# H3 reference-video starter — free download

Turn the movement in a short video into a different character using one photo.
This is a normal, editable ComfyUI workflow. No ModelPilot account, private
nodes, API key, paid adapter or ModelPilot runtime is required.

**Free means the files are free.** Running the model needs your own compatible
GPU or rented compute, electricity and model downloads. No GPU minimum, runtime,
cost or identity accuracy has been measured for this standalone starter.

## Run it

1. Install ComfyUI and its Python requirements using the official instructions:
   https://docs.comfy.org/installation/manual_install . Use a CUDA installation
   supporting H3 and the listed quantized weights. The source baseline is
   ComfyUI commit `6b747c0428c343e1417219641db93a4fb7cb69ae`; newer versions
   can change node behavior. Update your frontend if H3 nodes are missing.
2. Download the four files linked in `models.json` from Comfy-Org/MiniMax-H3.
   Put each in the stated `ComfyUI/models/…` directory. These are large model
   files, not included in this download. Restart ComfyUI after installing them.
3. Copy `motion.mp4` and `character.png` into `ComfyUI/input/`.
4. Drag `workflow.json` onto the ComfyUI canvas. Check that Load Video shows
   `motion.mp4`, Load Image shows `character.png`, and the four loaders show
   the filenames in `models.json`.
5. Read the editable instruction, then click **Run** once you are ready to use
   your compute. The generated MP4 is saved under `ComfyUI/output/video/`.

For API users, `workflow-api.json` is the equivalent prompt graph. Submit it
under the ordinary ComfyUI `/prompt` API's `prompt` key after installing the
same files. It does not run automatically or call a hosted service.

## What it does

- `<Video 1>` supplies movement, timing, camera and background.
- `<Picture 1>` supplies the fictional adult character's appearance and outfit.
- Output: **480 × 832, 24 fps, 124 frames (5.17 seconds), 20 steps**, fixed seed
  1414213562, `res_multistep` sampler and `simple` scheduler.
- Input video: the first 124 frames resampled to 24 fps, fit inside 320 × 576,
  with audio removed. The supplied reference photo is unchanged.
- H3 generates video and audio jointly. Source audio is not preserved. The
  instruction requests no music or dialogue; that is not a guaranteed mute.

This deliberately small example demonstrates native reference conditioning.
It **does not contain pose extraction, tracking, background restoration, face
repair, long-video windows, worker management or billing**. It is not a replay
of Pose Efficient v1, H1, HM4 or a published study result. Those options have
different conditioning and delivery logic. Face identity, clothing, movement
and the background can all drift; a prompt is not a hard preservation mask.

To use your own inputs, replace the two files and update the loader selections
and prompt. Keep source footage at 24 fps and choose an H3-valid frame count
(`length = 17n + 5`, e.g. 124). Match the intended duration rather than silently
cutting a longer source. The native reference-video path supports 2–15 seconds;
this starter does not add stitching or extension for longer videos.

## Verification and limits

The package is checked against real registered ComfyUI node schemas and its
included media is decoded and saved through core video nodes on CPU. The
included API graph is checked against the same UI connections and settings.
**A fresh GPU generation of this exact standalone package has not been run.**
Schema/media checks cannot establish model fit, GPU compatibility, output
quality or generation speed. No saved ModelPilot result is presented as the
output of this smaller standalone recipe.

## Credits and licenses

Workflow adapted from the repository's standard `minimax_h3_r2v.json` template,
based on Comfy Org's workflow templates (MIT; included `LICENSE-WORKFLOW.txt`).
H3 weights are separately licensed: read `MINIMAX-H3-LICENSE.txt` and the model
card. ModelPilot's own authorization does not license another person's use.

Motion source: Andres Daza, Pexels,
https://www.pexels.com/video/a-woman-in-a-sports-bra-and-shorts-is-dancing-19192491/
under https://www.pexels.com/license/ . This sample is a cropped/resampled,
silent excerpt, not a performer endorsement. Character reference: ModelPilot's
fictional AI-generated adult in plum dancewear, supplied for this example.
No celebrity photo is needed. See `manifest.json` for file hashes and sources.