🎬 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 |

Codex sends a reference-led request to local ComfyUI and returns the generated video

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: ```bash 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: ```bash 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](https://github.com/Larryvrh/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](https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora) 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 ```text $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 ```mermaid 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

```text $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

```text $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

```text $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](docs/skill-config.md) for every control flag, configuration file, model override, generation setting, and ready-to-copy combination. ## 🧰 Nine skills, one production system | | Skill | Role | |---:|---|---| | 🎬 | [`minimax-h3-drama-producer`](skills/minimax-h3-drama-producer/SKILL.md) | Produce or resume a complete multi-shot video project | | 🧪 | [`minimax-h3-profile-distiller`](skills/minimax-h3-profile-distiller/SKILL.md) | Distill local reference videos into a safe, reusable production profile | | 🧭 | [`minimax-h3-adviser`](skills/minimax-h3-adviser/SKILL.md) | Choose the workflow, improve a prompt, or diagnose a failed result | | ✍️ | [`minimax-h3-text-to-video`](skills/minimax-h3-text-to-video/SKILL.md) | Invent a complete shot from language | | 🖼️ | [`minimax-h3-frame-to-video`](skills/minimax-h3-frame-to-video/SKILL.md) | Animate an exact first frame or bridge exact first/last frames | | 🎛️ | [`minimax-h3-reference-to-video`](skills/minimax-h3-reference-to-video/SKILL.md) | Bind identity, design, style, motion, camera, performance, or audio references | | ✂️ | [`minimax-h3-video-editor`](skills/minimax-h3-video-editor/SKILL.md) | Make a precise change while stating what must remain unchanged | | ⚙️ | [`minimax-h3-comfyui`](skills/minimax-h3-comfyui/SKILL.md) | Prepare, validate, submit, monitor, and fetch local H3 workflows | | 🪄 | [`qwen-image-edit`](skills/qwen-image-edit/SKILL.md) | 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](skills/qwen-image-edit/references/comfyui-workflow-install.md). Its default path copies the prompt payload directly into the workflow. All boolean bracket controls share the aliases documented in [skill configuration](docs/skill-config.md). ## 🔄 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: ```bash 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
Framing atlas

🎥 Camera motion
Camera motion atlas

💡 Lighting
Lighting atlas

The generated MiniMax H3 prompt remains text-only; the atlases are conversation aids. ## 🎚️ Built-in production profiles Profiles are declarative data, never executable code: ```text 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 ```text outputs// ├── 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 [`adapter/open-webui`](adapter/open-webui/). The producer and profile distiller are Codex-specific. ---

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