New.char finetuning and scoring are now available in the cloud
Integrations · ComfyUI

OmniChar characters in your ComfyUI graphs

ComfyUI-Omnichar adds eight nodes that load a .char into the workflows you already run. The same face, body and clothes carry across FLUX.2 dev, FLUX.2 Klein 9B and 4B, Krea 2 and MiniMax H3, and a character's voice carries to MiniMax H3.

Open source under GPL-3.0. Needs Python 3.10 or newer.

The example FLUX.2 Klein 9B workflow in ComfyUI: Load Character with sia.char feeds Decode Character, whose refs go into Character Reference Latent nodes on the positive and negative conditioning, and Save Image shows sia sitting on a studio floor.
Example · flux_klein_9b_image_char.json. Nodes tagged Omnichar come from this package.
Install

Four steps to a loaded character

In ComfyUI's Python environment
cd ComfyUI/custom_nodes
git clone https://github.com/omnichar/ComfyUI-Omnichar
cd ComfyUI-Omnichar
pip install -r requirements.txt
  1. 01

    Clone the node

    In ComfyUI/custom_nodes, git clone the ComfyUI-Omnichar repository.

  2. 02

    Install the reader

    Run pip install -r requirements.txt in ComfyUI's environment. It installs omnichar-sdk, which has no other dependencies.

  3. 03

    Restart ComfyUI

    The Omnichar nodes appear under their own badge after a restart.

  4. 04

    Add characters

    Put .char files in ComfyUI/models/characters/. Load Character lists whatever is in that folder.

Set INLINE_CHARACTERS_DIRto OmniChar Studio's characters folder and both read the same files. Install docs →

Wiring

One character, wired to each kind of model

Load Character feeds Decode Character, which turns the file into a prompt and an ordered reference list. Where the references go next depends on how the model reads them.

FLUX.2

Decode Character's refs go into a Character Reference Latent node on both the positive and the negative conditioning.

MiniMax H3

refs go into Character References Split, or one Character Reference node per numbered slot.

Krea 2 and other LoRA models

Apply Character LoRA patches MODEL and CLIP with the adapter the character carries.

Anything that takes one batch

Read the references output: the same resolved set, in the same order.

Reference

The eight nodes

NodeWhat it doesInputsOutputs
Load CharacterPicks a .char from models/characters and checks its license if it has one.char, char_path, outputchar
Decode CharacterTurns a character into a prompt and an ordered reference list.char, style, clip, prompt, arch, max_references, size_from, fit, voice, audio_positionconditioning, references, refs, sheet, prompt, voice
Character ReferenceOne reference by slot number.refs, indeximage, role, count
Character References SplitFans references out to numbered slots, as MiniMax H3 takes them.refsimage_0 to image_4, count
Character Reference LatentFeeds references to FLUX.2 as latents, on the positive and negative side.conditioning, refs, vaeconditioning
Apply Character LoRAApplies the LoRA the character carries, for Krea 2 and other adapter models.model, clip, char, strength, arch, min_key_coveragemodel, clip
Encode CharacterBuilds a new character from reference photos and an optional voice clip.name, description, resolution, face / body / cloths (3 slots each), voicechar
Save CharacterWrites the .char to models/characters.char, filename, overwritepath
Every input explained in the nodes reference →
Voice

A character can carry a voice for MiniMax H3

Wire about 30 seconds of clean speech into Encode Character's voice input. On Decode Character, use the token style and wire voice into the H3 node's ref_audio_0.

With voice on auto, the clip is sent only when the prompt has dialogue
PromptVoice on auto
sia faces the camera and says "I built this look one frame at a time."Sent
close-up of sia as she whispers to someone off cameraSent
sia walks along the lake at duskNot sent

H3 speaks new words in a voice like the clip, so expect a likeness rather than a copy. The match is on words, so set never for a silent shot whose prompt still mentions speech, or always to force it in.

Voice workflows →
Licensed characters

Checked before each run

The license comes with the file

A character licensed to you carries its license inside the .char. Load Character sends the license ID and your secret to OmniChar before each run, and the run goes ahead only if the license covers this character and output. The secret stays on your machine in ComfyUI/user/omnichar/licenses.json, never in a workflow, image or log. Unlicensed characters never touch the network.

How likeness licensing works →

GPL-3.0 nodes, Apache-2.0 reader

The nodes are GPL-3.0-or-later. The omnichar-sdk package inside the repository is Apache-2.0, so other tools, closed-source ones included, can read .char files too.

What a .char file holds →

Questions

How do I install the OmniChar node in ComfyUI?

Clone github.com/omnichar/ComfyUI-Omnichar into ComfyUI/custom_nodes, run pip install -r requirements.txt in ComfyUI's environment, and restart ComfyUI. It needs Python 3.10 or newer.

Where do .char files go?

In ComfyUI/models/characters/. Load Character lists whatever is in that folder. Set INLINE_CHARACTERS_DIR to OmniChar Studio's characters folder and both read the same files.

Can I build a character inside ComfyUI?

Yes. Encode Character takes face, body and wardrobe references (three slots each), a description and an optional voice clip, and Save Character writes the .char. A file built in OmniChar Cloud or Studio loads the same way.

Does a character need the internet in ComfyUI?

Only a licensed one. Before each run, Load Character sends the license ID from the file and your license secret to OmniChar to check it. An unlicensed character never touches the network.

Bring your characters into the ComfyUI workflows you already run