Back to home

MOSS-Image, explained: Base vs SFT

Use SFT for a finished image, Base if you plan to train. Download links, recommended settings and what it still cannot do are below, each checked against the official repository, model cards and technical report.

MOSS-Image is an open-weight text-to-image model built for text-rich images: posters, menus, signs and other layouts where the words have to be right and in the right place. It was released on 16 October 2026 by OpenMOSS, the model team behind the MOSS series, in two variants — Base and SFT — with weights, code and a technical report.

This page collects what the official sources say, checked on 9 November 2026. Every number below is linked to its source at the bottom. This studio is an independent service and is not affiliated with the OpenMOSS team.

Who released it, and under what licence?

The MOSS-Image repository and both model cards list the Apache-2.0 licence, which permits commercial use with attribution and no per-image fee. The technical report is on arXiv (2610.23374) and the weights are on Hugging Face and ModelScope. OpenMOSS also runs a hosted API through its MOSI platform, and this studio hosts the same weights behind its own interface.

How is it built?

MOSS-Image is a 5B single-stream diffusion transformer, pretrained from scratch rather than fine-tuned from another image model. The official repository describes a 42-layer transformer paired with a Qwen3.5-9B text encoder and a FLUX.2 VAE, which is why prompts can run up to 4K tokens.

On top of the backbone, the paper introduces an intention-aware prompt-enhancement module that turns a short request into a detailed caption or a layout specification before generation. The public weights include the backbone; the enhancement step is part of the hosted system.

MOSS-Image-Base vs MOSS-Image-SFT: which should I use?

Want a finished image? Use SFT. Plan to train? Use Base.

SFT is the variant most people want. It is fine-tuned from Base on curated data, and the official documentation uses it first because it groups text more cleanly and lays out scenes better. Base has no post-training or CFG distillation, which makes it the starting point for LoRA and full fine-tuning.

Base vs SFT at a glance

MOSS-Image-BaseMOSS-Image-SFT
What it isThe pretrained checkpoint, no post-trainingA supervised fine-tune of Base on curated data
Text fidelityShort quoted lines render cleanly; long or unusual lines are where errors showSame short-line accuracy, with clearer text grouping
CreativityWider, less directed compositionsMore organized layouts; the model card's own example is a tidier whiteboard
StepsYours to set. The user guide suggests 50 or more, CFG 3.0 (range 3.0–6.0), shift 5, flow-matching EulerThe card lists the same defaults: 50 steps, CFG 3, shift 5, Euler
GPU memorySame weights class. One 1024×1024 image at 50 steps peaked at 3.65 GiB of GPU memory in low-memory mode; no source states a minimumSame figure. Switching checkpoints does not change the memory budget
Use it forLoRA and full fine-tuning, where post-training would get in the wayPosters, menus, signs and anything where the words have to read

Who should use which. If you want a finished image, use SFT. If you plan to train on the weights, use Base — the user guide explains its lack of post-training as a deliberate choice so the training signal stays intact. Both sit in the same selector in this studio at the same cost, and our own three-prompt test found no consistent winner on short quoted text, including mixed Chinese and English. The differences showed up in long or unusual lines, and two runs of the same prompt did not always agree, so generate two or three candidates and proofread each one.

Switching between them is a single loader change in ComfyUI, or one click in the model selector here.

How well does it render long and Chinese text?

The paper reports the headline numbers. On LongLongTextBench (LLTB), a benchmark introduced with the model for long-form English and Chinese text, MOSS-Image reports state-of-the-art results among open-weight models with an average text coverage of 96.3%. On RealLayoutArena (RLA), which measures end-to-end layout-rich generation, it ranks second in the system comparison and its backbone ranks first among generators evaluated with shared enhanced prompts.

"Coverage" is not the same as "correct". Our own tests on generated posters found that short quoted lines came out cleanly, while long addresses, small print and unusual punctuation were where errors appeared. The exact-text guide covers what we check and how to phrase a prompt for it.

How does JSON layout control work?

A layout is a JSON object with a summary (canvas size, overall caption) and a list of items. Each item has a label (text, object or character), a caption, and a bounding box; text items also carry the words to render. Coordinates are centred on the canvas and written as yxyx, so on a 1024×1024 canvas every value sits between −512 and +512.

Paste the JSON into the prompt field and set the output size to match the canvas you designed. The JSON layout guide walks through a full example, and the layout builder writes the JSON for you from boxes you draw.

