Generated Computers
Locals: an Apple Silicon engine that runs local LLMs up to 3x faster. A product of Generated Computers.
Translation pending
自托管 AIGC 编排台:ComfyUI / Ollama / 语音 / OpenAI 兼容 /v1。不是 liberastra.com.br。
Translation pending
Developer toolsPython · TypeScriptGitHub @liberalchangOnlineOwner unconfirmed
Have you tried it? One tap is enough. No sign-in needed.
Owners of public repositories under a personal GitHub account can sign in to verify ownership and manage the introduction themselves. No administrator approval is needed.
Sign in, generate a verification file, and commit it to this repository’s default branch within 30 minutes. Files on other branches or forks do not count. Then check the file below to manage the project without administrator approval. Keep the file to retain your access.
If needed, you will go to sign-in. After signing in, return here and press your chosen verification button again.
Run /nomorevibe verify in your project folder to verify ownership and manage the introduction yourself.
Verify in one minuteAsk about other verification options
LiberAstra 编排台 — Self-hosted AIGC orchestration for ComfyUI, Ollama, and OpenAI-compatible APIs (image, video, speech). MIT.
Translation pending
Taken from the repository README
Loading the full README.
Read the full README ↗LiberAstra 编排台 — self-hosted AIGC orchestration
只编排、不推理:把 ComfyUI、Ollama 和语音接到同一个工作台(Web / CLI / MCP / OpenAI /v1)。
仓库:https://github.com/liberalchang/LiberAstra
GitHub Pages 主页:https://liberalchang.github.io/LiberAstra/
- 首页为产品落地页;文档骨架位于
/docs/(示例:https://liberalchang.github.io/LiberAstra/docs/)
搜索这个名字时,liberastra.com.br 是另一个项目。这里是开源的自托管 AIGC 编排台。
astra CLI、MCP、OpenAI 兼容 /v1。docker compose,或 Windows 安装包。需要自备 ComfyUI、Ollama,或兼容的云端 API。LiberAstra 不替代它们,也不内置模型。
在仓库根目录:
cp .env.example .env
docker compose up -d --build
浏览器打开 http://localhost:8000。首次创建管理员,再到「设置 → 服务提供方」填写各服务地址。
Windows:下载 LiberAstra-Setup.exe,装完打开 http://127.0.0.1:8000。本地语音和深度图在「资源 → 插件」里安装。
环境变量、分置和模型地址:文档索引。
功能演示(约 6 分钟):
https://github.com/user-attachments/assets/7d134651-f796-43b1-8c5a-3a038228ca9d
画布导演台演示:
https://github.com/user-attachments/assets/8334b7ba-5259-416b-a717-68c27e35fd7b
下面只区分职责,没有速度或画质对比。
| LiberAstra | ComfyUI | 典型聊天界面 | |
|---|---|---|---|
| 角色 | 编排工作台 | 推理 / 工作流引擎 | 对话 |
| 模态 | 图、视频流水线、语音 | 偏图和视频图 | 偏文本 |
| 接入 | Web + CLI + MCP + /v1 | 插件与自身 API | API / 插件 |
| 模型在哪跑 | 不跑,交给外部服务 | 在本机执行图 | 通常是云端或本地 LLM |
设计上参考过 Pixelle-Video、Dreamifly、Liber_SpeechAPI,边界见 参考项目。
要什么硬件? 编排进程本身不需要 GPU。出图、出视频需要一台能跑 ComfyUI 的机器,可以在局域网另一台。只用在线 API 和 edge-tts 时,可以没有本地显卡。视频拼接需要 ffmpeg:容器镜像里已有;Windows 安装包需要系统里能找到 ffmpeg。
模型文件放哪? 放进 ComfyUI 的 models\,或用 Ollama 拉取。不要放进 LiberAstra 的数据目录。
Docker 里访问本机的 ComfyUI / Ollama。 用 http://host.docker.internal:端口。localhost 指容器自己。另一台机器用那台机器的局域网 IP。
8000 端口被占用。 在 .env 里设 LIBERASTRA_PUBLISH_PORT=18000,然后打开 http://<主机>:18000。容器内仍监听 8000。
