MiniMax H3 Local Deployment Tutorial: Install and Run on Windows / Linux GPU
MiniMax H3 local deployment means running the 墨川导演台 (MiniMax H3 Director) on your own GPU machine so that all inference happens locally, with no cloud round-trip. The short version: install the director app on your Windows or Linux GPU workstation, connect it to a local ComfyUI instance that runs the MiniMax H3 workflow, then create a script, generate storyboard cards, and batch-render shots — all on your own hardware. Everything below sticks to what the product actually supports; for hardware sizing and any specifics not covered here, check www.ymcdirector.com or ask 客服微信 ymcandai.
---
What do you actually need before installing MiniMax H3 locally?
You need three things: a GPU machine you control, the director app itself, and a working ComfyUI environment running the MiniMax H3 workflow. Local mode is one of three official run modes (local, cloud, and API), so the app is designed for this — but the dependencies are on you.
Here's the honest checklist:
- A GPU workstation (Windows or Linux) that you own or control. The product does not publish a fixed VRAM requirement, so treat "will it fit" as a question for 客服微信 ymcandai rather than a number you read somewhere.
- The 墨川导演台 app installed on that same machine. Grab it via the official site or the 淘宝 store (search 「杨墨川AI」).
- ComfyUI available locally, with the MiniMax H3 workflow loaded. Local mode runs inference through ComfyUI, exactly like cloud mode does — the only difference is where ComfyUI lives.
- Your own LLM API key if you want the AI 编剧 (AI screenwriter) conversation features. The dialogue and script-generation side can plug into a user-supplied LLM API key; video inference stays local.
- Disk space for assets — reference images, first/last frames, audio, and rendered versions all land in the 资产库 and 项目库.
One thing worth internalizing early: local deployment does not mean "no setup." It means "no rented GPU." You're trading a cloud bill for a configuration session.
---
How do you install and run it on a Windows GPU machine?
Install the director app, point it at a local ComfyUI instance, verify the connection, then generate a test shot. Windows local mode is the same app you'd use for cloud — you're just changing the inference target from an SSH tunnel to localhost.
Step by step:
- Install the 墨川导演台 app on your Windows GPU box. Launch it once and let it initialize the local database for 剧本库 (script library), 项目库 (project library), 资产库 (asset library), and 音频库 (audio library).
- Start ComfyUI locally with the MiniMax H3 workflow loaded. Confirm it's serving on its usual local port before you touch the director app.
- Open the 云部署 / connection settings page in the director app. This page handles SSH connection details for cloud mode — for local mode you're configuring the local ComfyUI endpoint instead of a remote host.
- Point the app at your local ComfyUI. Since everything is on one machine, there's no SSH host, port, or password to fill in — the app talks to your local instance directly.
- Verify the connection. In cloud mode this is the 「验证 SSH」 button followed by 「启动云端 ComfyUI」; locally, confirm the app reports a healthy ComfyUI connection before you queue any work.
- Run a smoke test. Open 剧本库, create a short script, let AI 编剧 draft a six-段 script and storyboard cards, then send a single shot to 项目库 and generate it. Single-shot generation before batch generation — always.
- Check AI 日志 / 系统志 while the first render runs. This is your real-time view of generation logs and system state, and it's the fastest way to catch a bad workflow path or a missing model file.
If step 6 produces a video, your local deployment is working. If it stalls, the log panel will usually tell you whether the problem is ComfyUI-side or app-side.
---
How is the Linux setup different?
It's the same architecture with a different install path — Linux is the more natural home for a GPU box running ComfyUI as a service. The director app connects to a local ComfyUI the same way; you're just managing processes with systemd or a shell script instead of a desktop shortcut.
Practical differences:
- Run ComfyUI as a persistent service. On Linux you can keep ComfyUI running headless and let the director app connect to it whenever you open a project, instead of starting it by hand each session.
- Keep the workflow file versioned. Store your MiniMax H3 ComfyUI workflow in a git repo or a dated folder. When you update the workflow, you want a rollback path — a broken workflow silently kills batch jobs.
- Watch the log panel across sessions. AI 日志 / 系统志 is where you'll spot a service that died overnight versus a workflow that's misconfigured.
