H3-colab-studio

H3 Colab Studio

English | 繁體中文

An agent skill and plugin for building and running a MiniMax H3 Prompt-to-Video Workspace bridged to Google Colab G4 (ComfyUI).

A single MiniMax H3 generation renders one continuous take with no internal cuts. To produce cohesive multi-shot stories without watermarks, this workflow pairs three parts:

1. Control-Surface Web App (app-blueprint.md): A persistent job queue and live prompt-draft preview. The UI is explicit that GPU generation runs on Colab rather than inside the browser app.
2. Official H3 Multi-Shot Prompt Structure (prompt-spec.md): Converts user ideas in any language into structured English shot blocks with verbatim character/palette anchors and a shared fixed noise_seed.
3. Verified Colab G4 Runbook & Stitching (colab-runbook.md, stitch.sh): Runs all shots in a single Colab G4 session (RTX PRO 6000 Blackwell, ~96 GB VRAM), applies two critical ComfyUI fixes before queueing, stitches MP4 clips locally via ffmpeg, verifies stream metadata with ffprobe, and tracks exact Colab compute-unit burn.

Repository Layout

.
├── skills/
│   └── h3-colab-studio/               # Universal Agent Skill (Antigravity, Claude Code, OpenSkills)
│       ├── SKILL.md                   # Entrypoint instructions with YAML frontmatter
│       ├── skill.yaml                 # Packaged skill manifest
│       ├── references/
│       │   ├── app-blueprint.md       # Persistent job-queue web app specification
│       │   ├── colab-runbook.md       # Verified Colab G4 + ComfyUI setup & failure fixes
│       │   └── prompt-spec.md         # Official MiniMax H3 prompt structure & multi-shot guide
│       └── scripts/
│           └── stitch.sh              # ffmpeg concat + ffprobe stream verification
├── plugins/
│   └── h3-colab-studio/               # Native Google Antigravity Plugin bundle
│       ├── plugin.json
│       ├── README.md
│       ├── assets/logo.svg
│       └── skills/h3-colab-studio/
├── dist/
│   └── h3-colab-studio.skill          # Ready-to-import .skill archive
├── .claude-plugin/                    # Claude Code plugin marketplace manifests
└── install.sh                         # One-click local installer

Prerequisites

• Google Colab Compute Units: A Google account with Colab compute units (for example, Google AI Ultra or Colab Pro/Pro+) able to select the G4 runtime (NVIDIA RTX PRO 6000 Blackwell Server Edition, ~96 GB VRAM).
• Local FFmpeg: ffmpeg and ffprobe installed locally for multi-shot concatenation and output verification (brew install ffmpeg on macOS, apt install ffmpeg on Ubuntu/Debian).

Installation

Option 1: One-Click Installer (Antigravity & Claude Code)

git clone https://github.com/Rea1Felix/H3-colab-studio.git
cd H3-colab-studio
chmod +x install.sh
./install.sh all

Available modes:

• ./install.sh all — Installs both the Antigravity standalone skill (~/.gemini/config/skills/) and native plugin (~/.gemini/config/plugins/).
• ./install.sh skills — Installs the standalone skill to ~/.gemini/config/skills/h3-colab-studio.
• ./install.sh plugins — Installs the native Antigravity plugin to ~/.gemini/config/plugins/h3-colab-studio.
• ./install.sh claude — Installs the skill to ~/.claude/skills/h3-colab-studio.

Option 2: OpenSkills / Agent Skills CLI

openskills install Rea1Felix/H3-colab-studio

Or with npx skills:

npx skills add Rea1Felix/H3-colab-studio

Option 3: Packaged .skill Archive

Import dist/h3-colab-studio.skill directly into platforms that accept .skill zip bundles.

Key Verified Findings (Colab G4 + ComfyUI)

• Measured G4 Burn Rate: ~8.94 compute units/hour (~1.49 units per 10 minutes), slightly below the ~10.05 units/hour estimate shown in the Colab resource panel.
• Two Mandatory Pre-Flight Fixes (documented in colab-runbook.md):
  1. Missing Turbo LoRA: Download lightx2v/Minimax-h3-Turbo (minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors, ~1.96 GB) into /content/ComfyUI/models/loras before queueing, even when Turbo is toggled off.
  2. SaveVideo Format Check: Verify the SaveVideo node's format field is explicitly set to mp4 before every run so finished renders are not dropped at save time.
  3. Avoid keboqi/minimax-h3 Prism Studio UI: Use the direct ComfyUI notebook (dseditor/minimax-h3-colab, H3_Basic_ComfyUI.ipynb) instead.

Usage Examples

Once installed, ask your agent naturally:

• Build the workspace app:

  "Use h3-colab-studio to build the local prompt-to-video job-queue web app following the blueprint."

• Draft a multi-shot H3 prompt:

  "Optimize this idea into a 3-shot MiniMax H3 prompt (5s per shot, 1376×768, fixed seed): a deep-sea diver discovers an ancient bioluminescent clockwork cathedral."

• Execute & stitch a queued job:

  "Run H3 job #1 on Google Colab G4 following the runbook, stitch the downloaded shots with stitch.sh, and backfill the output specs and compute units used."

License

Released under the MIT License.