What Is MOSS-Image? Architecture, Benchmarks and License Explained
A plain-English guide to OpenMOSS's 5B open-source text-to-image model, what it is good at, how it was trained and where its limits are.
MOSS-Image is a 5-billion-parameter diffusion transformer for text-to-image generation, released under Apache-2.0 by OpenMOSS on October 16, 2026, and built to render very long Chinese and English text and follow JSON layout instructions. It is text-to-image only. There is no image editing, no reference-image mode and no official hosted API yet.
This guide explains who built it, how it works, what the published benchmarks say and where it falls short. Every number below links to its source: the GitHub repository, the arXiv technical report or the Hugging Face model card. For a shorter overview, see the MOSS-Image model guide.
Who made MOSS-Image?
MOSS-Image was built by MOSI Intelligence together with the Shanghai Innovation Institute (SII), Fudan University and Shanghai Jiao Tong University. The technical report lists "MOSI.AI" as its author, with Zhe Xu and Donghua Yu as co-first authors and Xie Chen and Xipeng Qiu as advisors.
Xipeng Qiu leads the MOSS team at Fudan. The code lives under the OpenMOSS organization on GitHub; the weights live under OpenMOSS-Team on Hugging Face. The OpenMOSS organization on Hugging Face is an unrelated group, so check the name before downloading.
When was MOSS-Image released?
MOSS-Image was released on October 16, 2026. Code, weights and day-0 support for SGLang, vLLM-Omni and ComfyUI shipped the same day, and the 35-page technical report was submitted to arXiv as 2610.23374. The ComfyUI Wiki news post covers the launch.
Two checkpoints were published at launch: MOSS-Image-Base and MOSS-Image-SFT. Both share the same architecture and inference pipeline. A free demo runs on a Hugging Face Space.
What is the MOSS-Image architecture?
MOSS-Image is a single-stream diffusion transformer (DiT) with 5 billion parameters, trained from scratch. It uses a frozen Qwen3.5-9B language model as its text encoder and a frozen FLUX.2 VAE for the latent space. The 5B count covers only the diffusion generator, not the text encoder or VAE.
All figures below come from Appendix A of the arXiv report and the Hugging Face model card.
| Component | Specification |
|---|---|
| Generator | 5B single-stream DiT, trained from scratch, 42 layers, hidden size 3072, 24 heads × 128 dims, FFN 8192 |
| Text encoder | Qwen3.5-9B (frozen). Last-layer token states (4096-dim) are projected to 3072 through a linear layer and a two-layer text-only transformer |
| VAE | FLUX.2 VAE (frozen), loaded from FLUX.2-klein-4B. 32-channel latent, 8× downsampling |
| Positional encoding | 4D RoPE over reference index, sequence, width and height. The reference axis is reserved for future reference-image generation |
| Feed-forward | SiTU-GLU, adopted from Kimi K3 |
| Patching | 2×2 latent patches, equivalent to a 16 px stride. A 1024² image is 4096 tokens |
| Text limit | 4096 tokens |
| Optimizer | Muon-AdamW hybrid. A 1B ablation reached the same loss in 29-33% fewer steps than AdamW alone |
Because the text encoder and VAE are standard components, ComfyUI only had to add the DiT as a new model type. That is a large part of why community integration happened within days.
What is the difference between MOSS-Image-Base and MOSS-Image-SFT?
MOSS-Image-Base is the raw pre-trained model with no post-training and no inference-parameter distillation such as baked-in CFG. MOSS-Image-SFT is the same model after supervised fine-tuning on curated data. Most users should run SFT. Base exists as a clean starting point for LoRA and full fine-tuning.
The official examples describe the SFT checkpoint as producing cleaner text grouping and more organized whiteboard-style layouts. The GitHub README explains that Base intentionally keeps all training information intact so the community can train adapters on top of it. If you only want to generate images, pick SFT and follow the recommended settings.
How good is MOSS-Image at rendering long text?
Long text rendering is the model's strongest published result. On LLTB (LongLongTextBench), a benchmark introduced in the paper, MOSS-Image scores a Mean-6 of 96.3, with 97.1 on English long text and 92.4 on Chinese. The next model listed, Qwen Image 2512, scores 64.9, a 31.4-point gap.
LLTB has 240 prompts, 120 in English and 120 in Chinese, averaging 1,339.8 characters each. The report says that is about 11 times the length of the older LongTextBench. Ideogram 4 scores 63.5 on LLTB when refusals count as zero.
On the shorter LongTextBench (average 120 characters), MOSS-Image scores 97.63. Open-source baselines sit between 96 and 98.9, so the paper treats that benchmark as saturated. Line breaks are controlled with \n or a real newline in the prompt. The Chinese score is lower than English because the training captions are mostly English. Prompt examples are collected in our MOSS-Image prompt guide.
How does JSON layout control work in MOSS-Image?
MOSS-Image accepts a JSON object as its prompt. The JSON lists items with a label (text, character or object), a caption, optional text content and a bounding box. On the 7Bench layout benchmark, MOSS-Image scores 0.88 mIoU and 0.70 Layout AUC. Ideogram 4 scores 0.44 and 0.20.
Bounding boxes use the yxyx format, meaning [y_min, x_min, y_max, x_max], in abs_centralized coordinates. That is absolute pixels with the canvas center as the origin, x to the right and y downward, so a 1024×1024 canvas spans -512 to +512 on both axes. The paper argues that zero-centered absolute coordinates generalize to unseen resolutions better than coordinates normalized to a [0, 1000] range.
On 7Bench, GPT Image 2, Nano Banana Pro, Seedream 5.0 Pro and Qwen Image 3.0 Pro all land in the 0.14-0.17 range according to the report. The feature is labeled Experimental. ComfyUI core includes a MOSS-Image Layout Prompt node that converts drawn boxes into this JSON. The full field reference is in our layout JSON spec.