本地语音和深度图。 装好主程序后,在「资源 → 插件」里安装并启用 Liber Speech(端口 5555)或 Liber Depth(端口 5556)。Release 不再附带 LiberSpeech-Setup.exe。只做在线朗读时可以不装语音插件,用 edge-tts。
装完页面是空的。 先创建管理员,再添加服务提供方。没有 Provider 时,生成页没有可调用的后端。
| 文档 | 内容 |
|---|---|
| 文档索引 | 使用、部署、CLI、API |
| 部署(英文) | Docker、Windows、语音插件、模型放哪 |
| 在 Claude / Cursor 里用 | Skill 与 MCP |
| 参与贡献 | 怎么跑前端和后端、怎么提 PR |
平台本身不做模型推理。界面按侧栏分组。各页和画布功能节点的提示词共用 PromptField(原始稿 / 优化稿、场景预设、提交优化任务,刷新后仍可接着看)。功能模板与场景预设在「设置 → 提示词模板」里改,保存后下一次优化即生效。
总览:按图像 / 视频 / 语音切换,写一句提示词即可进入对应功能。下方是图像、视频、语音和工具四组常用入口。
侧栏「画布」。在无限画布上把输入、工作流包、语音和拼接工具连成流程图,按节点执行。
{{变量名}} 引用,/ 可插入已连接节点;媒体标签如 <Picture 1> 原样交给工作流。图片输入可标记、涂蒙版、裁剪,并保留版本。t2i 工作流包。内置 Qwen-Image 2.1(官方 / 越狱)与传统 checkpoint 包。参数区可换 UNET、文本编码器、VAE;越狱包可开关 Uncensored LoRA。也支持在线图像 API。`<image1>` / `<image2>` 把多人放进同一场景;局部重绘是原图加蒙版。官方包与越狱包成对。另有 PoseStudio 姿势换姿/姿势生图、Union 深度控制、参考图调色和换衣包。upscale_basic 走 RealESRGAN 保真放大;seedvr2_3b_image_restore 做 SeedVR2 生成式修复。i2v)、配音与 BGM。i2v。内置 MiniMax H3 官方、越狱、越狱+提示词增强;另有 i2v_ref 多参考、minimax_h3_orbit 360° 环绕和 minimax_h3_extender 长视频包,可开 Turbo。static_ / image_ / video_)。tts 工作流。astra CLI、MCP Server、OpenAI 兼容 /v1(chat / audio / images)。单用户,API Token;Provider 密钥加密保存。两种方式都只安装 LiberAstra 编排程序。Ollama、ComfyUI、语音插件是独立组件,装在本机或另一台主机均可。
更细的变量说明见 docs/deployment.md,分置示例见 docs/分置部署教程.md。安装包、Ollama、ComfyUI 和各工作流模型的下载地址见 部署相关资源下载。英文部署见 docs/deployment.en.md。
只起 LiberAstra 时,Ollama / ComfyUI / 语音仍用已经装好的地址。需要把可选组件也放进同一套 Compose 时:
docker compose --profile ollama up -d # 额外起 Ollama,宿主机 11434
docker compose --profile comfyui up -d # 额外起 ComfyUI,宿主机 8188(需 NVIDIA Container Toolkit)
docker compose --profile speech up -d # 额外起语音插件,宿主机 5555(需 NVIDIA)
docker compose --profile postgres up -d # 额外起 PostgreSQL;同时设置 LIBERASTRA_DATABASE_URL
语音插件也可以只拷 services/speech 到显卡主机,在该目录执行 docker compose up -d --build。LiberAstra 里 Base URL 填 http://<那台主机>:5555。
常用参数(写在 .env,环境变量优先于该文件):
| 参数 | 默认 | 作用 |
|---|---|---|
LIBERASTRA_PUBLISH_PORT | 8000 | 宿主机对外端口。8000 被占用时改为 18000 等,容器内仍听 8000 |
LIBERASTRA_HOST / LIBERASTRA_PORT | 0.0.0.0 / 8000 | 进程在容器内的监听地址,一般不用改 |
LIBERASTRA_DATA_DIR | ./data(容器内为 /data) | 数据库、作品、上传、用户工作流、日志 |
LIBERASTRA_SECRET_KEY | 空则写入 data/.secret_key | JWT 与 Provider 密钥加密。换机器要连这个文件一起拷 |
LIBERASTRA_DATABASE_URL | 空则 SQLite data/liberastra.db | 改用 PostgreSQL 时填写,并启用 postgres profile |
LIBER_SPEECH_DEVICE | auto | 语音插件设备:auto / cuda / cpu |
LIBER_SPEECH_API_KEY | 空 | 语音插件鉴权。LiberAstra 的 Provider 密钥填同一个值 |
LIBER_SPEECH_TTS_MODEL | multilingual | 语音插件默认 TTS:turbo / standard / multilingual |
LIBER_SPEECH_ASR_MODEL | large-v3 | 语音插件默认识别模型 |
从 Releases 下载 LiberAstra-Setup.exe。安装包自带 Python 运行时和已构建的页面,目标电脑不用另装 Python 或 Node。数据在 %LOCALAPPDATA%\LiberAstra\data。安装程序会放行本机防火墙的 TCP 8000。
仓库里不提交安装包本体。需要自行重新打包时:
powershell -ExecutionPolicy Bypass -File packaging\windows\build.ps1
产物:packaging\windows\output\LiberAstra-Setup.exe。构建机需要 Node.js 20+、外网,以及 Inno Setup 6。视频成片还需要系统里能找到 ffmpeg(winget install Gyan.FFmpeg)。
