宜博 / lemonboss

Your AI film director. Give me a script, and I'll give you the final cut — an automated pipeline: storyboard → MiniMax H3 generation → one-click JianYing assembly.

Features

• 🎬 Automatic storyboarding (WenWu director engine): Asks for your story, then generates a storyboard with shot-to-shot continuity anchors, generates character four-view sheets (three full-body angles + one face close-up, via image generation or web-sourced reference), and writes H3 prompts — in either the official six-part English format or the WenWu Chinese director format, with shot density adapting to any duration.
• 🎞️ MiniMax H3 workflow routing: Automatically routes to Ref2VA (reference images), T2V (pure text), or I2V (first/last frames), and patches prompt, duration, steps, seed, resolution, and output prefix by class_type — never hard-coded node IDs.
• ✅ Parameter confirmation: Summarizes all parameters into one table for your review before submission; invalid values (duration 5-15s, steps 4-40, etc.) are rejected by backend validation.
• ☁️ Cloudflare tunnel: Detects local 8188 / existing tunnels / AutoDL automatically, and starts or reuses a TryCloudflare tunnel when remote access is needed.
• 📥 Batch submission & resumable monitoring: Uploads reference images, submits each segment, polls and downloads results with automatic retries and persistent state.
• ✂️ JianYing auto-assembly: Uses jianying-editor to build a JianYing draft (ordered clips + dissolve transitions), with optional automatic MP4 export.
• 🖥️ First-use onboarding: Asks about auto-update, detects ComfyUI, evaluates hardware, guides model downloads from ModelScope, and asks — with a clear purpose for each — whether to install helper skills and the Comfy MCP server.

Installation

Install from skills.sh (recommended):

npx skills add TFboy1/oh-my-minimaxh3-director --skill oh-my-minimaxh3-director

Or clone for local development:

git clone https://github.com/TFboy1/oh-my-minimaxh3-director.git

First Use

On first invocation, the skill confirms the following with you, explaining the purpose of each item before asking:

1. Enable auto-update? (recommended);
2. Is ComfyUI installed? (otherwise a desktop download tutorial is provided);
3. Run hardware detection to see whether your machine can run MiniMax H3;
4. Are the MiniMax H3 models downloaded? (otherwise a ModelScope tutorial is provided);
5. Optional dependencies: h3-prompt-writing (prompts), jianying-editor (assembly), comfy-mcp (ComfyUI management).

Full tutorial: references/setup-guide.md.

Usage

In Codex, just say:

Use $oh-my-minimaxh3-director to turn this script into a video

Then provide a script file or story text. Pipeline:

script/story → story requirements → storyboard (with continuity) → character
four-view reference sheets → H3 prompts (official / WenWu) → H3 workflows
→ parameter confirmation → submit & generate → download clips
→ JianYing draft → (optional) auto export

Manual step-by-step:

# 1. Probe ComfyUI (open a tunnel when remote access is needed)
python scripts/probe_comfy.py --workspace <workspace root> --write
python scripts/ensure_cloudflared.py --workspace <workspace root>

# 2. Build per-segment workflows and the parameter table
python scripts/build_workflows.py --project <project dir> --workspace <workspace root>

# 3. Submit and monitor (resumable)
python scripts/submit_jobs.py --project <project dir> --workspace <workspace root>
python scripts/monitor_jobs.py --project <project dir> --workspace <workspace root>

# 4. Generate and run the JianYing assembly script
python scripts/generate_assembly.py --project <project dir> --title <film title>
python <project dir>/assemble_jianying_<film title>.py

Hardware Requirements

Run python scripts/check_hardware.py (or --remote-url <URL> to evaluate a remote GPU):

VRAM	Verdict	Suggestion
	Not recommended locally	Cloud GPU (AutoDL 24GB+) or Comfy Cloud
12-16GB	Entry-level	int8/nvfp4 quantized + 4-step Turbo, 0.4MP shorts
16-24GB	Recommended	0.4-0.6MP, 10s clips
24GB+	High-end	0.6-1MP, 10-15s clips

Models total ~40GB; NVIDIA GPU + CUDA required; reserve 60GB+ disk.

Project Structure

oh-my-minimaxh3-director/
├── SKILL.md                    # main skill workflow
├── README.md
├── banner.svg                  # repo banner
├── agents/openai.yaml          # UI metadata
├── assets/templates/           # H3 templates (ref2va / t2v / i2v, API format)
├── references/
│   ├── setup-guide.md          # first-use setup: ComfyUI / models / hardware / deps
│   ├── storyboard-schema.md    # storyboard JSON & prompt spec
│   ├── workflow-routing.md     # template routing & field mapping
│   └── cloudflared.md          # tunnel operations
├── scripts/                    # pipeline scripts
└── evals/evals.json            # test cases

FAQ

• node_errors on submit: usually a missing custom node or wrong model filename on the remote ComfyUI; the error message names it — install/fix accordingly.
• Local OOM: lower resolution (0.4MP → 0.2MP), shorten clips, switch to 4-step Turbo, or move to a cloud GPU.
• Unstable tunnel: ensure_cloudflared.py --stop then restart, or fall back to the AutoDL direct address.
• Custom templates: convert UI-format workflows with convert_ui_workflow.py and put the API-format JSON into templates_dir.

Credits

This project stands on the shoulders of these open-source projects:

• MiniMax H3 (MiniMax-AI): the model and the h3-prompt-writing skill, under the MiniMax H3 community license;
• Comfy-Org/minimax-H3: ModelScope repackaged models for fast downloads in China;
• ComfyUI: video generation engine, GPL-3.0;
• jianying-editor-skill (luoluoluo22): JianYing assembly via JyWrapper, MIT;
• comfy-mcp (Comfy-Org): ComfyUI MCP server, AGPL-3.0-or-later OR Commercial;
• cloudflared (Cloudflare): remote tunnels, Apache-2.0.

Full list: CONTRIBUTORS.md

License

The skill code is open-sourced under the MIT license — see LICENSE. Model weights follow the MiniMax H3 community license; upstream components (ComfyUI, comfy-mcp, jianying-editor-skill, cloudflared) keep their own licenses.