Ming Image 0.1 Design 12GB — 6B int8 Text to Image 2K + Edit 1-3 Images + real RGBA transparency
1
23
1
☕ One coffee = one happy creator: https://buymeacoffee.com/the_frizzy1
Ming Image 0.1 Design, a 6B design model from inclusionAI / Ant Group, on a 12 GB card. Text to image up to 2K and editing with one to three images in one node, plus a switch for a real transparent background — Ming outputs RGBA and Save Image keeps the alpha. Put every word you want rendered in quotes, that is how you get the text right.
What you need
✔ Custom nodes — None. Core ComfyUI only.
✔ ComfyUI version — 0.39 or newer (Ming Image support). 0.39.2 tested.
✔ VRAM — 12 GB. No peak figure was measured — see "What did not work".
✔ System RAM — 9.8 GB peak, in a 32 GB machine
✔ Tested on — RTX 3060 12 GB · 32 GB DDR4 · ComfyUI 0.39.2 (Comfy Desktop)
✔ Licence — Workflow MIT · Ming Image 0.1 Design: MIT
Models to download
Click a file name to download it. Put it in the folder shown, inside ComfyUI/models/.
Required — 3 files, 19.3 GB
diffusion_models/
File: ming_image_0.1_design_int8_convrot.safetensors — 6.2 GB (all modes)
Download:
https://huggingface.co/Comfy-Org/Ming-Image/resolve/main/diffusion_models/ming_image_0.1_design_int8_convrot.safetensors
Put it here:
ComfyUI/models/diffusion_models/
text_encoders/
File: ming_image_0.1_ling_mini_2.0_w4a8.safetensors — 12.8 GB (all modes (32 GB RAM))
Download:
https://huggingface.co/Comfy-Org/Ming-Image/resolve/main/text_encoders/ming_image_0.1_ling_mini_2.0_w4a8.safetensors
Put it here:
ComfyUI/models/text_encoders/
vae/
File: ming_image_vae_bf16.safetensors — 0.3 GB (all modes)
Download:
https://huggingface.co/Comfy-Org/Ming-Image/resolve/main/vae/ming_image_vae_bf16.safetensors
Put it here:
ComfyUI/models/vae/
The official templates use the int8 text encoder instead, ming_image_0.1_ling_mini_2.0_int8_convrot at 19.5 GB. It works too, pick it in the Text encoder field. I use the w4a8 one so the whole thing fits 32 GB of system RAM. The template's prompt enhancer (Qwen3.8 27B, 17.3 GB more) is left out on purpose: your prompt goes in exactly as typed.
Modes
Text to image — prompt uses, 1 MP: 16 s · 2K: 1 min 11 s (median of the warm runs) rtx 3060
Edit (your images) — Image 1 (+ 2, 3 optional) uses, 1 MP: 34 s (1 image) · 56 s (2 images) rtx 3060
Transparent background puts the model card's phrase "RGBA, 4-channel, transparent background." in front of your prompt.
Comfy Kitchen attention against PyTorch attention, sampling only: 12 s against 13 s at 1 MP, 1 min 05 s against 1 min 06 s at 2K. It buys you about a second — that is not the reason to use this.
What did not work
The transparency leaks. Even normal images can come out partly transparent: my 1K poster had alpha down to 199 on 45 % of its pixels. Flatten onto a background if you need an opaque image.
"Invalid image file" before the run starts means an image input you are not using is un-bypassed. Ctrl+B it.
"Needs Image 1" means you picked Edit and Image 1 is still bypassed.
Text spelled wrong is one image per seed. Put the exact words in quotes and try another seed.
No peak VRAM was measured on this one, only system RAM at 9.8 GB. It runs on a 12 GB card; I do not have an nvidia-smi figure to quote.
The two times above are medians of the warm runs. There is no cold-run number.
Get the workflow
GitHub: https://github.com/Thefrizzy1/The_frizzy1-Workflows/tree/main/workflows/ming/image-0.1-design-12gb
Hugging Face mirror: https://huggingface.co/The-frizzy1/The_frizzy1-Workflows/tree/main/workflows/ming/image-0.1-design-12gb
More videos on this stuff: https://www.youtube.com/playlist?list=PLG_da2oqTx7k

