简体中文 · ⚙️ Skill configuration · 🎬 Example Gallery · ⚡ Turbo Report

🎬 MiniMax-H3 Drama

Turn a story, campaign brief, or reference pack into a finished drama video—inside Codex.
👤 Character design · 🏙️ Scene design · 🧩 Storyboards · 🎞️ MiniMax H3 video · ✂️ Post-production · ✅ QC

Codex Plugin · MiniMax H3 · GPT-Image · ComfyUI · FFmpeg

MiniMax-H3 Drama is a Codex-first video production plugin, not just a prompt collection. Codex plans the production, creates reusable visual sources of truth, routes every shot to the right MiniMax H3 workflow, monitors local ComfyUI, and delivers a resumable project with picture, sound, captions, and QC.

✨ Why creators use it

Highlight	What you get
🎭 A complete production, not one clip	Brief → character/product bible → scene masters → storyboard → keyframes → shots → master
🧬 Cross-shot consistency	Identity, wardrobe, product geometry, locations, props, screen direction, light, and audio are tracked explicitly
🧠 Profile-driven direction	Built-in short-drama and commercial grammar, plus reusable profiles distilled from your own reference videos
🛠️ Local, deterministic finishing	Official H3 workflows through local ComfyUI; versioned FFmpeg editing, mixing, captions, exports, and QC
⚡ Turbo by default	T2V, I2V, and R2V use the MiniMax H3 Turbo LoRA at 6 steps; [turbo=false] keeps the original 20-step graphs available
🔁 Resumable by design	Every prompt, input hash, workflow, prompt_id, take, selection, assumption, and artifact stays in the project ledger

From a natural-language request and references to a returned video without leaving the Codex task.

See More Drama Examples in Gallery

🚀 Quick start

1. Install for Codex

This repository is packaged as a Codex Plugin with nine skills and a pinned local ComfyUI MCP connection. Install the complete bundle from its repository marketplace:

codex plugin marketplace add chiphoton/MiniMax-H3-Codex-Drama
codex plugin add minimax-h3-drama@chiphoton
codex plugin list --json

Start a new Codex task after installation so the bundled skills and MCP server are loaded.

For a skills-only installation:

npx skills add chiphoton/MiniMax-H3-Codex-Drama --all -g -a codex -y

The skills-only route does not install the bundled ComfyUI MCP connection.

Migrating from 0.1.x? Remove the old minimax-h3-prompt-skills plugin before installing minimax-h3-drama so the specialist skills are not registered twice.

2. Prepare the local runtime

Install Node.js and npm: the plugin launches the pinned comfyui-mcp@0.49.3 package through npx, and its first launch may need network access to populate the npm cache. Run ComfyUI at http://localhost:8188 with compatible MiniMax H3 models and nodes. Install FFmpeg/FFprobe for assembly and QC.

Turbo is enabled by default. Install ComfyUI-MiniMax-H3-Turbo through ComfyUI-Manager (search MiniMax-H3 Turbo) or under ComfyUI/custom_nodes/, restart ComfyUI, and place minimax_h3_turbo_v4_step600_ema.safetensors from the Turbo LoRA repository in ComfyUI/models/loras/. Use [turbo=false] for the original workflow when those Turbo dependencies are intentionally unavailable.

The plugin checks the environment before expensive work; it does not silently download models, install ComfyUI nodes, start services, or restart ComfyUI.

3. Direct your first drama

$minimax-h3-drama-producer
Create a 45-second 9:16 workplace reversal drama from my attached story.
Preserve my characters and dialogue. Design a consistent character sheet and
office scene master, build an 8-shot storyboard, generate every shot, add
captions and sound, and deliver the finished vertical master.
Use the tiktok-short-drama profile. [mode=fast]

Remove [mode=fast] when you want to approve the production plan and visual lock before generation.

🐍 One production, folded into a compact flow

flowchart TB
    subgraph R1["Plan and lock"]
        direction LR
        A["① 💡 Story / brief"] --> B["② 🧠 Resolve profile"] --> C["③ 👤 Character & product bible"] --> D["④ 🏙️ Scene masters"]
    end
    subgraph R2["Design and generate"]
        direction RL
        E["⑤ 🧩 Storyboard & keyframes"] --> F["⑥ 🎯 Route every shot"] --> G["⑦ 🎞️ H3 + ComfyUI"] --> H["⑧ ⭐ Select takes"]
    end
    subgraph R3["Finish and deliver"]
        direction LR
        I["⑨ ✂️ Edit, voice, captions & sound"] --> J["⑩ 🔍 Technical + visual QC"] --> K["⑪ 📦 Master + reproducible project"]
    end
    R1 --> R2 --> R3

Guided mode pauses after the production plan and again after the visual sources of truth. Fast mode records conservative assumptions and continues automatically.

