New.char finetuning and scoring are now available in the cloud
.char files

One character in one .char file, across models

A .char holds everything a model needs to draw the same person: the reference photos, a locked description, any LoRA trained from those photos, and a voice clip if you add one. Build it in OmniChar Cloud, OmniChar Studio or ComfyUI, and open it in the other two.

Characters← Characters
Name
Eval Subject
Description
a woman with a light scatter of freckles, natural unretouched skin
References (3)
  • Reference 1: the character facing the camera against a grey backdrop1
  • Reference 2: the character facing the camera, a second frontal take2
  • Reference 3: the character in profile, hair tied back3
Drop images here to add more. Order is what the prompt refers to.
OmniChar Studio · Characters panel
Inside the file

Your photos, plus what each model needs from them

A zip with manifest.json first. Only the references and the description are originals; everything else is built from them and rebuilt when they change.

Example · A character with a Krea 2 LoRA
emmy.char
  manifest.json              payload index, fingerprints, training record
  refs/000..004.png          your reference images, immutable
  derived/face_000..004.png  face crops, 512px
  text/description.md        the locked description, also the trigger
  scoring/                   per-reference embeddings and their centroids
  payloads/flux2-klein/      references resized for FLUX.2
  payloads/krea2-lora/       the trained Krea 2 adapter
refs/
Your original images, untouched. Everything else is derived from these.
text/
A short locked description, prepended to every prompt that uses the character.
derived/
A face crop per reference, cut at 512px by the detector.
scoring/
A 128-d face embedding and a 768-d subject embedding per reference, plus a centroid of each. What a take is measured against.
payloads/<model>/
References resized for a model that reads them, such as FLUX.2. Cache: deleted and rebuilt whenever the sources change.
payloads/<model>-lora/
A trained adapter for a model that has no reference channel, such as Krea 2, with the base it trained against and the strength to apply it at.
Reference order

The numbers in your prompt match the images

References are sorted by role, face first. FLUX.2 addresses them by position, so “the jacket from image 2” means the second reference in the file. Cloud, Studio and the ComfyUI node all keep that order.

  • Reference photo: the character facing the camera in soft key lightimage 1

    Frontal

  • Reference photo: the character turned to a near-profile viewimage 2

    Profile

Prompt text omnichar-sdk writes for each model family
ModelPrompt starts with
FLUX.2Images 1 and 2 show Ada, ...
MiniMax H3<Picture 1> <Picture 2> show Ada, ...
Seedance@Image1 and @Image2 show Ada, ...
A LoRAAda, a woman in her thirties ...
Models

How each model reads a character

A model either reads the reference images directly or uses a small adapter trained once from the same photos. One file can carry both.

ModelHow it takes the characterTraining
FLUX.2Reads the references directly.None
MiniMax H3 Reference to VideoReads the references directly, and the voice clip if there is one.None
Krea 2Uses a LoRA trained once from the same photos.LoRA, once
MiniMax H3, other nodesUses a trained LoRA; these nodes have no reference channel.LoRA, once
Nano Banana Pro (hosted)References and description are sent with the request.None
Seedance 2.0 (hosted)Build and wardrobe only. It rejects any reference with a face in it.None
Portability

Where a .char opens

  1. 01

    OmniChar Cloud

    Build a character from 1-3 photos in the browser. It downloads as a .char whenever you want the file.

    Try in browser →
  2. 02

    OmniChar Studio

    The Characters panel reads and writes models/characters on your own machine. Pick one on a generation node.

    Install Studio →
  3. 03

    ComfyUI

    Put the file in ComfyUI/models/characters and choose it on a Load Character node. Set INLINE_CHARACTERS_DIR to share one folder with Studio.

    ComfyUI integration →

In your own code

omnichar-sdk reads a .char in Python with only the standard library. It is Apache-2.0, so closed-source tools can use it.
# pip install omnichar-sdk
from omnichar_sdk import Character

char = Character.open("Ada.char")
char.get_description()                   # locked description
char.get_references(arch="flux2-klein")  # in prompt order
char.get_lora()                          # adapter, if any
char.get_voice()                         # voice clip or None

From any language

The same package installs a command. Add --json for the whole record.
$ omnichar-sdk inspect Ada.char
Ada  (2 references, format 1)

references: 1 body, 1 face
reference sets: flux2-klein
adapter z-image: rank 16, 600 steps, strength 0.8
Sharing and licensing

A character can travel with its terms

Publish a character for others to use

The showcase lists characters people have published, each with sample renders and the file to download.

Browse the showcase →

A licensed .char carries its license

An Ed25519-signed license block sits inside the file, with its scope and dates. OmniChar checks it before each generation, and the owner can revoke it.

How likeness licensing works →

Questions

What is a .char file?

A zip with manifest.json first. Inside are the reference photos (face, body and wardrobe), a locked description, face crops and embeddings used for scoring, any LoRA trained from the same photos, and an optional voice clip of about 30 seconds.

Can I read a .char without OmniChar?

Yes. omnichar-sdk opens one in Python with no other dependencies and is licensed Apache-2.0, so closed-source tools can use it. The same package installs an omnichar-sdk command whose inspect --json output any language can read.

Which models can use a character?

FLUX.2 and MiniMax H3 Reference to Video read the references directly. Krea 2 and the other MiniMax H3 nodes use a LoRA trained from the same photos. Nano Banana Pro gets the references and description with the request, and Seedance 2.0 takes build and wardrobe only. A model with no reference channel and no adapter cannot take a character.

Does every .char contain a LoRA?

No. FLUX.2 needs none, so a character built for it is references and a description. A LoRA is added when you train one for Krea 2 or MiniMax H3, and the file records the base model it trained against and the strength to apply it at.

What happens when the format changes?

Your reference images and description are the originals in the file. The resized references and adapters are derived from them, so if the format moves on they are rebuilt from your originals rather than migrated.

Build a character once, then take it from model to model