of every model tested, and ranks second only to GPT Image 2 in a blind design arena, above Seedream 5.0 Pro, Qwen Image 3.0 Pro and Nano Banana Pro.
- 1.
Layout control
Boxes for titles and objects, in pixels from the canvas center.
- 2.
Bilingual text
Chinese and English in one image, with the line breaks you set.
- 3.
Open weights
Apache 2.0, Base and SFT, 27 GB of full weights each.
- 4.
Photo and anime
Realistic photos and anime art from the same model.
Examples
MOSS Image examples, made on this site
Every image here came out of the MOSS Image generator above on its first run. Load any prompt with one tap, then change the words.
Hand lettering, prices lined up A chalkboard menu on an easel outside a small coffee shop on a sunny cobblestone street, potted plants beside it. Hand-lettered white and yellow chalk text, neatly arranged: "TODAY'S BREWS Oat Latte $4.50 Cold Brew $3.80 Honey Cortado $4.20 Ask about our seasonal pie!" A small chalk drawing of a steaming cup in the corner. Shallow depth of field, warm morning light.
Chinese calligraphy and small print, in 2K 一张极简的中式茶饮海报。米白色宣纸背景，左上方竖排的黑色书法大字"山间一盏茶"，右下角一枚红色方形印章，印章里写着"秋"。画面中央是一只青瓷茶杯，杯口飘着淡淡热气，旁边散落几朵金黄的桂花。底部居中一行细小的宋体字："十月新品 · 桂花乌龙 · 每杯 18 元"。大面积留白，雅致安静。
An infographic with four labeled steps A clean flat-vector infographic poster titled "HOW MOSS GROWS" in bold dark green letters at the top. Below the title, four numbered panels arranged in a vertical zigzag, each with a simple illustration and a short label: "1 Spores land on damp stone" "2 Green threads spread out" "3 Tiny shoots rise up" "4 Capsules release new spores" Thin dotted arrows connect the panels. Cream background, palette of moss green, olive and soft brown. At the bottom, a small footnote: "Most mosses need shade and steady moisture."
A realistic 35mm portrait A candid 35mm film photograph of an elderly fisherman mending a blue net on a weathered wooden pier at dawn. Soft golden side light, mist over the harbor behind him, deep wrinkles and calloused hands in sharp focus, shallow depth of field, natural skin texture, subtle film grain.
Anime from a tag list best quality, highres, anime key visual, 1girl, short silver hair, yellow raincoat, riding bicycle, rainy city street at night, neon shop signs, puddle reflections, motion blur, dynamic low angle, cel shading, vivid colors
A product label you can read Studio product photo of a matte kraft-paper coffee bag standing on a slab of grey stone, a small sprig of moss beside it. The front label is printed in black ink: a small leaf logo, the large words "MOSSY PEAK ROAST", and below them "Single Origin · Ethiopia · 250 g". Soft window light from the left, gentle shadow, clean beige backdrop.
Layout editor
MOSS Image layout editor: put every word where you want it
Drag a box for the headline, another for the product, and say what goes in each: switch the generator above to Layout. It writes the JSON layout MOSS Image was trained on, so you never type a coordinate.
The image MOSS Image made from this exact layout, with its boxes on top.
Whole image
A cheerful flat-illustration poster for a weekend farmers market, warm cream background, moss green and tomato red accents, bold headline at the top.
- WEEKEND MARKETHuge bold dark-green headline, centered at the top.
- Sat & Sun · 9 AM to 3 PMMedium tomato-red subtitle under the headline.
- A wicker basket overflowing with tomatoes, carrots and leafy greens.
- A striped canvas market stall with hanging string lights.
- Old Mill SquareSmall dark-green footer text, centered at the bottom.
Layout JSON925 chars
{
"summary": {
"meta": [
"poster"
],
"canvas": "single",
"caption": "A cheerful flat-illustration poster for a weekend farmers market, warm cream background, moss green and tomato red accents, bold headline at the top."
},
"items": [
{
"label": "text",
"caption": "Huge bold dark-green headline, centered at the top.",
"text": "WEEKEND MARKET",
"bbox": [
-390,
-520,
-250,
520
]
},
{
"label": "text",
"caption": "Medium tomato-red subtitle under the headline.",
"text": "Sat & Sun · 9 AM to 3 PM",
"bbox": [
-225,
-360,
-170,
360
]
},
{
"label": "object",
"caption": "A wicker basket overflowing with tomatoes, carrots and leafy greens.",
"bbox": [
-110,
-540,
370,
-29
]
},
{
"label": "object",
"caption": "A striped canvas market stall with hanging string lights.",
"bbox": [
-140,
30,
370,
540
]
},
{
"label": "text",
"caption": "Small dark-green footer text, centered at the bottom.",
"text": "Old Mill Square",
"bbox": [
385,
-200,
425,
200
]
}
],
"bbox_coord_space": "abs_centralized",
"bbox_format": "yxyx"
}Prompting
Three ways to write a MOSS Image prompt
In the technical report, short requests scored lowest with human judges, detailed and expanded prompts scored far higher, and JSON layouts scored highest. Pick the style that fits the job.
