MOSS Image model card: it writes your posters, typos thrown in

In short

• An open image model with 5 billion parameters, built for text in images and for layout S7S4. On our 30 professional visuals in 9:16, the SFT version gets an exact result 15 times: 10 out of the 25 cases where there was text to write, plus the 5 where it doesn't add any R5.
• In our tests, one image costs $0.0108 at 50 steps, or $0.0014 at 8 steps with the DMD2 LoRA (medians, our calculation: GPU time × hourly rate) R5.
• Apache 2.0 license: commercial use allowed, no country excluded S3S7. Limitation: the Civitai page states that image editing and reference images are not part of "this release" S7, and the model adds unrequested gibberish to a third of our images R5.

What it can do

Text in the image. That's its promise: "Generate clear, accurate text within images" S4. The lab announces an "average text coverage of 96.3%" on LongLongTextBench (a measure the abstract doesn't detail), a benchmark it created itself, and the best score among open models on that benchmark S13. Its number, its benchmark. In our tests, on 30 real-world cases (posters, ads, menus, social media posts), MOSS SFT is exact 15 times. Careful: 5 of those cases asked for no text at all. Of the 25 that did, it gets everything right 10 times. It never gets everything wrong: the other 15 cases are "partial" (one wrong word, one added word, a lost punctuation mark) R5.

A case with no text requested counts as "exact" if the model doesn't add any. Checked at full resolution, October 23, 2026. R5

Advertising and product photos. A short name on a product, it can do: "SILICE" on a bottle, "ONDE" on headphones, "Atelier Milky" on a label R6R5. Things go wrong when the image leaves room for extra text: it fills it with gibberish (letters that don't form any word) that nobody asked for. Our review finds it on 10 images out of 30, with or without DMD2 R5. A fix? I don't know of any LoRA or setting that reliably prevents it, so I haven't tested one, and I'm not going to make one up. The only safe habit: read every piece of text in the image at full size before you publish.

A failure I'm showing you: the real brand. For the shoe ad, we asked for an invented brand, "FOULÉE". Accelerated MOSS (DMD2) writes it correctly, but it also draws the swoosh of a real sports brand on the sports bra and on the sock R5. For an ad, that's the most awkward failure: your brand ends up next to someone else's. Along the way, it copies a piece of the prompt into the image, "INVEND FRAND" for "invented brand", and adds a line of gibberish at the bottom R5.

MOSS's mistakes have a style of their own. It writes "L'ADLISSION" instead of "L'ADDITION" (and "L'ADDISSION" with DMD2), and slips a "cinèma" with a grave accent in underneath. It replaces "Avant" ("Before") with "Aafore", a fake English word R6. And on the ad for an invented gadget, it adds lines in Chinese that nobody asked for, with or without DMD2 R6. Its Hugging Face page declares two languages, English and Chinese S1. Just saying.

Layout. MOSS accepts a JSON prompt that places each piece of text and each object in a box within the image S4S5. The publisher's guide labels this mode "Experimental" S5. Not tested by us yet R7.

Retouching and editing. Not announced. The Civitai page, in its section about its own online generator, states: "Image editing and reference images are not part of this release" S7. The GitHub repository talks about "controlled editing" through the layout: you change the JSON (text, descriptions, boxes) to adjust what's in the image S4. Nothing there talks about retouching an existing photo, and we haven't tested it R7.

Recommended settings

• The guide: it's written for the Base version; the SFT page points to it for size, sampling and prompts S1S5.
• Language: English is recommended, Chinese and other languages are supported S7. Our prompts were in English, with the French text to write in quotes R6.
• Text: put quotes around the text you want written, and line breaks where you want them S7S5.
• Layer skipping (one of the model's internal layers skipped during part of the computation): according to the guide, without it the image looks "very realistic, but it may also cause some details to collapse" S5. Our images ran without it R2. We'll redo them with it.
• Prompt enhancement: the publisher provides a "Prompt Enhancer", a skill for agents that turns an idea into a detailed description or a layout S4. We didn't use it: our prompts went in as they were R2.

Hardware and cost

Sizes in gigabytes (billions of bytes). Our calculation from the file sizes. S2S7S8

The GPU. None of the official pages we read gives the GPU memory needed to generate images S1S4S7. The only figure, "as little as 6GB VRAM", is about training LoRAs with the KohakuMossTrainer tool S4. We used two 32 GB cards: an RTX 5090 (shared benchmark) and an RTX PRO 4500 Blackwell (real-world cases) R3R5. We didn't measure peak memory R7.

• Time: on the benchmark, 17.5 s of compute per image at 50 steps; 1.9 s with DMD2, so 9 times faster R3. The very first image is slower, while the model loads: 53.8 s and 44.5 s R3.
• Price: the RTX 5090 was at $1.19/h, the RTX PRO 4500 at $0.72/h R3R5. The 30 real-world cases cost us $0.335 with SFT and $0.086 with DMD2, waiting time included (our calculation) R5.

License

MOSS Image is under Apache 2.0: "The model weights and bundled inference code are released under the Apache License, Version 2.0" S1. The text itself:

• Clause 2: "a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work" S3.
• Clause 4: if you redistribute, "You must give any other recipients of the Work or Derivative Works a copy of this License" and "You must cause any modified files to carry prominent notices stating that You changed the files" S3.
• Clause 6: the license does not give the right to use "the trade names, trademarks, service marks, or product names of the Licensor" S3.
• Clause 7: the model is provided "AS IS", without warranty S3.

