MiniMax H3 on Colab

Use the bundled notebook and scripts/runner.py to generate Ref2VA video from local reference images. The Colab CLI must be installed and authenticated with an available compute-unit balance and GPU allocation.

Workflow

1. Inspect the requested images and any supplied prompt file. Keep image order stable: image 1 maps to <Picture 1>, image 2 to <Picture 2>, through image 9.
2. If the user supplied a prompt file, pass it through unchanged. Otherwise write a complete Ref2VA prompt with subject_definitions, summary, retention_analysis, detailed_description shot blocks, overall_soundscape, and non_diegetic_music; include dialogue in its shot with <d>[Chinese] ...</d> when requested. Use the bundled runner's compose_ref2va_prompt helper for structured shot input.
3. Check Colab access and balance with python3 scripts/runner.py usage --json. Sufficient compute units do not guarantee that Colab will allocate the requested GPU or high-memory runtime. Report allocation failures plainly.
4. Put the jobs in a UTF-8 JSON manifest and run one batch. The runner uploads each job's images and prompt, executes the bundled notebook sequentially on the same live session, downloads each MP4, and stops the session when it created it. If reusing a session, pass --stop-on-complete when the user wants it stopped after the batch.
5. Confirm each output exists and report its local path. Preserve completed clips if a later job fails.

Example manifest:

{
  "jobs": [
    {
      "title": "intro",
      "reference_images": ["/absolute/path/girl.jpg"],
      "prompt_file": "/absolute/path/prompt.txt",
      "duration_seconds": 12,
      "output_name": "intro"
    }
  ]
}

Run from this skill directory:

python3 scripts/runner.py batch --manifest /absolute/path/jobs.json --gpu A100 --timeout 10800 --output-dir /absolute/path/outputs

Use --no-high-mem when high-memory allocation is unavailable or not desired. For a named existing session, add --session SESSION --stop-on-complete if the requested workflow should end that session.

Operational limits

• Each job requires 1–9 non-empty reference images, a non-empty UTF-8 prompt, and a duration from 4–15 seconds.
• Keep all jobs in one batch to reuse the same Colab session and loaded ComfyUI/model state.
• Do not blindly retry a timed-out colab exec: the remote kernel may still be working. The runner stops the session during cleanup and records completed jobs before reporting the failure.
• Do not expose OAuth credentials or runtime tokens in prompts, manifests, browser state, or logs.
• The bundled assets/MiniMax_H3_Turbo_Colab.ipynb is the inference notebook used by the runner.