MOSS-Image Model Guide
Everything verified about OpenMOSS's open-source 5B text-to-image model, with every number linked to its source.
MOSS-Image is a 5-billion-parameter text-to-image diffusion transformer released under the Apache-2.0 license by OpenMOSS on October 16, 2026. It is built for long English and Chinese text rendering and for layout control through JSON bounding boxes, rather than for raw photorealism. This page collects what the official repository, model cards and technical report actually say, with sources, and is updated as the project changes.
Last verified: October 20, 2026. MossImage is an independent site and is not affiliated with OpenMOSS.
What is MOSS-Image?
MOSS-Image is an open-weight image generation system made of a 5B single-stream diffusion transformer, a frozen Qwen3.5-9B text encoder and the frozen FLUX.2 VAE. OpenMOSS published the code, two checkpoints (Base and SFT) and a 35-page technical report on the same day, with day-0 support in ComfyUI, SGLang and vLLM-Omni.
| Fact | Value | Source |
|---|---|---|
| Official name | MOSS-Image (checkpoints: MOSS-Image-Base, MOSS-Image-SFT) | GitHub |
| Release date | October 16, 2026 | GitHub News |
| Developer | OpenMOSS; paper authored as "MOSI.AI" with affiliations MOSI Intelligence, SII, Fudan University and Shanghai Jiao Tong University | arXiv 2610.23374 |
| License | Apache-2.0 for code and weights; the paper is CC BY 4.0 | Hugging Face |
| Parameters | 5B in the diffusion generator, excluding the text encoder and VAE | arXiv, Appendix A |
| Text encoder | Qwen3.5-9B, frozen | arXiv |
| VAE | FLUX.2 VAE (from FLUX.2-klein-4B), 32 latent channels, 8x downsampling | arXiv |
| Max prompt | 4,096 tokens | GitHub |
| Hugging Face org | OpenMOSS-Team (not "OpenMOSS", which is an unrelated account) | Hugging Face |
| Online demo | Hugging Face Space | Demo |
Who made MOSS-Image?
The GitHub organization is OpenMOSS and the Hugging Face organization is OpenMOSS-Team. The technical report lists "MOSI.AI" as the author, with MOSI Intelligence, SII, Fudan University and Shanghai Jiao Tong University as affiliations. The project leads are Zhe Xu and Donghua Yu; the advisors are Xie Chen and Xipeng Qiu, whose group at Fudan also created the MOSS language model series.
What is the architecture?
MOSS-Image is a 42-block single-stream transformer with hidden size 3072, 24 attention heads of 128 dimensions, and an FFN of 8192, trained from scratch. Text features come from the final layer of a frozen Qwen3.5-9B and pass through a linear projection and two text-only transformer blocks before joining the image tokens.
| Component | Specification |
|---|---|
| Backbone | 5B single-stream DiT, 42 blocks, hidden 3072, 24 × 128 heads, FFN 8192 |
| Feed-forward | SiTU-GLU layers (the paper cites Kimi K3) |
| Position encoding | 4D RoPE over reference index, sequence, width and height |
| Latent patch | 2 × 2, so one token covers 16 × 16 pixels; a 1024² image is 4,096 tokens |
| Optimizer | Hybrid Muon-AdamW; a 1B ablation reached the same loss with 29-33% fewer steps than AdamW |
| Training data | ReLAION, OmniCorpus, DataComp and others, about 3 billion images after deduplication; captions mostly English |
| Final checkpoint | Step 310,000, EMA weights |
Source: arXiv 2610.23374, Appendix A.
What is the difference between MOSS-Image-Base and MOSS-Image-SFT?
Base is the raw pretrained model with no post-training and no distilled inference parameters, published so the community can train LoRAs on an unaltered foundation. SFT is fine-tuned from Base on curated data with the same architecture and pipeline, and it is the checkpoint most people should run. The official example shows SFT producing cleaner text grouping and a more organized whiteboard layout.
How good is MOSS-Image at rendering text?
On LongLongTextBench (LLTB), a benchmark introduced in the technical report with 240 tasks averaging about 1,340 characters each, MOSS-Image reaches an average text coverage of 96.3. The next open-weight model in the comparison, Qwen Image 2512, scores 64.9. English long-form text scores 97.1 and Chinese long-form text 92.4.
| Benchmark | MOSS-Image | Comparison | Source |
|---|---|---|---|
| LLTB Mean-6 (long text, EN + ZH) | 96.3 | Qwen Image 2512: 64.9; Ideogram 4: 63.5 | arXiv |
| LLTB English long | 97.1 | — | arXiv |
| LLTB Chinese long | 92.4 | — | arXiv |
| LongTextBench (short text, ~120 chars) | 97.63 | Public baselines: 96.05-98.85 | arXiv |
The gap between English and Chinese exists because the training captions are predominantly English, with no dedicated Chinese caption annotation.
How does layout control work?
Instead of describing positions in prose, you give MOSS-Image a JSON prompt that lists items with labels, captions, optional text and bounding boxes in absolute pixel coordinates centered on the canvas. On the 7Bench layout benchmark this scores 0.88 mIoU and 0.70 layout AUC, against 0.44 and 0.20 for Ideogram 4 and 0.14-0.17 for GPT Image 2, Nano Banana Pro, Seedream 5.0 Pro and Qwen Image 3.0 Pro.
