binwu1/drama-series-agent

选用指南

AI 根据 README 整理 · 2026/10/14 14:01:08

适用场景

适用于从一句话创意生成连续短剧剧集，涵盖剧情开发、剧本编写、角色设定和视频渲染的完整流程。

查看 README 依据

# Drama Series Agent [](LICENSE) [](https://www.python.org/) [](https://github.com/comfyanonymous/ComfyUI) **English** · [中文](README.zh-CN.md) > From a one-line idea to a chained vertical short-drama episode — > **develop → write → cast → Ref2VA render** — in one series-scoped agent. https://github.com/user-attachments/assets/52c170e2-bc91-41d2-8b3d-3f2de95c023c

In one series chat you can walk through **series settings → episode script → cast portraits → video render**, all under the same `series_id`, then continue into the next episode. --- ## Why this project? | Pain | What we ship | |------|----------------| | Ideas die in notes | [`skills/`](skills/) for develop, write, and H3 wiring | | Scripts that cannot be shot | Auto-built shot tables for ComfyUI (with cast / voice refs) | | One-off clips, no shot continuity | **Prev-shot tail → next first frame** — auto chain, **no manual editing** |

使用方式

提供Web工作台和命令行界面，通过聊天交互控制创作流程，也可使用CLI渲染单集。

查看 README 依据

- **Where LLMs come from** — presets: **ModelScope (friendly free tier)**, Qwen (DashScope), OpenAI, Claude, DeepSeek, Moonshot, local Ollama; any OpenAI-compatible API works - **Skill packs** — develop / write / H3 prompt skills under [`skills/`](skills/), editor-agnostic - **Local render** — MiniMax H3 workflows in [`workflows/selfhost/`](workflows/selfhost/); you bring ComfyUI + models - **Cast library** — `data/cast/{series}/` portraits, optional voice refs - **Web UI / CLI** — chat and jobs in the browser; or `drama-series-render` for a headless episode <p align="center">

| Surface | URL | Notes | |---------|-----|--------| | **Web workbench** | **http://localhost:5173/** | Default Vite port (`web/`). Chat, series settings, scripts, cast, jobs, system config. | | API (OpenAPI) | http://127.0.0.1:8000/docs | Backend the UI calls | | API base used by UI | `http://127.0.0.1:8000/api/series` | Override with `web/.env`: `VITE_HERMES_API_BASE=...` | Open **http://localhost:5173/** after both processes are up. If the page loads but chat/settings fail, the API is not running on port **8000** (start it first, or point `VITE_HERMES_API_BASE` at your API).

Without the helper scripts: ```bash # terminal 1 — API python scripts/run_api.py --host 127.0.0.1 --port 8000 # terminal 2 — workbench (http://localhost:5173/) cd web npm install npm run dev ``` ### Render one episode (CLI) 1. Cast: `data/cast/{series}/{character}.png` (+ optional `voices/{character}.wav`) 2. Run files: `templates/{series}/run/EP001.episode-run.jsonl` + `EP001.episode-meta.json` 3. ComfyUI at `COMFYUI_URL` (default `http://127.0.0.1:8188`) ```bash drama-series-render \ --project templates/<series> \ --episode EP001 \ --cast-series <series> \ --context-ir off \

部署要求

需要Python 3.10+，ffmpeg，以及自行安装的ComfyUI和MiniMax H3模型。可选Node 18+用于Web工作台，可选OpenAI兼容LLM用于智能体对话。

查看 README 依据

<img src="docs/assets/main.png" alt="Hermes workbench main UI" width="920" /> </p> <p align="center"><em>Workbench: chat on the left, per-episode progress and render jobs on the right</em></p> --- ## Quick start ### Requirements - Python **3.10+** - **ffmpeg** (see [Install ffmpeg](#install-ffmpeg) below) - Optional: Node **18+** for the React workbench - Optional: OpenAI-compatible LLM for the agent chat (e.g. ModelScope) ### ComfyUI (required for video render — install yourself) This repo does **not** bundle ComfyUI or MiniMax H3 weights. You must:

1. **Install [ComfyUI](https://github.com/comfyanonymous/ComfyUI) yourself** and keep it running (default `http://127.0.0.1:8188`). 2. Install the **MiniMax H3 Ref2VA custom nodes** your graphs need (e.g. `MiniMaxH3ReferenceToVideo` and related loaders / optional SageAttention · BlockCache nodes used by the `*_fast` / `*_turbo` tiers).

3. **Download every model file referenced by the workflow JSON** into the matching ComfyUI folders (`models/diffusion_models`, `models/text_encoders`, `models/vae`, `models/loras`, … — follow ComfyUI’s usual layout). Typical names from [`workflows/selfhost/`](workflows/selfhost/): | File (examples) | Used by | |-----------------|---------| | `minimax_h3_ref2va_pruned_int8_convrot.safetensors` | UNET / Ref2VA (fast · lora · turbo) | | `minimax_h3_fl2va_pruned_int8_convrot.safetensors` | Quality tier `video_minimax_h3_r2v.json` |

已知限制

未安装ComfyUI和模型时，视频渲染功能不可用，但其他功能（如设置开发、剧本编写、角色管理）仍可使用。

查看 README 依据

Without a ready ComfyUI + models, you can still develop settings, write scripts, and manage cast in the workbench — **video jobs will fail** until the above is done. ### Install ffmpeg ffmpeg is required after each shot to extract the **tail frame** (chain continuity) and to concat shots into `master.mp4`. Resolution order in code: `FFMPEG_PATH` → `PATH` → bundled **`imageio-ffmpeg`** (pulled in by `pip install`). **Recommended: install a system binary** (clearer for debugging): <details> <summary><strong>Windows</strong></summary> ```powershell # Option A — winget