How does MOSS-Image rank in human preference tests?
In RealLayoutArena, a blind pairwise human-preference test from the paper, MOSS-Image ranks second end-to-end with a Bradley-Terry Elo of 1698.70, behind GPT Image 2 at 1844.35. Seedream 5.0 Pro scores 1661.16, Qwen Image 3.0 Pro with prompt enhancement 1614.84, and Nano Banana Pro 1590.19.
The arena covers 61 tasks: 20 posters, 20 infographics and 21 UI screens. The MOSS-Image entry in the end-to-end ranking uses a GPT-6 prompt enhancer. In a second setting where every model receives the same enhanced prompts, MOSS-Image ranks first at 1807.34, ahead of ERNIE Image at 1748.51 and SenseNova-U1 at 1689.23.
The paper also measures how prompt style affects Elo for MOSS-Image: short prompts 1081.70, detailed prompts 1509.21, natural-language enhanced prompts 1506.20 and JSON layout prompts 1622.87. The practical takeaway is that a layout JSON beats a well-written paragraph.
One caution: the report does not publish GenEval, DPG-Bench or OneIG scores for the final 5B model, only for small ablation models. Any general-benchmark number you see quoted for MOSS-Image should be checked against the arXiv source.
How was MOSS-Image trained?
MOSS-Image was trained on roughly 3 billion deduplicated images drawn from ReLAION, OmniCorpus, DataComp and similar sources, recaptioned with Gemma 4. Training ran on 512 devices on what the team calls a "heterogeneous GPU platform". The specific GPU model is not published.
The captions are mostly English with no dedicated Chinese annotation pass, which the report gives as the reason for the gap between English and Chinese long-text recall.
What are the known limitations of MOSS-Image?
MOSS-Image is text-to-image only. It cannot take a reference image or edit an existing image; the paper lists image-conditioned generation as future work. High-quality layouts also depend on an external LLM to expand the prompt, and errors in that LLM carry through to the image.
The other limits, drawn from Section 8 of the report and the USERGUIDE:
- Chinese long text scores lower than English (92.4 vs 97.1 on LLTB).
- The Base model needs 50 or more steps and careful CFG. There is no official distilled checkpoint, so it is not fast.
- JSON layout control is labeled Experimental.
- The team's own framing, as summarized by ComfyUI Wiki, is that the model sells text and structure rather than raw photorealism.
- Official VRAM requirements and inference speed are not published.
- An open-source layout prompt enhancer is announced as "coming soon" but not released.
How can you try MOSS-Image today?
The fastest path is the official Hugging Face Space, which needs no setup. For local use, ComfyUI merged support into core on launch day (PR #17095 by Kijai), and Comfy-Org publishes a repacked checkpoint. Diffusers loads it with a custom pipeline.
Requirements listed on GitHub: Python 3.10 or newer, PyTorch 2.9 or newer, NVIDIA CUDA, diffusers 0.38 or newer and transformers 5.3 or newer. Weight sizes: DiT bf16 10.2 GB, DiT int8 5.4 GB, Qwen3.5-9B int8 9.8 GB or w4a8 7.3 GB, FLUX.2 VAE 340 MB. The Gradio app offers three memory modes: GPU-resident, component offload and sequential offload.
SGLang and vLLM-Omni both had day-0 support; the vLLM recipe is the quickest server setup. There is no hosted API: as of October 20, 2026 no Hugging Face Inference Provider, fal.ai, Replicate, WaveSpeed or Kie.ai deployment exists. Our ComfyUI guide walks through the local workflow. If you just want to make a poster now, try the MossImage generator.
How fast is the community adopting MOSS-Image?
Four days after release, the GitHub repository had 973 stars and 45 forks, and Hugging Face already listed more than 30 fine-tunes and LoRAs across the two checkpoints. Community distillations include a 4-step "rapid-moss" LoRA and an 8-step DMD LoRA.
Snapshot from October 20, 2026, per Hugging Face:
| Checkpoint | Likes | Monthly downloads | Adapters | Fine-tunes | Other |
|---|---|---|---|---|---|
| MOSS-Image-SFT | 414 | 3,336 | 16 | 14 | 18 quantized versions, 21 Spaces |
| MOSS-Image-Base | 371 | 2,774 | 15 | 22 | — |
Other early derivatives include the Kromoss full fine-tune and the KohakuMoss LoRA trainer, all appearing between October 17 and 19.
FAQ
Is MOSS-Image free for commercial use?
Yes. Both the code and the weights are released under Apache-2.0, which permits commercial use, modification and redistribution. The technical report itself is CC BY 4.0. See the license file on GitHub.
Can MOSS-Image edit images or use a reference photo?
No. The model is text-to-image only. Reference-image and image-editing support are listed as future work in the report, and the 4D RoPE reference axis is reserved for that purpose.
How much VRAM does MOSS-Image need?
The team has not published a VRAM figure. The bf16 DiT alone is 10.2 GB, and the official Gradio app includes component-offload and sequential-offload modes for machines with less memory.
Which checkpoint should I download, Base or SFT?
Download MOSS-Image-SFT for generation. MOSS-Image-Base is meant as a fine-tuning starting point and has had no post-training.
Does MOSS-Image support Chinese text?
Yes. LLTB includes 120 Chinese prompts and MOSS-Image scores 92.4 on them, compared with 97.1 on English. English prompts are still recommended because the training captions are mostly English.
Is there an API for MOSS-Image?
Not yet. As of October 20, 2026 there is no official hosted endpoint and none of the major inference marketplaces list the model. You can self-host with the Diffusers pipeline, ComfyUI, SGLang or vLLM-Omni.