Commercial use. The license contains no clause that forbids it S3. The Civitai page sums it up like this: "Commercial use, derivatives and LoRAs are allowed; keep the licence notice with any redistribution"; commercial use of images is ticked there S7. The Apache text talks about the "Work" (the weights and the code); it contains no clause about the images produced S3.

Territory. No country excluded: the license is "worldwide" S3.

Download

Official links only:

• Hugging Face, OpenMOSS: MOSS-Image-SFT to generate, MOSS-Image-Base to train LoRAs S7. Diffusers format, for the publisher's Python code S1. Repository not gated: no terms to accept before downloading S2.
• GitHub, OpenMOSS: OpenMOSS/MOSS-Image, the code, the user guide and the news S4S5.
• Comfy-Org (files for ComfyUI): Comfy-Org/MOSS-Image, the same weights repackaged as single files, with workflow templates S9.
• Civitai, official mirror: MOSS Image, published by the CivitaiOfficial account, files taken from Comfy-Org S7. That's where we got ours, through Civitai's official API; every SHA-256 hash we recomputed matches the announced one R1.
• You need a recent ComfyUI. MOSS support was added to ComfyUI on October 16, 2026, and "As of October 19, 2026, the v0.41.3 release does not include this support" S6. Here: version 0.30 doesn't run it, 0.42.0 does R1. Another detail from the same page: ComfyUI and the publisher's Python code don't give the same image with the same seed ("matching seeds and parameters do not guarantee identical images") S6.

Compatible LoRAs we use

DMD2 (8 steps). A speed-up LoRA: it teaches MOSS to do in 8 steps what it does in 50 S10. It was distilled from the Base version and also works on SFT S10S8.

• Origin: created by lodestones, in the Kromoss repository: "8-steps-dmd2-lora.safetensors (recommended)" S10.
• License checked: Apache 2.0. The repository's README: "The weights and the bundled inference code […] are released under the Apache License, Version 2.0"; the repository's LICENSE file is the full Apache 2.0 text S10S11.
• Our file: the slimmed-down version (0.63 GB) published on Civitai by the tsolful account (the "MOSS Image Turbo LoRAs" page, a repackaging, not the author's page), which states: "I did not train, distill, or create the original Turbo/DMD LoRA weights" and gives "License: Apache-2.0" S8R1.
• Author's settings: 8 steps, CFG 1, Euler, strength 1, only one speed-up LoRA at a time S10S8. "4 steps also works" S10. Us: 8 steps, CFG 1, strength 1 R2.
• What it changes: 9 times faster on the benchmark R3, but 13 exact texts out of 30 instead of 15, the only clear anatomy defect among our 60 MOSS images (an arm melting into the shirt), and 3 images discarded out of 30 versus 1 R5R6.

Three other speed-up LoRAs are installed here but not tested yet: akhaliq's SFT 4 steps, and lodestones' SFT 8 steps and Base 8 steps R1. Their Hugging Face pages all show "license: apache-2.0" S14S15S16.

Our test images

The shared benchmark: 6 identical prompts for every model in the catalog, the same neutral woman ("a woman in her late twenties with shoulder-length brown hair"), seed 2210, 1152 × 1536, one run R2R3.

• "BONJOUR": spelled right, both times. With a smiley without DMD2, a heart with it. Nobody asked for either R4.
• Night: the face is well lit by the phone screen, with and without DMD2 R4.
• Background text: MOSS invents shop signs, "TOS HIIOCRA" with SFT, "DOSELAS" with DMD2 R4.
• The face: same seed, same prompt, and the DMD2 portrait shows a different woman from the SFT portrait R4.
• Outfits: in the 12 benchmark images, everyone is dressed normally R4. Across the 60 MOSS real-world cases, 3 images were set aside because the outfit didn't follow our rules: never shown. A 4th had been set aside for the real logo of a sports brand; I'm showing it to you above, because that failure is the model's own doing R5.

Milky's verdict

MOSS SFT is top of the class for short texts: a brand name, a price, a label. It lights a night scene well. It's slow, but at one US cent per image, I'm not going to cry about it. As soon as a poster has three blocks of text, it fills the gaps with gibberish, and it sometimes writes a wrong title in huge letters. With DMD2, you gain a factor of 9 in speed and lose a bit of accuracy. How I'd use it: DMD2 to hunt for the idea, SFT for the final image, and always proofread every letter before publishing. Who it's for: a small business that wants a poster or a product photo with a short name, with a downloadable model and a simple license.

What we don't know yet

• What MOSS gives with layer skipping at layer 22, set the way the publisher recommends: our images ran without it.
• What the JSON layout is worth, and the publisher's prompt enhancement.
• The minimum GPU memory it really needs.
• What several seeds per prompt would give: here, a single run.
• The Base version, Kromoss and the three other speed-up LoRAs.
• Where MOSS ranks in our Arena: no duels yet.

Words to remember

• Steps: the number of "denoising" passes used to build the image. The more there are, the longer it takes.
• CFG: how closely the model sticks to the prompt. 3 by default for MOSS S1.
• LoRA: a small file added to a model to change its style, its character or its speed.
• Distillation (DMD2): teaching a model to do in 8 steps what it used to do in 50 S10.
• int8, bf16: two ways of storing the model's numbers. int8 takes about half as much space S7.
• Gibberish: made-up letters that don't form any word.

Source passages