Omnichar packages a character's references, description and optional voice clip into a reusable .char file. Its ComfyUI nodes can unpack that file into MiniMax H3's reference inputs, making it easier to reuse the same character preparation across shots. The file is a portable input asset; the generated video still depends on H3, the selected checkpoint and the scene prompt.
This guide follows the ComfyUI-Omnichar README, its voice generation workflow, and the author's voice announcement, reviewed October 10, 2026. These are documentation-based instructions. We have not installed the nodes or reproduced voice consistency. H3Vid does not currently import .char files.
What the character file actually contains
The project describes a .char as reference images plus a fixed character description, sometimes with a trained LoRA. The updated encoder can also store reference speech. Loading the file therefore restores a prepared set of inputs rather than requiring you to locate each image and recording separately for every scene.
A stored description does not impose a hard identity lock on generated pixels. Reference selection, ordering and prompt interpretation still matter. Nor does the presence of a LoRA mean the same adapter weights work with every supported model. The package provides an explicit Apply Character LoRA node; check the intended architecture before using it. Portability of the container and compatibility of its contents are separate questions.
The README lists MiniMax H3, Krea 2 and several FLUX.2 variants. Its voice route is specifically described for H3 reference-to-video. Do not assume every image model can consume speech because it can read the character's pictures.
Prepare the local environment and assets
Start with a working MiniMax H3 ComfyUI setup. Omnichar requires ComfyUI, Python 3.10 or later and its SDK reader. The repository instructs users to clone the node into ComfyUI/custom_nodes, install requirements.txt, then restart ComfyUI. Use the Python environment that actually launches ComfyUI; installing dependencies into a different system interpreter will not make them available to the running app.
Store saved character files in ComfyUI/models/characters/. If you already use Omnichar Studio, the documented INLINE_CHARACTERS_DIR option lets both tools share a character directory. Keep a copy of the original file when experimenting with a new description or reference set.
Choose visual references with clear roles: face, body and clothing. Encode Character accepts three slots for each role and sorts references by role, with face first. Review the resulting numbered sheet before writing a prompt that mentions particular reference positions. A stale assumption about which image occupies a slot can send H3 contradictory instructions.
Build the voice character, then connect it to H3
The repository includes a voice encoding workflow. Its documented chain is:
- Load the visual references and a speech recording you are permitted to use.
- Connect Load Audio to Encode Character's
voiceinput. - Save the encoded character with Save Character.
- Load that file and pass its
charoutput to Decode Character. - Use Decode Character's
tokenstyle for the documented H3 voice route. - Connect
refsto Character References Split, then connect the numbered image outputs to the H3 reference-image slots. - Connect Decode Character's
voiceoutput to H3'sref_audio_0, and its prompt output to the prompt input.
The references output is a combined image batch. The refs output is the structured reference collection consumed by the split node. Confusing the two can leave a graph with the wrong input type or lost ordering. The current supplied voice workflow visibly connects the decoded audio to the H3 node; a successfully loaded .char alone does not establish that audio reached generation.
Choose a speech sample and an explicit dialogue policy
The current README recommends approximately 30 seconds of clean speech without music. The author's announcement discusses short samples, but that does not establish a universal five-second minimum or equivalent quality at every length. Prefer one speaker, clear pronunciation and a representative tone. Multiple voices make the desired identity less clear.
Decode Character offers auto, always and never voice behavior. In auto, dialogue-related words determine whether the clip is forwarded. This is a word match, so a negative instruction such as “no speaking” can still trigger it. Use never for a deliberately silent shot and always when you explicitly want the reference sent.
For a first experiment, use one visible speaker, a static camera and a short English line. For example: “The character faces the camera and says, ‘The next scene starts here.’” This is an editorial test prompt, not a demonstrated output. The author reports weaker behavior with non-English speech; test the actual language you need rather than extrapolating an English example.
Reference voice is different from an exact soundtrack
The README describes H3 producing new words with a voice resembling the recording. That differs from matching mouth motion to a fixed, approved audio track. If exact wording, timing or an unchanged performance is essential, evaluate a dedicated lip-sync workflow against that requirement.
The author says lip sync comes from H3 itself; Omnichar does not add a separate mouth-animation pass. Listen for pronunciation, speaker drift and unexpected voices, then inspect mouth timing independently. A convincing face is not proof of correct dialogue.
For a useful local comparison, keep the character, checkpoint, prompt and seed fixed. Save one speaking result with the reference voice, one without it, and a silent shot using never. Then change only the scene and repeat the speaking line. Record which references were selected and the node revision. This separates wiring mistakes from cross-shot consistency problems and gives you a concrete next step: retain the character asset only when both its visible identity and its audio meet your own production requirements.
The cover is an AI-generated concept illustration, not an Omnichar screenshot, MiniMax H3 output or project test result.