☕ One coffee = one happy creator: https://buymeacoffee.com/the_frizzy1
MOSS-Image on a 12 GB card, four ways: text to image, layout with bounding boxes, editing one picture, and a new picture in the shape of your photo. Four workflow files, core ComfyUI nodes only, with the prompt enhancer from the official templates. No video on this workflow yet — the one above is how to use my workflows, if you are new to them.
What you need
✔ Custom nodes — None. Core ComfyUI only.
✔ ComfyUI version — 0.43 or newer. The layout file's box editor needs Nodes 2.0 (ComfyUI icon → Nodes 2.0).
✔ VRAM — 12 GB — 11.0 to 11.5 GB peak at 1 MP, tested, not estimated
✔ System RAM — 32 GB · ComfyUI itself peaked at 5.2 GB
✔ Tested on — RTX 3060 12 GB · 32 GB RAM · ComfyUI 0.43.1 · sage attention
✔ Licence — Workflow MIT · MOSS-Image: Apache-2.0
Models to download
Click a file name to download it. Put it in the folder shown, inside ComfyUI/models/.
All six files, 14.3 GB total. Text to image and layout only need the first three.
diffusion_models/
File: moss_image_sft_int8_convrot.safetensors — 5.41 GB
Download:
https://huggingface.co/Comfy-Org/MOSS-Image/resolve/main/diffusion_models/moss_image_sft_int8_convrot.safetensors
Put it here:
ComfyUI/models/diffusion_models/
text_encoders/
File: qwen3.5_9b_w4a8_gptq.safetensors — 7.34 GB
Download:
https://huggingface.co/Comfy-Org/MOSS-Image/resolve/main/text_encoders/qwen3.5_9b_w4a8_gptq.safetensors
Put it here:
ComfyUI/models/text_encoders/
vae/
File: flux2-vae_small.safetensors — 0.25 GB
Download:
https://huggingface.co/Comfy-Org/MOSS-Image/resolve/main/vae/flux2-vae_small.safetensors
Put it here:
ComfyUI/models/vae/
loras/
File: moss_image_edit_lora_rank_64_bf16.safetensors — 0.30 GB (edit)
Download:
https://huggingface.co/Comfy-Org/MOSS-Image/resolve/main/loras/moss_image_edit_lora_rank_64_bf16.safetensors
Put it here:
ComfyUI/models/loras/
File: moss_image_color_depth_control.safetensors — 0.31 GB (control)
Download:
https://huggingface.co/Comfy-Org/MOSS-Image/resolve/main/loras/moss_image_color_depth_control.safetensors
Put it here:
ComfyUI/models/loras/
geometry_estimation/
File: moge_3_vitl_fp16.safetensors — 0.74 GB (control)
Download:
https://huggingface.co/Comfy-Org/MoGe/resolve/main/geometry_estimation/moge_3_vitl_fp16.safetensors
Put it here:
ComfyUI/models/geometry_estimation/
The four files
..._moss-image-t2i-12gb_v1.0.0.json — text to image, a short prompt is fine, the enhancer writes the long caption · ..._moss-image-layout-12gb_v1.0.0.json — text to image with boxes, say what goes where · ..._moss-image-edit-12gb_v1.0.0.json — edit one picture with the edit LoRA · ..._moss-image-control-12gb_v1.0.0.json — new picture in the shape of your photo, MoGe reads it, no map needed.
Lighter than the official templates: the w4a8 text encoder (7.3 GB) instead of int8, and MoGe 3 ViT-L (0.74 GB) instead of ViT-G (2.5 GB) in the control file.
Speed on my RTX 3060
Wall is queue to finished. VRAM is the whole card as nvidia-smi reports it, about 0.7 GB of that is Windows.
Text to image, first run after starting ComfyUI (loads the files) — 1 MP, 1 min 16 s wall, 11.3 GB peak vram, 3.8 GB peak ram (comfyui)
Text to image, new prompt (enhancer + image) — 1 MP, 55 s wall, 11.3 GB peak vram, 3.9 GB peak ram (comfyui)
Text to image, same prompt, new seed — 1 MP, 43 s wall, 11.0 GB peak vram, 3.9 GB peak ram (comfyui)
Layout (template boxes) — 1 MP, 45 s wall, 11.5 GB peak vram, 3.9 GB peak ram (comfyui)
Edit, one image — 1 MP, 1 min 11 s wall, 11.3 GB peak vram, 4.0 GB peak ram (comfyui)
Control (MoGe mask) — 1 MP, 58 s wall, 11.3 GB peak vram, 5.2 GB peak ram (comfyui)
Everything here is 1 MP. There is no 2K row, so there is no 2K number to give.
What did not work
Layout text. The knight and the ruin went where the boxes said. The text did not: the template asks for 5 text blocks, I got one garbled title, one garbled line and the other 3 empty, on seed 1. Boxes for objects, yes. Boxes for words, no.
Edit and control looked right.
No box editor in the layout file unless you are on Nodes 2.0.
Control follows the photo too loosely? In the MoGe subgraph set Render MoGe Geometry to depth instead of mask.
11.0 to 11.5 GB peak on a 12 GB card at 1 MP. It fits, it barely fits.
Get the workflow
CivitAI only lets a version carry two files, so the download box here has the text-to-image .json on its own, and a .zip with all four workflows in it: text to image, layout boxes, edit LoRA and MoGe control. GitHub and Hugging Face below have the four as separate files.
GitHub: https://github.com/Thefrizzy1/The_frizzy1-Workflows/tree/main/workflows/moss/moss-image-12gb
Hugging Face mirror: https://huggingface.co/The-frizzy1/The_frizzy1-Workflows/tree/main/workflows/moss/moss-image-12gb
More videos on this stuff: https://www.youtube.com/playlist?list=UUGKDBMhTGqvlBZE9cF5d50Q

