MiniMax H3 Colab Skill
Codex skill that makes short MiniMax H3 videos from your images on a Google Colab GPU, with a job queue and result download
- 136 ★ on the repository
- Last commit
Extended access
Why this level
- Runs a local runner and the Colab CLI and uploads your images and prompt to a Google Colab session
- Spends paid Colab compute units and keeps a Google token in the Colab CLI configuration
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/minimax-h3-colab-skillDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Runs a local runner and the Colab CLI and uploads your images and prompt to a Google Colab session
- Spends paid Colab compute units and keeps a Google token in the Colab CLI configuration
This is third-party code. Review the repository files before installing.
What it does
MiniMax H3 Colab Skill lets a Codex agent create short MiniMax H3 videos from local reference images through Google Colab. The local machine only prepares and uploads inputs, model inference runs on a Colab GPU, and the finished MP4 downloads to the path you choose. The repository holds SKILL.md, the MiniMax_H3_Turbo_Colab.ipynb notebook, the scripts/runner.py runner, the run_colab_inference.sh launcher, the install.sh installer and tests with a fake Colab CLI. The default Ref2VA mode takes one to nine ordered images that the prompt references as <Picture 1> and onward; the first_frame mode takes one image as the first frame. The prompt file is uploaded unchanged. If you give none, the agent writes a full prompt with subject definitions. The default clip length is 12 seconds and H3_DURATION_SECONDS changes it from 4 to 15. The default request is an A100 with high memory. The COLAB_AUTH, COLAB_GPU, COLAB_HIGH_MEM, COLAB_EXEC_TIMEOUT and COLAB_SESSION_NAME variables control sign-in, GPU and session. python3 scripts/runner.py usage --json shows the Colab unit balance before a run.
Who it is for. People with a Google Colab account with compute units who want to generate short image-guided videos from Codex without a local GPU.
Good fit when
- You have one to nine images of characters or scenes and want a clip from them by a text description
- The video should start from a ready frame: first_frame mode
- A series of clips should run in sequence with a job manifest
- Before running you need to know the Colab compute unit balance
- You have no local GPU and Colab with an A100 is available on your plan
Not a fit when
- You need the model to run on a local computer: all computation happens in Colab
- The Colab account has no compute units or the needed GPU: sufficient balance does not guarantee a GPU allocation
- You need long clips: duration is limited to 4 to 15 seconds
- The agent is not Codex: the skill is written and installed for Codex
Example request
Use minimax-h3-colab: take two character images from the refs folder, write an 8 second prompt and save the video to out/intro.mp4Limitations
- Codex, Python 3.11 or newer and uv are required, and the Colab CLI (google-colab-cli) needs Python 3.12 or newer; the author validated version 0.7.4.
- A Google account with Colab compute units is required: an A100 or an equal high-memory runtime may need a suitable plan and balance, and sufficient balance does not guarantee a GPU allocation.
- First sign-in goes through the Colab CLI OAuth2: the CLI stores the token and it must not go into the repository, a prompt or a manifest.
- The optional ffprobe checks that the MP4 has video and audio streams.
How to disable. Delete the ~/.codex/skills/minimax-h3-colab folder (or $CODEX_HOME/skills/minimax-h3-colab). Colab sessions are stopped with the runner stop command and the Colab CLI token lives in its configuration.
FAQ
How do I install MiniMax H3 Colab Skill?
Run git clone https://github.com/killkli/minimax-h3-colab-skill.git, enter the folder and run ./install.sh. Then uv python install 3.12 and uv tool install --python 3.12 google-colab-cli.
How do I authorize the Colab CLI?
Run colab --auth=oauth2 usage and follow the URL and code in the output. The application default credentials option is set with COLAB_AUTH=adc.
How do I run one video?
Run ./run_colab_inference.sh --image /path/ref.png --prompt /path/prompt.txt --output /path/result.mp4. The --image flag can repeat from one to nine times.
What video length can I set?
From 4 to 15 seconds through H3_DURATION_SECONDS, 12 seconds by default.
Does the model run on my computer?
No, the local computer only prepares and uploads inputs. Inference runs on the Colab GPU and the MP4 downloads back.
Discussion
No questions or reviews yet.
Related
SkillsThe skill retells a reference video, waits for your confirmation and writes MiniMax H3 prompts per scene and shot.
Three Codex and Claude Code skills: search MiniMax H3 cases and full prompts, pick tutorial routes and prepare submissions
The skill turns a short idea into a full MiniMax-H3-Image prompt and can render the picture through the MiniMax API.
The official mmx CLI of the MiniMax platform with an agent skill: text, images, video, speech, audio transcription, vision and search
Local video with MiniMax H3 on your own GPU: the open-video CLI and a skill that lets an agent write prompts and run clips.
A Codex skill that renders images, SVGs and scripts as hand-drawn whiteboard videos, with 30 styles and narration
README in short
The README in English and Traditional Chinese explains that the repository holds a complete Codex skill and gives a four-command quick start. Detailed sections cover requirements, installing the skill and Colab CLI, single runs, batch mode, environment variables, session behavior, limits and troubleshooting.
README badge
Original SKILL.md
---
name: minimax-h3-colab
description: Create short MiniMax H3 first-frame or reference-to-video clips from local images through Google Colab CLI. Use when a user asks for image-guided H3 video generation or a sequential batch of clips.
---
# MiniMax H3 on Colab
Use the bundled notebook and `scripts/runner.py` to generate MiniMax H3 video from local images. Jobs default to Ref2VA; choose `first_frame` for FL2VA first-frame conditioning. The Colab CLI must be installed and authenticated with an available compute-unit balance and GPU allocation.
## Workflow
1. Inspect the requested image(s) and any supplied prompt file. `reference` accepts 1–9 ordered images mapped to `<Picture 1>` through `<Picture 9>`. `first_frame` requires exactly one image and uses FL2VA first-frame conditioning.
2. If the user supplied a prompt file, pass it through unchanged. Otherwise, for `reference`, 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. For `first_frame`, describe motion and shots from the supplied opening image using FL2VA prompt conventions.
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:
```json
{
"jobs": [
{
"title": "intro",
"inference_mode": "reference",
"reference_images": ["/absolute/path/girl.jpg"],
"prompt_file": "/absolute/path/prompt.txt",
"width": 1344,
"height": 768,
"duration_seconds": 12,
"output_name": "intro"
}
]
}
```
Omit `inference_mode` to keep the compatible `reference` default. Width and height default to 1344×768; each must be a positive multiple of 32 and their area must not exceed 1344×768 pixels. A first-frame job must have exactly one reference image.
Run from this skill directory: