← sources

One .char model, Consistent face, body & cloths: Minimax H3

source: reddit (reddit:StableDiffusion)published 2026-09-27 14:53ingested 2026-09-28 02:39

retained text

Hey Guys,

Based on my last [post](https://www.reddit.com/r/StableDiffusion/comments/1wp60lr/comment/pbz6tzr/?context=3) on stills, I thought to experiment the same .char model with Minimax H3.   
I created a character model sia.char which encodes references, prompts around face, body & cloths. 

Checkout previous post for full detail. 

[](https://preview.redd.it/virtual-tryon-consistent-face-body-cloths-via-portable-char-v0-jn2nimb3thrh1.png?width=2744&format=png&auto=webp&s=6ea66897058d461451e438d7b81ff15255e289cb)

**How it works?**

\- Build .Char: You drop max 9 reference reference, I prefer to use a ratio 2:2:1(face:cloths:body). YuNet finds the face, SFace takes a per-reference face signature, DINOv2 takes a subject signature, and the references get cleaned and normalised. All of that packs into a single portable file, a .char.  
*Only face/ref is required body & cloths link is optional.*

\- Generation: At generation, the file(.char) feeds its references into Flux's own native multi-reference channel and prepends a locked description to the prompt.

**Prompting Guide**

* **Name your character**: Give your character a name e.g. under encode character(Click adjust icon on the bottom side of the node), I have used name **sia**, so when passing prompt, I only have to say, **sia walking on the beach**.
   * *Again providing prompt like a woman or any features specific details like black hairs etc will only mislead the generation.*
* **Describe character features**: Encode all of the character features in encode character prompt & trigger your character with a name in generation prompt.
   * *Avoid describing same things in generational prompt.*
* **Handling Character drift**: e.g. if you want specific style or cloth e.g. half sleeves, sleeveless, add it to the generational prompt. There can be a slight drift in clothing as body shot also has cloths, which interferes with clothing references.
   * *Each refs should be unique, face should not have body or vice versa, same applies for clothing.*
* **Portability:** Once character is built, you can use the same character with only simple prompt & generation graph.

I have generated face with Flux Klein 9b & outfit pictures were taken from [Zara website](https://www.zara.com/in/en/short-dress-with-contrast-bow-p05029207.html?v1=556949208&v2=2420896)

Note: For best result, pass cropped references, so that model takes the required shot, model gets confused if cloth slot also has a face or face slot has cloths.

**Requirements**

Nvidia GPU: 24GB+ VRAM & 64 GB RAM(Run Locally)

**Workflows:** 

[Flux 2 Klein(Still)](https://www.omnichar.org/workflows/virtual-try-on-with-portable-consistent-character-flux2-klein-9b-4b), [Minimax H3(Video)](https://www.omnichar.org/workflows/minimax-h3-guided-consistent-characters-via-reference-identity-face-body-cloths)  
Note: Minimax H3, now also support int8 model varient

**Github Repo:** [https://github.com/omnichar/OmniChar](https://github.com/omnichar/OmniChar) (GPLV3, Opensource)

**Inputs:** Attached in the Flux2 workflow page

**Limitation:** Reference conflicts e.g. if two reference/input images has two different faces, it might conflict in generation, provide well cropped body & cloth images. Face images are crossed automatically by Sface.  
  
Happy to hear any suggestions or feedbacks.

document fetched 2026-09-28 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1wrm7gt/one_char_model_consistent_face_body_cloths