Where do I download the two checkpoints?

Both checkpoints are public. Download the one that matches the choice above; the file names on each page say which is which.

ComfyUI's own templates load a different set of files, an int8 checkpoint plus a text encoder and a VAE, from Comfy-Org/MOSS-Image. The ComfyUI guide lists which file each of the four workflows needs.

How can I run it?

There are four documented routes. The weights run natively in ComfyUI, which had day-0 support; they also run through a Diffusers pipeline, a bundled Gradio app, and inference servers including SGLang and vLLM-Omni.

Start here: MOSS-Image-SFT, about 1–2 million pixels (1024×1024 is the published example), 50 steps, CFG 3, shift 5, flow-matching Euler. Base uses the same starting point, with CFG anywhere from 3.0 to 6.0. These are the repository's suggested defaults, not a setting that is best for every prompt.

The same page documents skip-layer guidance (SLG) on the negative branch at layer 22 — layers 24 and 26 also help, and turning SLG off looks more realistic but can collapse detail.

The memory numbers are published too: one 1024×1024 image at 50 steps peaked at 3.65 GiB of GPU memory and 14 GiB of system RAM in low-memory mode. The ComfyUI guide lists the four official workflows, the exact files they load and the settings each one uses.

Requirements: resolution, dimensions and memory

Output should total 1M, 1.5M or 2M pixels; the model extrapolates from 0.25M (512×512) up to 4M (roughly 2K). Any aspect ratio works, but width and height must be positive multiples of 16 — the Python API and manual dimension entry reject anything else, while the Gradio app rounds for you. Prompts can run to about 4K tokens. On memory, no primary source states a minimum VRAM figure; what the repository does publish is the 3.65 GiB peak above, and its advice for a tight card is to lower the resolution or use sequential or CPU offload.

If you would rather not run anything locally, this studio runs the hosted weights behind a form: type a brief, get a poster, and skip the setup.

What can't it do yet?

The official material does not document image editing or image-to-image workflows, so treat it as a text-to-image model rather than an editor. There is no documented way to change one line of an existing image while keeping the rest.

A wrong word has two fixes, and neither edits the pixels. If most of the text is wrong, regenerate: put the exact words in quotes, say where each line goes, and ask for nothing else. If the picture is right and one word is wrong, cover that word in a design tool and type it again. The exact-text guide shows the prompt shape, and three real failures show what each kind of mistake looked like.

A blurry result is usually the output size, not a broken render. Stay inside the 1–2 million pixel range above instead of generating small and enlarging, and keep the lettering large. The blur test separates a deliberately soft background from lettering that went soft.

Text quality is also not a guarantee. Coverage numbers are measured against benchmark prompts, and our own runs produced two different results for the same prompt on different runs — generate two or three candidates for anything that matters and proofread each one. Licence terms are Apache-2.0, but the model still shares the biases of its training data.

What else is worth reading?

  • Running MOSS-Image for free — the weights are free to self-host, which is not the same thing as a free hosted API; what each option actually costs.
  • Where AI posters get the text wrong — three real runs that each went wrong in one specific way, and the prompt change that fixed it.
  • A poster blur test — how to tell a blurred background from blurred lettering, with the wording held identical between the two versions.

Sources

Last checked: 9 November 2026.

Frequently asked questions about MOSS-Image

Is MOSS-Image free to use commercially?

Yes. Both model cards and the repository list the Apache-2.0 licence, which allows commercial use with attribution and no per-image fee. A hosted API such as MOSI, or this studio, sets its own prices.

Should I run Base or SFT?

Use SFT for generation: it is fine-tuned from Base on curated data and groups text more cleanly. Use Base if you plan to train a LoRA or fine-tune, because it has no post-training or CFG distillation.

Can MOSS-Image edit an existing image?

The official material documents text-to-image generation and JSON layout control, not editing or image-to-image. To change a line of text, regenerate the image from a tighter prompt.

How big is MOSS-Image, and where can it run?

It is a 5B single-stream diffusion transformer with a Qwen3.5-9B text encoder and a FLUX.2 VAE. It runs in ComfyUI, Diffusers, the bundled Gradio app, SGLang and vLLM-Omni; a 1024×1024 image at 50 steps peaks around 3.65 GiB of GPU memory in low-memory mode.

Make something worth looking at

What do you want to say?

Start with a headline. Give it a little space. See where it takes you.

Open the studio