MOSS-Image settings: steps, CFG, shift, SLG and resolution

The official defaults, what each one does, and image sizes the model accepts. From the user guide and README.

Setting	Official	ComfyUI template	Notes
Steps	50	25	Best results at 50 or more. 30 is a usable middle ground
CFG	3.0	2.5	Range 3–6. 3 is enough for long prompts; raise it for short ones
Shift	5.0	5.0	Model default
Sampler	Euler	er_sde, simple	The official plugin uses Euler / sgm_uniform
Skip-layer guidance	Skip layer 22 on the negative branch	Layer 22, 0 to 0.5	Layers 22, 24 and 26 all help. Off looks more photographic but details can collapse
Negative prompt	Empty in the Python API	Empty	Suggested: nsfw, worst quality, bad anatomy, watermark, jpeg artifact
Prompt length	Up to 4,096 tokens		Long, detailed captions are what it was trained on

Skip-layer guidance, briefly

There are two variants. The default, “two-pass” SLG, skips one transformer layer (22, counting from 0) when computing the negative branch of CFG. Standard SLG instead runs at scale 3 between 1% and 100% of sampling. In ComfyUI the native SkipLayerGuidanceDiTSimple node does the first; to do the second, swap in SkipLayerGuidanceDiT with scale 3, start 0.01, end 1 and rescaling 0. Kijai notes SLG can overblow anime styles; try it off or try layer 24 or 26 there. Some official examples use layer 26.

Prompt formats

• Natural language, English or Chinese. English has the most training data.
• Tags, in this order: quality → score → meta → year → character → series → artist → general. Example: best quality, highres, hatsune miku, vocaloid, @foo bar, 1girl, blue hair. The Gradio app prefills best quality, masterpiece.
• Layout JSON, with boxes for text and objects. Build one in the layout builder.
• Mixed: natural language and structured input can be combined.

For text in the image, put it in quotes and break lines yourself with real newlines or \n; the model wraps text on its own, but not always where you want. See the chalkboard example.

Resolution

Aim for 1, 1.5 or 2 megapixels. The model stretches from 0.25 MP (512×512) to about 4 MP. Any aspect ratio works, but width and height must both be multiples of 16, or the pipeline raises an error.

Rounded to the nearest multiple of 16 on each side.

Memory modes in the Gradio app

GPU resident keeps everything on the card; Component moves parts between CPU and GPU as needed; Sequential uses the least memory and is the slowest. Out of memory: choose Sequential or lower the size. Compile DiT is off by default and unavailable in Sequential.