MiniMax H3 Cloud Deployment Tutorial: SSH to a GPU Server and Run ComfyUI

2026-11-01 · English · https://blog.ymcdirector.com

MiniMax H3 Cloud Deployment Tutorial: SSH to a GPU Server and Run ComfyUI

You can run MiniMax H3 on a rented GPU cloud server by registering with the right invite code, deploying the private 墨川导演台 image, then pasting the SSH host, port, and password from your cloud instance into the 导演台's Cloud Deployment page. Once "Verify SSH" passes and "Start Cloud ComfyUI" turns green, you're ready to write a script and generate video—no local GPU required.

This guide walks through the exact flow using 仙宫云 (Xiangongyun) as the test platform, plus what to check when things don't connect. Everything here reflects the real 墨川导演台 (MiniMax H3 AI Director) workflow—for anything not covered, confirm with 官网 www.ymcdirector.com or 客服微信 ymcandai.

---

Why run minimaxh3导演台 on a cloud GPU server instead of locally?

Because cloud deployment moves all the heavy inference off your own machine, so you can generate on a laptop while a rented GPU does the work. The 导演台 supports three real run modes:

  • Local deployment — inference runs on your own GPU computer.
  • Cloud deployment — the app connects to a GPU server you rent, and inference travels over an SSH tunnel.
  • API calls — you access 导演台 capabilities through an interface.

Cloud mode is the practical pick if your local machine is modest, if you want to keep a project rendering while you edit elsewhere, or if you're collaborating and need a shared, always-on backend. The 导演台 uses ComfyUI workflows for cloud inference, which is what gives you reference-image consistency, shot-to-shot continuity, and batch queuing. Your GPU rental is billed by your cloud provider—pricing is not something the 导演台 sets, so check with the provider or 客服微信 ymcandai for current guidance.

One detail worth knowing up front: the 导演台 isolates local tunnel ports automatically, so multiple users on the same machine don't collide. That matters if you're running a small team off one workstation.

---

How do I deploy the 导演台 image on 仙宫云 step by step?

Follow the registration and invite-code flow exactly—the private image is not public, so a fresh registration with the correct code is required. Here's the sequence:

  • Register on 仙宫云 at https://www.xiangongyun.com/register/B1H9M6. You must register anew and enter invite code B1H9M6.
  • Complete real-name verification on the platform.
  • Send your 仙宫云 account ID to customer service — it looks like ac401472-db02-5df3-xxxx-xxxxxxxxxxx. This is what unlocks the private image for your account.
  • Wait for the test account to be activated. Once it is, the 导演台 image appears in your image list.
  • Deploy an instance using that image.
  • Open the instance's SSH panel (the SSH button on the right side of the running instance) and copy the host, port, and password.

That's the cloud side done. Don't skip step 3—without your account ID on file, the image won't show up and you'll assume something is broken when it's simply not provisioned yet.

---

How do I connect the 导演台 to my GPU server over SSH?

Open the Cloud Deployment page in 导演台 and paste the three values you copied from 仙宫云, then verify and launch. Concretely:

  • Enter the SSH Host exactly as shown (an IP or hostname).
  • Enter the SSH Port — cloud providers rarely use 22, so copy it, don't guess.
  • Enter the SSH Password from the instance panel.
  • Click Verify SSH. A pass means the tunnel handshake succeeded.
  • Click Start Cloud ComfyUI and wait for the green success indicator.

Once green, the ComfyUI backend is live behind the tunnel and the 导演台 is talking to it. From here the rest of the app behaves normally: AI 编剧 writes a six-段 script and storyboard draft into your 剧本库, AI 制图 produces first/last frames and character reference images, and 项目库 queues single-shot or batch video generation against the MiniMax H3 model.

A few practical notes on the connection:

  • Copy, don't retype. Passwords from cloud panels often contain characters that are easy to mistype.
  • Verify before you launch. If Verify SSH fails, Start Cloud ComfyUI will not help—fix the credentials first.
  • Watch the AI 日志 / 系统志 panel. It shows live generation logs and system status, which is the fastest way to tell whether a stall is a connection problem or a queue problem.

---

What should I check when SSH verification or ComfyUI startup fails?

Most failures come down to three things: wrong credentials, an instance that isn't actually running, or an image that was never unlocked for your account. Work through this checklist:

  • Is the instance running? A stopped instance gives you stale or dead SSH details.
  • Did you copy host, port, and password from the current session? Restarting an instance can change them.
  • Was your account ID submitted and the test account activated? If the image never appeared, this is almost always why.
  • Did you register with invite code B1H9M6? The image is private; the code is the gate.
  • Is Verify SSH passing but ComfyUI not going green? Give it a moment, then check AI 日志 / 系统志 for the actual error rather than retrying blindly.

If you've confirmed all of the above and it still won't connect, send the details to 客服微信 ymcandai rather than troubleshooting in the dark.

---

How do I go from a green light to an actual generated video?

Write a script in AI 编剧, let it land in your 剧本库 as storyboard cards, then generate from 项目库—the cloud backend handles inference. The workflow that works well:

  • Draft in AI 编剧. Describe your concept in conversation; the AI replies and writes a six-段 script plus storyboard draft straight into the 剧本库. It handles Chinese, English, Spanish, Arabic, and Japanese.
  • Lock your assets. Upload character, scene, and prop reference images into 资产库—this is what keeps faces and locations consistent across shots.
  • Generate frames in AI 制图. Produce first frames, last frames, and character reference images via text-to-image or image-to-image.
  • Add audio. Use 音频库 for voice cloning, lip-sync, background music, and sound effects, and AI 配音 (IndexTTS 2.5) for multilingual lines with emotion control.
  • Queue the video. In 项目库, run a single shot to test, then batch-generate the rest. Versions are tracked per project.
  • Review in AI 日志 / 系统志 if anything stalls.

Test with one shot before committing to a batch—it's the cheapest way to confirm your reference images and prompts are behaving before you queue twenty clips.

---

FAQ

Do I need my own GPU to use MiniMax H3 cloud deployment?

No. Cloud deployment runs inference on the GPU server you rent, connected over SSH. Local deployment is the option that uses your own hardware.

Can I use a different cloud provider?

The 导演台 supports 通用云端 GPU deployment across AutoDL, 阿里云, 腾讯云, 仙宫云, and custom SSH servers. The 仙宫云 flow above is the documented test path with the private image; for other providers, confirm specifics with 客服微信 ymcandai.

What if I get stuck mid-setup?

Reach 客服微信 ymcandai, browse the 淘宝 store by searching 「杨墨川AI」, or check 官网 www.ymcdirector.com and the blog at blog.ymcdirector.com. Email works too: info@ymcdirector.com.

---

Bottom line: register with invite code B1H9M6, get your account ID to support, deploy the image, paste SSH details into the 导演台, verify, launch, and generate. The whole loop is short once the image is unlocked—most delays happen before you ever touch the SSH fields.

MiniMax H3 Director · Official Website →

WeChat QR

WeChat: ymcandai · Taobao: 杨墨川AI

Articles