How Do You Run Moss Image Locally in ComfyUI or SGLang?

How do you run Moss Image locally?

Download the SFT checkpoint from Hugging Face and run it in ComfyUI, which supports the model natively. SGLang and vLLM-Omni are the other supported runtimes, and the reviewer ran his demos in them. Start at the defaults of 50 steps and CFG 3, and aim for 1 to 2 megapixels. If generation feels slow, community 8-step and 4-step LoRAs and quantized builds are already available.

What is Moss Image?

Moss Image is the first image model from Open Moss, the team behind Moss TTS and the Mova video model. It is a 5 billion parameter diffusion transformer. The prompt is read by a 9 billion parameter Qwen-family language model acting as the text encoder, and prompts can run to around 4,000 tokens. Older image models used much smaller text encoders, which is part of why they struggled with long, complicated prompts.

The license is the other draw. Both weights and code are Apache 2.0, so you can use it commercially and build it into products. Several other strong open-weight image models ship under non-commercial research licenses or gated community licenses. FLUX 3 Klein, for example, comes with an under-$5M commercial license. The reviewer also noted the model can produce both safe-for-work and NSFW content.

Which checkpoint should you pick?

Moss ships two checkpoints. They have different jobs.

If you only want pictures, grab SFT. If you plan to train a style or a person, you want the base model. Community work so far, including the full fine-tune mentioned below, builds on base.

Which runtimes support Moss Image?

Three are confirmed:

1. ComfyUI. Native support, and probably the route most people will take. Standard inpainting also works here.
2. SGLang. Useful if you want to serve the model or script generation.
3. vLLM-Omni. The reviewer used this one to run his demos.

The sources don't give exact install commands, node names or VRAM requirements, so check the model card and each runtime's docs for current instructions. The 5B size and the community quantized builds suggest it is within reach of consumer cards, but verify against the model card rather than trusting a guess.

What settings should you use?

The reference settings are simple.

• Steps: 50
• CFG (classifier-free guidance): 3
• Checkpoint: SFT
• Resolution: anything from about 0.25 megapixels up to 4 megapixels, with 1 to 2 megapixels as the sweet spot

Aspect ratio is flexible. Wide banners and portrait phone wallpapers work, so you aren't locked into squares. Going toward 4 megapixels costs more time, and the quality gain over the 1 to 2 megapixel range isn't documented in the sources. Start in the middle and move up only if you need the pixels.

Why is it slow, and how do you speed it up?

At stock settings, 50 steps is a lot of denoising passes. The reviewer put generation at around 30 seconds per image on his setup, and he expected older, smaller GPUs to take noticeably longer. Your number will depend on your hardware and resolution.

Within about a week of release, the community had already produced:

• 8-step and 4-step LoRAs that cut the step count sharply.
• Quantized versions that you can run in ComfyUI.
• A full fine-tune of the base model.

In his demo, applying the 8-step LoRA brought generation time well under the 30-second baseline. He also toggled a "turbo" mode in his own gallery app, which made edits visibly quicker. Separately, the Moss team has said it is working on a few-step distillation, so an official turbo checkpoint may follow.

Practical advice: add a step-reduction LoRA to your workflow, lower the step count to match what that LoRA's page recommends, and compare against a 50-step render of the same seed. Distilled LoRAs typically trade some fine detail for speed, so judge the result on your own prompts. For final renders, you can always go back to 50 steps.

How do you control text placement with JSON layouts?

Text is the main reason to use this model. The lab's own long-text benchmark puts it well above other open models and only a few percentage points behind GPT Image 2, and just behind Nano Banana. Treat that carefully. It is the lab's own benchmark, and the reviewer found no third-party evals yet. He also said it can break down when you pack too many words onto a page, and that larger models may still beat it.

To improve reliability, use the layout feature. Alongside a normal prompt, you give a JSON description of the scene, then list elements (text or objects), each with a bounding box. The model was trained to respect this format, so it isn't a bolt-on trick. In the demo, a plain generation drifted and got parts of a poster's text wrong, while the boxed version stayed accurate.

Layout control also makes iteration cheap:

• Change a couch from red to green by editing one element's description.
• Change "Player One" to "Ready Player One" on a poster and regenerate.
• Swap a cat for a tiger.

Results aren't perfectly isolated. Other details can shift, such as how a hoodie sits. If you need something specific, say so in the element descriptions. You can also have an LLM write the JSON from a plain-language brief, and Open Moss provides its own layout prompt enhancer for this.

How do you add a person or style with a LoRA?

Because Moss Image has no editing mode, you can't hand it a reference photo. Whatever it doesn't know has to be trained into the weights. The reviewer showed this with a public figure the model didn't recognize. Well-known actors came out fine, but this person didn't. A custom LoRA trained on a modest, imperfect set of scraped images fixed it. The outputs skewed younger because many source photos were older, which is a reminder that your dataset shapes the result.

For your own training, use the base checkpoint. At the time of the review, not all official training code examples were out, so expect to lean on community scripts for now. A LoRA is also the practical route for a house style or a product look.

Is Moss Image worth running locally?

It depends on what you need.

Good fit:

• Posters, infographics, menus and other text-heavy images, in English or Chinese
• Anything where you need exact placement via layout boxes
• Commercial projects that need a permissive license
• Portraits, product shots and stylized work such as watercolor or anime, at default settings

Poor fit:

• Editing existing photos. It does text-to-image only.
• Dense pages with many words, where text can still fall apart
• Fast iteration on slow hardware, unless you use a step-reduction LoRA