What is MOSS-Image?

MOSS-Image is an open-weight text-to-image model family from OpenMOSS, released on 16 October 2026 under Apache-2.0. It pairs a 5B single-stream diffusion transformer with a frozen Qwen3.5-9B text encoder and the FLUX.2 VAE. Its specialty is text and placement: it renders long Chinese and English passages inside images, reads prompts of up to 4,096 tokens, and accepts a JSON layout that assigns every object and line of text its own bounding box.

You can try it in your browser, run it in ComfyUI or call the official API.

Specifications at a glance

	MOSS-Image
Developer	OpenMOSS (MOSI.AI)
Released	16 October 2026
Generator	5B single-stream DiT, 42 blocks, hidden size 3072
Text encoder	Qwen3.5-9B, frozen
VAE	FLUX.2 VAE, frozen (32 latent channels)
Prompt length	Up to 4,096 tokens
Prompt formats	Natural language, tag sequences, JSON layout, or a mix
Output size	About 0.25–4 MP at any aspect ratio; width and height in multiples of 16; 1–2 MP recommended
Checkpoints	MOSS-Image-Base, MOSS-Image-SFT
Default sampling	Flow-matching Euler, 50 steps, CFG 3, shift 5, negative-branch layer skip at 22
License	Apache-2.0

How it works

The generator is a single-stream transformer: text tokens from the Qwen3.5-9B encoder and image latents from the FLUX.2 VAE are processed together in one sequence, with joint self-attention in each of 42 blocks. Because the text encoder is a large language model, the prompt can be long and detailed — a full poster brief, a tag list, or structured JSON.

According to the technical report, the model was pretrained from scratch on about 3 billion images from public datasets (including ReLAION, OmniCorpus and DataComp) plus text-rich and poster data. Apart from deduplication and removing AI-generated images, the team did not filter the data further; watermarks, blur and compression artifacts are described in the captions instead. Captions were rewritten into short, medium and long versions, with layout JSON giving explicit boxes. Training used a hybrid Muon–AdamW optimizer.

Base vs SFT: which checkpoint to use

Both checkpoints share the same architecture and settings, so you can switch between them freely.

• MOSS-Image-SFT is fine-tuned from Base on curated data. In the official whiteboard comparison it groups text more clearly and organizes the layout better. Use SFT for everyday generation; ComfyUI’s default templates load it too.
• MOSS-Image-Base has no post-training and no CFG distillation. It keeps more variety, reads booru-style tag prompts well, and is the recommended starting point for LoRA training. Short prompts on Base benefit from a higher CFG (3–6).

Base output from the official comparison. Open the SFT version to compare text grouping.

Benchmarks

All numbers below come from OpenMOSS’s own technical report. They have not been independently reproduced.

Long text rendering — LongLongTextBench

LLTB asks models to render literary passages of short, medium and long length in English and Chinese, then measures how much of the text appears correctly (Mean-6 coverage across the six bins).

Model	Mean-6 text coverage
MOSS-Image	96.3
Qwen-Image-2512	64.9

MOSS-Image is the best open-weight model on this benchmark, 31.4 points ahead of Qwen-Image-2512.

Layout accuracy — 7Bench

7Bench measures how well generated objects land inside the boxes they were assigned (224 images, 576 target objects).

Model	mIoU	Layout AUC
MOSS-Image	0.679	0.703
Ideogram 4.0	0.243	0.205
Qwen Image 3.0 Pro	0.169	0.176
Seedream 5.0 Pro	0.167	0.179
Nano Banana Pro	0.149	0.163
GPT Image 2	0.143	0.155

Every model received the same target boxes. GPT Image 2, Nano Banana Pro, Seedream 5.0 Pro and Qwen Image 3.0 Pro were given the original JSON prompts; Ideogram 4 ran from its public FP8 checkpoint with the boxes converted to its own JSON schema. mIoU counts a missed object as zero. Ideogram’s own report lists about 0.69 mIoU on 7Bench, which OpenMOSS attributes to differences in model variant, prompt format and scoring.

Human preference — RealLayoutArena

In blind pairwise votes on realistic, layout-heavy requests, the full MOSS-Image system scored 1698.70 Elo, second only to GPT Image 2 and ahead of Seedream 5.0 Pro. When every generator received the same enhanced prompt, the MOSS-Image backbone ranked first of nine at 1807.34 Elo, ahead of ERNIE Image (1748.51). The same study found JSON layout prompts scored highest of the prompt strategies tested (1622.87 vs 1506.20 for natural language).

What it is good at

• Text inside images: posters, menus, signs, labels, whiteboards and infographics with Chinese, English or mixed text.
• Long, detailed prompts: up to 4,096 tokens, enough for a complete design brief.
• Placement: JSON layouts with a bounding box per element. Our layout editor writes them for you.
• Styles beyond text: OpenMOSS highlights anime and photorealistic output; see the official night-street and illustration samples on our home page.
• Efficiency and fine-tuning: a 5B generator is light enough for consumer GPUs with quantized weights, and Base is designed for LoRA training.

Known limitations

• Text-to-image only out of the box. The official models take no image input. Editing currently needs a community LoRA on ComfyUI’s development branch.
• Layout is experimental. Boxes are strong guidance, not hard walls; in our tests text sometimes ran taller than its box. See real examples.
• Long text still needs proofreading. Coverage of 96% means some passages contain errors; rare characters and very small text are hardest.
• English prompts are safest. The report says training captions are “predominantly in English, without dedicated Chinese-caption annotation”. Chinese prompts work, and Chinese text renders well, but results may vary more.
• SFT can show vertical color banding on some images. A community DeBand LoRA reduces it.
• 50 steps is the quality default. That makes it slower than distilled models; a community 8-step DMD LoRA trades some quality for speed.

Ways to use MOSS-Image

Route	Best for	Cost
Online generator	Trying it now, any ratio, up to 2 MP	3 free images, then credits
Hugging Face Space	Free use with your HF quota	Free; limited by your daily ZeroGPU time
ComfyUI	Local workflows, layouts, LoRAs	Free on your GPU
Diffusers / Python	Scripts and custom pipelines	Free on your GPU
Official MOSI API	Apps and automation	Free for a limited time during beta; API key required

Timeline

• 16 Oct 2026 — MOSS-Image-Base and MOSS-Image-SFT released with code and weights; day-0 support in ComfyUI, SGLang and vLLM-Omni.
• 18 Oct 2026 — KohakuMossTrainer released for LoRA and full fine-tuning.
• 19 Oct 2026 — Technical report published on arXiv (2610.23374).
• 20 Oct 2026 — MOSS-Image available on the official MOSI API platform. ComfyUI adds a community color and depth control adapter.
• 22–23 Oct 2026 — ComfyUI’s development branch adds support for MOSS-Image edit LoRAs (not yet in a stable release).