MiniMax H3 Cloud Deployment: SSH Into a GPU Server and Run ComfyUI Workflows

2026-10-29 · English · https://blog.ymcdirector.com

Your laptop fan is spinning like a jet engine, your render queue is stuck at 12%, and somewhere a client is asking why the 4K shot still looks like it was filmed through a fogged bathroom mirror. If that sounds familiar, you're exactly the person who should stop fighting local hardware and let a rented GPU do the heavy lifting. 🎬

This guide walks through MiniMax H3 cloud deployment inside 墨川导演台 (MiniMax H3 AI Director), from connecting a cloud GPU box over SSH to launching ComfyUI workflows and generating your first shot. No hand-waving — just the actual steps, plus the small details that trip people up.

Why Bother With Cloud Deploy at All?

Think of the director's console as a control room and the GPU as the engine room. Locally, both live in your apartment — convenient, but the engine room is small and gets hot. With cloud deploy, you keep the control room on your desk and rent an engine room by the hour somewhere with better cooling.

墨川导演台 supports three real operating modes:

  • Local deployment — inference runs on your own GPU machine.
  • Cloud deployment — the console connects to a GPU server you rent yourself, and inference happens there over an SSH tunnel.
  • API calls — you wire the console's capabilities into your own pipeline through an interface.

The cloud path is the sweet spot for creators who want MiniMax H3 output quality without buying a workstation. The console talks to a cloud ComfyUI instance over SSH; the tunnel handles the connection, and each user's port is isolated automatically so multiple people on the same setup don't collide. That last part matters more than it sounds — shared GPU boxes are a classic source of "why is my job rendering someone else's character?" chaos.

One honest note upfront: this article won't quote VRAM requirements, generation times, or pricing, because those depend on the machine you rent and the shot you're making. Check the official site or ask 客服微信 ymcandai for current specifics.

The Actual Setup: From Registration to Green Light

The tested flow uses 仙宫云 (Xiangongyun) as the cloud provider. Here's the sequence, in order.

Step 1 — Register with the invite code. Go to the 仙宫云 registration page at https://www.xiangongyun.com/register/B1H9M6 and create a new account, making sure to enter invite code B1H9M6. Existing accounts won't work for this — you need a fresh registration.

Step 2 — Complete verification and send your account ID. After registration and identity verification succeed, send your 仙宫云 account ID to customer service. The ID looks like a UUID, for example ac401472-db02-5df3-xxxx-xxxxxxxxxxx. This is what gets your test account provisioned.

Step 3 — Find the director's console image. Once your test account is activated, the 墨川导演台 image becomes visible to you. It's a private image, not a public one — that's why the account-ID step exists. Deploy an instance using that image.

Step 4 — Grab your SSH credentials. When the instance is running, click SSH on the right side of the instance panel. You'll get three things: the connection host, the port, and the password. Copy all three.

Step 5 — Fill in the 云部署 page. In the director's console, open the 云部署 (Cloud Deployment) page and enter the values into SSH Host, SSH Port, and SSH Password.

Step 6 — Verify and launch. Click 验证 SSH (Verify SSH). Once it passes, click 启动云端 ComfyUI (Start Cloud ComfyUI) and wait for the green success indicator. Don't skip ahead while it's still starting — the tunnel needs to be live before anything downstream will work.

Step 7 — Write and generate. Now go write a script and test a video generation. That's the whole loop.

A useful analogy: the SSH credentials are like the address, door code, and key to a studio you've rented. The console doesn't need to know anything about the building — it just needs those three pieces to walk in and start shooting.

Practical tips for this stage:

  • Copy the port as a number, not as part of a pasted string. Stray spaces are the single most common cause of a failed SSH verification.
  • If verification fails, re-check the instance is actually running before touching anything else. A stopped instance looks identical to a wrong password from the console's point of view.
  • The tunnel isolates ports per user automatically, so you generally don't need to hand-configure port forwarding. If something feels off, restart the cloud ComfyUI step rather than fiddling with SSH settings.

What You Get Once ComfyUI Is Running

With the green light on, the console's cloud inference runs ComfyUI workflows on the remote GPU. In practice that means:

  • Reference-image consistency — upload character and scene references in the asset library so faces and locations stay stable across shots.
  • Shot-to-shot continuity — storyboard cards in the script library feed generation in sequence, which is what keeps a short drama from looking like a slideshow of strangers.
  • Batch queuing — queue multiple shots and let the remote machine chew through them while you work on the next scene.

The rest of the console's toolkit sits on top of that engine: AI screenwriting that drafts six-part scripts and storyboard cards in Chinese, English, Spanish, Arabic, or Japanese; AI image generation for first frames, last frames, and character references; IndexTTS 2.5 voiceover with emotion control; voice cloning, lip sync, music and sound effects in the audio library; and AI MV generation. The AI log / system monitor lets you watch generation logs and system status in real time — genuinely useful when a remote job stalls and you need to know whether it's the model or the network.

Two practical notes on workflow design:

  • Lock your references before batching. Generate and approve character reference images first. Changing them mid-batch wastes remote compute and produces inconsistent episodes.
  • Use single-shot generation to test, batch generation to produce. Iterate on one shot until the prompt and references are right, then queue the full sequence. This is the cheapest way to work on rented hardware.

If you're wiring the console into an existing pipeline instead, the API mode covers that. And for the AI screenwriting and dialogue features, you can plug in your own LLM API key.

Where to Get Help

Cloud deployment is the one part of this toolchain where a single wrong character in a password field can cost you an afternoon. If you hit a wall, the fastest routes are 客服微信 ymcandai, the Taobao store (search 「杨墨川AI」), or the official site at https://www.ymcdirector.com. You can also find 杨墨川 on Douyin, Xiaohongshu, and WeChat Channels, and reach the team at info@ymcdirector.com.

Once the SSH handshake succeeds and that green indicator appears, the rest is craft: writing better scripts, building stronger references, and letting a rented GPU do the rendering. Start with one test shot, confirm the pipeline end to end, and only then scale up to a full episode. 🚀

MiniMax H3 Director · Official Website →

WeChat QR

WeChat: ymcandai · Taobao: 杨墨川AI

Articles