1 · One sentence
Say it in a line, keep the words in quotes
This card came from eleven words: "A birthday card that says "Happy 30th, Lena!" with watercolor balloons". MOSS Image picks the lettering, colors and composition and prints the quoted words exactly.
- Put every word you want printed in quotes
- Best for cards, quick posters and social images
- Add a style word or two when you want a specific look
2 · A long description
Describe each block of text and where it sits
Name the masthead, every cover line, the price box and the art, with color and position for each. MOSS Image follows prompts up to 4K tokens, so a full brief fits.
- Give each text block its own sentence
- Use real line breaks where the text should wrap
- Lots of small print? Keep blocks short or use 2K
3 · A JSON layout
Draw boxes, get the exact composition
A layout lists the scene in summary.caption and every element in items, each with a label, a description, the text to print and a bbox in pixels from the canvas center. Build one in the layout editor above, or paste your own.
- bbox is [y_min, x_min, y_max, x_max]
- Coordinates match the 1K canvas; 2K scales them for you
- Each text box prints exactly the words in its text field
Versions
MOSS Image Base vs SFT
Both versions share the same 5B architecture. SFT is fine-tuned to give finished images from a prompt; Base is the raw checkpoint for training and research.
| Feature | MOSS Image SFTFor finished images | MOSS Image Base |
|---|---|---|
| Made for | Finished images straight from a prompt | LoRA training and research |
| Text and layout | Cleaner text grouping, tidier layouts | Same abilities, rawer output |
| Settings | 50 steps, CFG 3, no negative prompt | 50+ steps, CFG 3 to 6, negative tags help |
| Full weights | 27 GB (10 GB model, 17 GB text encoder) | 27 GB, same layout |
| Pick it when | You want posters, covers and photos now | You plan to train a LoRA or fine-tune |
Run locally
Run MOSS Image locally or online
The weights are open, so you can run MOSS Image on your own GPU. Here is what that takes, next to what this site does for you.
MOSS Image ComfyUI guide- Online here: no GPU and no download, results in your account
- Locally: free per image, needs a CUDA GPU and the disk space
- The official Hugging Face demo is free, with a queue and a daily cap
Setup details: MOSS-Image on GitHub and the ComfyUI guide.
# Python with Diffusers, NVIDIA GPU with CUDA$ pip install "diffusers>=0.38" "transformers>=5.3" accelerate safetensors# Full SFT weights: 27 GB$ hf download OpenMOSS-Team/MOSS-Image-SFT# ComfyUI: search "MOSS-Image" in ComfyUI Manager# Smaller builds: int8 model 5.4 GB, GGUF Q4_K_M 3.7 GBPricing
MOSS Image pricing: pay per image
A 1K image is two credits and a 2K image is six. Credit packs never expire, and every new account starts with 6 free credits.
Pay once, no subscription
Or subscribe and save
FAQ
MOSS Image questions
The model is free and open under Apache 2.0, so running it on your own GPU costs nothing per image. When you use MOSS Image online here, every new account gets 6 free credits: three 1K images or one 2K image, with no watermark. After that a 1K image costs $0.25 on the Starter pack and $0.17 on the Pro Pack.
Yes. The model is released under Apache 2.0, which allows commercial use. Images you make here with credits carry no watermark and are yours to use in ads, products and client work under our terms.
No. mossimage.co is an independent website, not affiliated with OpenMOSS or MOSI. We run the open MOSS-Image model on cloud GPUs and add a layout editor, 1K and 2K output and a place to keep your images. The official project is at openmoss.ai/MOSS-Image.
Not yet. MOSS Image is text-to-image: it makes new images from a prompt or a layout and does not take a photo as input. The OpenMOSS team lists reference-based generation as future work. Until then, describe the image you want, or use the layout editor to move and rewrite parts of a design.
Yes. It renders Chinese and English, and both languages in one image. Headlines, signs, labels and short lines come out with every character right. Dense blocks of small Chinese text are its weak spot at 1K: give each block fewer words, or switch to 2K, where small characters get four times the pixels. Put the words in quotes, and describe the scene in English, which has the most training data.
All three are open models. MOSS Image leads on layout: it reads a box for every text block and object and placed them with 0.88 mean IoU on 7Bench, the best score in its report. In the same report it ranked above Qwen Image 3.0 Pro in a blind design arena. It is a 5B model, smaller than the 20B Qwen Image and close to the 6B Z-Image.
Five aspect ratios: 1:1, 3:4, 4:3, 9:16 and 16:9. 1K is one megapixel (880 x 1168 at 3:4) for 2 credits and is ready in under 30 seconds; 2K doubles each side (1760 x 2352) for 6 credits and takes 60 to 90 seconds. Images download as JPG and stay in your account for 30 days.
A failed run shows a message and returns its credits to your account right away. Prompts that ask for sexual content are refused before anything is charged, so reword the prompt and try again.
Make your first MOSS Image poster
Type the words, pick a canvas, and get a 1K image back in under 30 seconds.