Ollama 安装包:https://ollama.com/download ,默认 http://<主机>:11434。示例:
ollama pull orcarouter/Qwen3.8-27B-Uncensored:q4_K_M
ollama pull huihui_ai/gemma-4-abliterated:12b
已有 GGUF 时,在文件旁写 Modelfile(一行 FROM ./模型文件.gguf),再执行 ollama create 本地模型名 -f Modelfile。Windows 路径用正斜杠。
ComfyUI 桌面安装包:https://www.comfy.org/download/ 。默认 http://<主机>:8188。权重放进 models\:
| 用途 | 下载 | 放到 |
|---|---|---|
| Qwen-Image 2.1 官方(UNET、8B 文本编码器、VAE) | https://huggingface.co/Comfy-Org/Qwen-Image-2.1 | diffusion_models/、text_encoders/、vae/ |
Qwen-Image 2.1 可选 UNET qwen-image-2.1-UC-int8_convrot.safetensors | https://huggingface.co/abenzerps/Qwen-Image-2.1-Uncensored-Quantized | diffusion_models/ |
| MiniMax H3 官方(主模型、文本编码器、视频/音频 VAE) | https://huggingface.co/Comfy-Org/MiniMax-H3 | diffusion_models/、text_encoders/、vae/ |
| H3 可选文本编码器(INT8,增强包用) | https://huggingface.co/ethanfel/Qwen3-VL-32B-Ultra-Heretic-MiniMax-H3-ComfyUI-INT8-ConvRot | text_encoders/ |
| H3 可选文本编码器(NVFP4) | https://huggingface.co/sakamakismile/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4 | text_encoders/ |
| H3 Turbo LoRA | https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora 、https://huggingface.co/lightx2v/Minimax-h3-Turbo | loras/ |
成套文件名见 视频指南。
语音与深度(可选)。 只做在线朗读时可以不装语音插件,用 edge-tts。需要本地合成、识别或深度图时,在「资源 → 插件」安装对应插件。插件和主程序在同一台 Windows 上时,语音地址是 http://127.0.0.1:5555,深度是 http://127.0.0.1:5556。
放到另一台显卡主机时,仍可单独跑 services/speech 或 services/depth 的 docker compose up -d --build,再把 Base URL 填成那台机器的局域网地址。不再发布 LiberSpeech-Setup.exe。
同一台机器的 Base URL 填 http://127.0.0.1:5555,别的主机填 http://<IP>:5555。
本地前后端、测试和 PR 约定见 CONTRIBUTING.md。
cd backend && uv venv .venv && uv pip install -e ".[dev]"
uv run astra serve --reload
cd frontend && npm install && npm run dev
CLI 示例:
astra tts "你好" --voice zh-CN-XiaoxiaoNeural -o hello.mp3
astra image gen --prompt "一只橘猫" -o ./out
astra video quick --topic "三个提高效率的小技巧" --voice zh-CN-XiaoxiaoNeural -o v.mp4
astra mcp
backend/ — FastAPI 后端、astra CLI、MCP、Alembic 迁移、内置工作流包(workflows/)frontend/ — React 19 + Vite + Tailwind v4 + i18nextdocker/、docker-compose.yml — 单镜像(ffmpeg + 中文字体)与可选 profilepackaging/windows/、packaging/speech/ — Windows 主程序与语音插件的构建脚本skills/liberastra/ — Agent Skilldocs/ — 使用与部署;研发笔记在 docs/internal/site/ — GitHub Pages 站点(产品落地页 + VitePress 文档输出).github/workflows/ci.yml — lint、测试、前端构建、镜像构建LiberAstra 以重新实现为主。协议与移植范围见 THIRD_PARTY_NOTICES.md。
| 项目 | 原 GitHub | 协议 | 本仓库如何参考 |
|---|---|---|---|
| Pixelle-Video | https://github.com/AIDC-AI/Pixelle-Video | Apache-2.0 | 视频流水线设计;分镜模板与素材分析工作流移植自其仓库 |
| Dreamifly | https://github.com/LastLighter/Dreamifly | MIT | 图像能力与 ComfyUI 工作流模板映射的设计参考 |
| Liber_SpeechAPI | https://github.com/liberalchang/Liber_SpeechAPI | AGPL-3.0 | 仅参考接口形态;本仓库 services/speech 为独立实现,不打包其源码 |
LiberAstra — self-hosted AIGC orchestration. It connects ComfyUI, Ollama, and speech services to one desk: Web UI, astra CLI, MCP, and an OpenAI-compatible /v1. The app does not run models. MIT.
