A nine-node ComfyUI workflow for MOSS Image. No custom nodes, no LoRA, no ControlNet — it even runs on an 8 GB laptop GPU.
WHAT YOU NEED (3 files, all from the MOSS base model)
- models/diffusion_models/ -> MOSS base int8
- models/text_encoders/ -> Qwen3.5 9B int8
- models/vae/ -> flux2-vae
THE FOUR NUMBERS THAT MATTER
Steps 50 · CFG 3.0 · euler / simple · Shift 5.0
Sizes: 1024x1408 portrait · 1408x1024 landscape · 1024x1024 square.
PROMPT STYLE — Danbooru tags, comma separated
masterpiece, best quality, very aesthetic, safe, 1girl, solo, <hair/eyes/outfit>, <pose>, <camera>, <scene>, <light>
THREE THINGS THAT SURPRISE PEOPLE
- It shortens long garments — write "floor length, long skirt"
- It drops hoods — write "hood up, head covered"
- "tiny figure" can erase the subject — use "full body, medium distance"
SPEED ON 8 GB 3070TI
~150 s at 1024x1408, ~130 s at 1408x1024, ~70 s at 1536x512.
IF VAEDECODE FAILS with comfy_kitchen / group_norm_stats3d
Update the kernel library inside ComfyUI's Python (python_embeded\python.exe -m pip install -U comfy-kitchen, needs 0.2.40+) and restart.