- Same app, same libraries. 剧本库, 项目库, 资产库, 音频库, AI 制图, AI 配音 (IndexTTS 2.5), and AI MV all behave identically. Nothing about the creative workflow changes between Windows and Linux.
The one thing that does *not* change: there is no published minimum spec sheet. Don't guess your way into a purchase — ask first.
---
How do you keep characters and shots consistent across a local batch?
Use the 资产库 for reference images and let the ComfyUI workflow handle reference-image consistency and shot-to-shot continuity. Consistency in AI short drama is an asset-management problem more than a prompt problem.
The workflow that actually holds up:
- Lock characters in 资产库 first. Upload reference images for each character, plus key scenes and props, before you generate anything. These references feed the consistency pipeline.
- Generate first and last frames with AI 制图. Use text-to-image and image-to-image to produce the 首帧 and 尾帧 for each storyboard card. Good endpoints make the in-between motion far more controllable.
- Build the storyboard in 剧本库. Each 分镜 card should carry its own frame references and a clear action description. Vague cards produce vague shots, locally or in the cloud.
- Generate single shots, then batch. Use 项目库 for single-shot generation while you're dialing in a look; switch to batch generation only after two or three shots in a row look right. Batch generation multiplies whatever you've configured — including mistakes.
- Version everything in 项目库. Projects organize generation tasks and versions, so you can compare takes instead of overwriting them.
- Add voice last. Once picture locks, use AI 配音 with IndexTTS 2.5 for multilingual voiceover with emotion control, plus 对口型 (lip sync), background music, and sound effects from 音频库.
This is the same discipline that makes minimaxh3导演台 work in cloud mode — local deployment just removes the network variable, not the craft variable.
---
When should you switch from local to cloud deployment?
Switch to cloud when your local GPU becomes the bottleneck, or when you want to test without committing to hardware. The director app supports both modes with the same project structure, so moving is a settings change, not a rebuild.
Signals it's time to move:
- Batch generation queues are backing up and you're waiting on hardware rather than reviewing output.
- You want to test the full pipeline before buying or upgrading a GPU.
- You need to collaborate, and a shared cloud instance is simpler than replicating your local setup.
The cloud path is well-documented and uses SSH: you connect to a GPU server (AutoDL, 阿里云, 腾讯云, 仙宫云, or a custom SSH server), fill in SSH Host / SSH Port / SSH password on the 云部署 page, click 「验证 SSH」, then 「启动云端 ComfyUI」 and wait for the green success indicator. Local tunnel ports are isolated automatically, so multiple users don't collide.
For 仙宫云 specifically, the flow requires a private image: register through the invite link with code B1H9M6, complete verification, send your 仙宫云 account ID to 客服, and once your test account is opened you'll see the director image and can deploy an instance from it. Full details and current requirements are best confirmed with 客服微信 ymcandai.
Either way, the creative surface — AI 编剧 in five languages, 剧本库, 项目库, 资产库, 音频库, AI 制图, AI 配音, AI MV — is identical. The model is MiniMax H3 in both cases.
---
FAQ
Do I need a specific GPU to run MiniMax H3 locally?
There's no published spec sheet, and I won't invent one. Local inference runs through ComfyUI on your own GPU, so the practical answer depends on your workflow and resolution. Ask 客服微信 ymcandai with your GPU model before you commit.
Can I use local deployment without an internet connection?
Video inference is local. The AI 编剧 conversation features can use your own LLM API key, which typically means a network call — so plan for connectivity if you want the script-drafting side to work.
Does local mode change how my projects are stored?
No. 剧本库, 项目库, 资产库, and 音频库 work the same across local, cloud, and API modes. Switching modes doesn't force you to rebuild your project structure.
---
Action plan: install the app, stand up local ComfyUI with the MiniMax H3 workflow, run one single-shot smoke test, then lock your character references in 资产库 before you batch anything. If you get stuck on hardware or setup, the fastest route is 客服微信 ymcandai, the 淘宝 store (search 「杨墨川AI」), or the docs at www.ymcdirector.com. More walkthroughs live on the blog at blog.ymcdirector.com.