🪄 Representative prompts

🎭 Produce a complete short drama

$minimax-h3-drama-producer
Turn this script and cast reference pack into a 60-second vertical short drama.
Keep the plot and dialogue unchanged. Design missing locations and shot coverage,
lock character continuity, then finish picture, dialogue, subtitles, music, and QC.

🧪 Distill a reusable house style

$minimax-h3-profile-distiller
Analyze these three local reference videos. Extract their pacing, shot grammar,
camera behavior, caption style, sound pattern, and QC rules into a reusable
profile. Exclude their people, branding, dialogue, plot, and music melody.

🎥 Design and run one H3 shot

$minimax-h3-adviser
Image 1 controls the actor's identity. Video 1 controls only the handheld camera
motion. Create an 8-second tense corridor reveal, then run it in ComfyUI.
[return=true] [load_workflow=true] [preview=true]

See skill configuration for every control flag, configuration file, model override, generation setting, and ready-to-copy combination.

🧰 Nine skills, one production system

	Skill	Role
🎬	<code>minimax-h3-drama-producer</code>	Produce or resume a complete multi-shot video project
🧪	<code>minimax-h3-profile-distiller</code>	Distill local reference videos into a safe, reusable production profile
🧭	<code>minimax-h3-adviser</code>	Choose the workflow, improve a prompt, or diagnose a failed result
✍️	<code>minimax-h3-text-to-video</code>	Invent a complete shot from language
🖼️	<code>minimax-h3-frame-to-video</code>	Animate an exact first frame or bridge exact first/last frames
🎛️	<code>minimax-h3-reference-to-video</code>	Bind identity, design, style, motion, camera, performance, or audio references
✂️	<code>minimax-h3-video-editor</code>	Make a precise change while stating what must remain unchanged
⚙️	<code>minimax-h3-comfyui</code>	Prepare, validate, submit, monitor, and fetch local H3 workflows
🪄	<code>qwen-image-edit</code>	Run the pinned one/two-reference Qwen consistency-edit workflow

The Qwen skill is explicit-only. Invoke $qwen-image-edit to run an edit, or $qwen-image-edit help to show its complete ComfyUI dependency and model guide. Its default path copies the prompt payload directly into the workflow. All boolean bracket controls share the aliases documented in skill configuration.

🔄 Compare and update installed skills

The project-level manager inventories deterministic skill hashes, compares the skills/ tree between Git refs, and previews plugin update commands before applying them:

python3 scripts/manage_skills.py list
python3 scripts/manage_skills.py diff v0.3.0 v0.4.0
python3 scripts/manage_skills.py update
python3 scripts/manage_skills.py update --apply

update is a dry run unless --apply is explicit. Start a new Codex task after updating so the refreshed plugin skills load.

🎨 Make visual direction concrete

The adviser can show visual atlases so “cinematic” becomes a controllable choice instead of a vague style word.

📐 Framing

🎥 Camera motion

💡 Lighting

The generated MiniMax H3 prompt remains text-only; the atlases are conversation aids.

🎚️ Built-in production profiles

Profiles are declarative data, never executable code:

base-video + one primary profile + explicit user overrides

Profile	Best for	Signature defaults
tiktok-short-drama	Conflict, emotion, reveal, payoff	9:16, hook-first pacing, shot keyframes, captions on
commercial-ad	Promise, proof, product hero, CTA	16:9 master, product geometry locks, deterministic brand text

Use minimax-h3-profile-distiller to learn a new production grammar from local references without copying their source-specific content.

📦 Reproducible output

outputs/<project>/
├── planning/ + profile/ + prompts/ + images/
├── workflows/ + clips/ + audio/ + subtitles/
├── edit/ + qc/ + final/ + logs/
└── project.yaml

Re-running the same project resumes completed work and versions new attempts instead of erasing history. A successful delivery includes the local master, media info, contact sheet, technical/visual QC report, timeline, and reproducible export instructions.

🔌 Notes

• Official pinned T2V, I2V, and R2V graphs are prepared deterministically; arbitrary custom workflow adaptation is intentionally out of scope.
• Voice and captions can be auto, on, or off. No paid voice service is called silently.
• FFmpeg is the v1 editing backend; selected clips, audio stems, captions, and timeline data remain available for manual NLE import.
• Prompt-only Open WebUI editions remain in <code>adapter/open-webui</code>. The producer and profile distiller are Codex-specific.

---

Bring the story. Let Codex build the production system around it. 🎬