Repository: https://github.com/liberalchang/LiberAstra
liberastra.com.br is a different project. This repository is the open-source orchestration desk.
/v1.You bring ComfyUI, Ollama, or a compatible API. LiberAstra does not replace them and does not bundle weights.
From the repository root:
cp .env.example .env
docker compose up -d --build
Open http://localhost:8000. Create the admin user, then Settings → Providers.
Windows: LiberAstra-Setup.exe, then http://127.0.0.1:8000. Install local speech or depth from Resources → Plugins.
Full steps, profiles, and model locations: docs/deployment.en.md. Index: docs/README.md.
Feature walkthrough, about 6 minutes:
https://github.com/user-attachments/assets/7d134651-f796-43b1-8c5a-3a038228ca9d
Canvas director demo:
https://github.com/user-attachments/assets/8334b7ba-5259-416b-a717-68c27e35fd7b
Roles only. There is no speed or quality benchmark.
| LiberAstra | ComfyUI | A typical chat UI | |
|---|---|---|---|
| Role | Orchestration desk | Inference / graph engine | Conversation |
| Modalities | Image, video pipeline, speech | Mostly image and video graphs | Mostly text |
| Access | Web + CLI + MCP + /v1 | Plugins and its own API | API / plugins |
| Where models run | Not here | On the ComfyUI host | Usually a cloud or local LLM |
Pixelle-Video, Dreamifly, and Liber_SpeechAPI are upstream references. See Reference projects.
Hardware. The orchestration process does not need a GPU. Image and video need a machine that can run ComfyUI, which can be another host on the LAN. Online APIs plus edge-tts need no local GPU. Video assembly needs ffmpeg (included in the Docker image; install it on Windows).
Where weights go. Into ComfyUI's models\ directory, or pull them with Ollama. Not into LiberAstra's data directory.
Docker reaching a service on the host. Use http://host.docker.internal:<port>. localhost inside the container is the container. Another machine is that machine's LAN IP.
Port 8000 is taken. Set LIBERASTRA_PUBLISH_PORT=18000 in .env and open http://<host>:18000.
Local speech and depth. After the app is installed, open Resources → Plugins and install Liber Speech (port 5555) or Liber Depth (port 5556). Releases no longer include LiberSpeech-Setup.exe. edge-tts covers online read-aloud without the speech plugin.
Empty generate pages. Create the admin account, then add a provider. Without one there is no backend to call.
Deployment · Docs index · Claude / Cursor · Contributing
Prompt fields share one control: raw and optimized drafts, scene presets, and an optimize action that survives refresh. Templates live under Settings → Prompt Templates.
Home. Pick image, video, or speech and start from one line.
Canvas. Open Canvas in the sidebar. Connect inputs, workflow packs, speech, and concat, then run the graph. Groups are maintained under Settings → Workflow Packs. Unchanged upstream nodes are reused. Save a preview into the asset library when you want to keep it. Text references use {{variable}}; media tags such as <Picture 1> pass through to the workflow.
Images. Text-to-image, including Qwen-Image 2.1 packs and classic checkpoints, plus online image APIs. Edit: full frame, up to four references, or inpaint with a mask. Upscale uses an upscale model.
Video. Quick create turns a topic or a fixed script into a short: storyboard, per-shot stills or text-to-video, TTS, subtitles, and BGM. The LLM visual-prompt switch and a pack's Guide enhancer are independent. Also: custom assets, talking avatar (bring your own avatar pack), image-to-video on MiniMax H3 with an optional Turbo LoRA, motion transfer, and storyboard templates (static_ / image_ / video_).
Speech. edge-tts, OpenAI-compatible TTS, Liber Speech, or a Comfy TTS pack. Speech-to-text exports txt, srt, vtt, or json. Voice library and a music library for BGM.
Platform. Asset library, job queue with SSE, cancel and retry. Single user. Provider keys are encrypted. Clients: Web UI, astra CLI, MCP, OpenAI-compatible /v1 (chat, audio, images).
The long pairing manual is Chinese: image guide and video guide. The Web UI itself is bilingual.
Re-implementation first. Licenses and what was ported: THIRD_PARTY_NOTICES.md.
| Project | Original GitHub | License | How we use it |
|---|---|---|---|
| Pixelle-Video | https://github.com/AIDC-AI/Pixelle-Video | Apache-2.0 | Video pipeline design; storyboard templates and analysis workflows ported |
| Dreamifly | https://github.com/LastLighter/Dreamifly | MIT | Image UX and ComfyUI workflow-mapping design reference |
| Liber_SpeechAPI | https://github.com/liberalchang/Liber_SpeechAPI | AGPL-3.0 | API shape only; services/speech is an independent implementation |