The format is documented on our layout JSON spec page. It is labeled experimental in the official user guide. ComfyUI exposes a MOSS-Image Layout Prompt node that turns drawn boxes into this JSON.
How does MOSS-Image compare in human evaluations?
In RealLayoutArena, a 61-task blind pairwise study of posters, infographics and UIs introduced in the paper, MOSS-Image with its prompt enhancer reached an Elo of 1698.70, second only to GPT Image 2 at 1844.35 and ahead of Seedream 5.0 Pro (1661.16), Qwen Image 3.0 Pro with prompt enhancement (1614.84) and Nano Banana Pro (1590.19). With shared enhanced prompts, the MOSS-Image backbone alone ranked first at 1807.34.
The same study measured how prompt style changes results: short prompts scored 1081.70, detailed prompts 1509.21, and JSON layout prompts 1622.87. Writing structured prompts matters more than most people expect.
The paper does not report GenEval, DPG-Bench or OneIG scores for the final 5B model, so we do not quote any.
What settings does the MOSS-Image team recommend?
Use the flow-matching Euler sampler with 50 or more steps, CFG 3 (raise toward 6 for short prompts), shift 5 and negative-branch skipping at layer 22. Output sizes must be multiples of 16; 1M to 2M total pixels is the recommended range, with 0.25M to 4M possible.
| Setting | Recommended value |
|---|---|
| Sampler | Euler (flow matching), scheduler sgm_uniform in ComfyUI |
| Steps | 50 or more |
| CFG | 3 (range 3.0-6.0) |
| Shift | 5 |
| Skip layers | uncond_skip_layers=[22]; 24 and 26 also work |
| Negative prompt | nsfw, worst quality, bad anatomy, watermark, jpeg artifact |
| Resolution | Multiples of 16; 1M-2M pixels recommended; 512² to 2K possible |
| Prompt language | English recommended; Chinese supported |
Full explanation in MOSS-Image recommended settings.
How do I run MOSS-Image?
You need an NVIDIA GPU with CUDA, Python 3.10 or later and PyTorch 2.9 or later. The easiest paths are the Hugging Face Space (no install), ComfyUI (native support merged on release day) and the Diffusers custom pipeline.
import torch
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained(
"OpenMOSS-Team/MOSS-Image-SFT",
custom_pipeline="pipeline_moss_image",
trust_remote_code=True,
torch_dtype={"default": torch.bfloat16, "vae": torch.float32},
)
Weight sizes from the Comfy-Org repack: DiT 10.2 GB in bf16 or 5.4 GB in int8; Qwen3.5-9B text encoder 9.8 GB in int8 or 7.3 GB in w4a8; FLUX.2 VAE 340 MB. The team has not published VRAM requirements or inference times. Step-by-step instructions are in our ComfyUI guide.
What are MOSS-Image's limitations?
The model only does text-to-image: it does not take reference images and does not edit existing images. The technical report lists image-conditioned generation and iterative refinement as future work. High-quality layouts in the paper rely on an external language model as a prompt enhancer, and the authors note that errors in that enhancer propagate to the image. Chinese long text scores lower than English, the layout JSON feature is marked experimental, and the Base checkpoint needs careful settings because nothing is distilled.
How fast is the community adopting it?
Four days after release, the GitHub repository had 973 stars, the SFT model card showed 414 likes and 3,336 downloads in the month, and Hugging Face listed 16 adapters, 14 fine-tunes and 18 quantized versions of SFT plus 15 adapters and 22 fine-tunes of Base. Community projects already include 4-step and 8-step distillation LoRAs, a full fine-tune called Kromoss and a LoRA trainer called KohakuMoss.
MOSS-Image vs other models
| Model | Open weights | Text-to-image | Image editing | Layout JSON | 7Bench mIoU |
|---|---|---|---|---|---|
| MOSS-Image | Yes, Apache-2.0 | Yes | No | Yes | 0.88 |
| Nano Banana Pro | No | Yes | Yes | No | 0.15 |
| GPT Image 2 | No | Yes | Yes | No | 0.14 |
| Ideogram 4 | No | Yes | Yes | No | 0.44 |
| Qwen Image 3.0 Pro | No | Yes | Yes | No | 0.17 |
7Bench values from arXiv 2610.23374, Table 14. Detailed comparison: MOSS-Image vs Nano Banana Pro.
FAQ
Is MOSS-Image free for commercial use?
Yes. The code and weights are released under Apache-2.0, which allows commercial use, modification and redistribution with attribution.
Does MOSS-Image have an official API?
No. As of October 20, 2026 there is no hosted API from OpenMOSS, and no inference provider on Hugging Face serves the model. You run it yourself or use the Hugging Face Space.
Can MOSS-Image edit images or use reference photos?
No. The backbone supports text-to-image only. Reference-image conditioning is reserved in the position encoding for future work.
Which checkpoint should I download?
MOSS-Image-SFT for generating images. MOSS-Image-Base only if you plan to train a LoRA or fine-tune.
How long can the prompt be?
Up to 4,096 tokens through the Qwen chat template, which is enough for several paragraphs of text to render.
Is MossImage the official MOSS-Image website?
No. MossImage is an independent generator and guide. The official homepage is openmoss.github.io/MOSS-Image and the official code is on GitHub under OpenMOSS.