Skip to main content
Each workflow is a JSON file in the ComfyUI-Omnichar repository. Drag one onto the ComfyUI canvas to open it.

​
Build a character

A character is a few reference images plus a description. Load a face, a body shot and one or two wardrobe shots into Encode Character. It writes a .char, Save Character puts it in ComfyUI/models/characters/, and Load Character picks it up from there. Download character_encode.json

​
Render with FLUX.2 Klein 9B

FLUX.2 reads references as latents. Decode Character’s refs goes into a Character Reference Latent node on both the positive and the negative conditioning. The result is an image. Download flux_klein_9b_image_char.json

​
Render with MiniMax H3

MiniMax H3 takes references in numbered slots. refs goes into a Character References Split node, which fills the slots in order. The result is a video. Download minimax_h3_char_video.json

​
Voice

A character can carry a voice for MiniMax H3 Reference to Video.
1

Add the voice

Wire Load Audio into Encode Character’s voice input. Use about 30 seconds of clean speech with no music.
2

Send it to H3

On Decode Character, set style to token and wire voice into the H3 node’s ref_audio_0.
With voice set to auto, the clip is sent only when the prompt has dialogue: The match is on words, so “no speaking” still counts as dialogue. Set voice to never for a silent shot like that, or always to force the voice in.