Ming-Image
Ming Image 0.1 Design by inclusionAI (MIT), packaged by Comfy-Org for ComfyUI: photorealistic, uncensored out of the box, and it writes clean text.
What it does well
Photographic nudes and portraits with natural skin texture, no LoRA needed
About 20 seconds per image at 1 megapixel on an 8 GB card with the int8 file
Clean typography on posters, covers and labels, in English and Chinese
Also anime and painted illustration
Plain-caption prompts: who, what she does, where, the light
Which file
8 GB cards and up: int8 convrot, 6.2 GB, the default
bf16, 12.3 GB: full precision, for merging and training
Settings
12 steps, cfg 1, euler, simple, about 1 megapixel (832x1248 or 1024x1024), the size the model is tuned for; 2048 px also works. The text encoder (ming_image_0.1_ling_mini_2.0_w4a8, 12.8 GB, in models/text_encoders) and the VAE (ming_image_vae_bf16, in models/vae) are both on this page. Needs a ComfyUI build with native Ming Image support: the main branch from 24 September 2026, or the next release.
Notes
The negative prompt does nothing at cfg 1
Nudity renders out of the box; explicit sex acts do not yet, the base has not learned them
The current files are Comfy-Org's, future release will include fine tuning
Don't use it on real people.
中文：基于 inclusionAI 的 Ming Image 0.1 Design（MIT 许可），由 Comfy-Org 打包的原版文件：开箱即可生成写实裸体照片，文字渲染清晰。8G 显存用 int8 convrot（6.2 GB），bf16（12.3 GB）用于融合与训练。文本编码器（12.8 GB）和 VAE 也在本页。12 步，cfg 1，约 100 万像素。需要原生支持 Ming 的 ComfyUI。
License
MIT, Copyright (c) 2026 inclusionAI.
Description
Ming Image 0.1 Design for ComfyUI: Comfy-Org's int8 convrot and bf16 files,
With the VAE and a workflow at 12 steps, cfg 1.
FAQ
Comments (13)
Can you provide smaller size int8 text encoder to use with smaller int8 model?
Is this model based on Qwen 2.1?
If yes can I use Qwen 2.1 text encoder with your model?
The base model is Ming Image, noctaluna added all the required files in the donwload section.
You need to update your ComfyUI because the support for that model was added few hours ago.
I am starting to use it, but it seems LCM sampler produces cleaner images
Post with Euler sampler:
https://civitai.red/posts/31220697
Post with LCM sampler
https://civitai.red/posts/31220919
This post is LCM with 16 STEPS
https://civitai.red/posts/31228588
I did some experiment and LCM looks pretty unique in how it behave : https://civitai.red/posts/31239398
Euler produce way too much noise. With LCM there's no noise but images looks sloppy. There should be something in the middle...
If this is a 12 step sampling process, then the solution to such a poorly trained baked-in-distillation is 12 step euler, followed by the final two steps of 16 step LCM with a different seed (i.e. start at 14, end at 16)
I did some experiment and LCM looks pretty unique in how it behave : https://civitai.red/posts/31239398
@VeerGeer that's what I did, 8 to 12 LCM refine seems ok
euler works better with ddim_uniform
And there is another option, the Z-Image LORAs work with Ming Image, probably there is a good one LORA for this purpose
Are you the one who archived the model and the workflow because CivitaiOfficial reposted it for its generator?
No, Civitai did it by themselves to give some space to the official release
@noctaluna FIY I don't need it, I just wondered because it seems the site didn't archive duplicates of other models they reposted (or transferred from theally, an ex manager) on CivtaiOfficial, but it happened to your upload