Deploy MiniMax H3 with Xiangong Cloud Image: From Signup to First Video
You can get MiniMax H3 running on Xiangong Cloud in about seven steps: register with invite code B1H9M6, get your account ID whitelisted by support, spin up an instance from the private director's image, copy the SSH host/port/password into the 墨川导演台 (MiniMax H3 Director) cloud deploy page, validate the SSH connection, launch ComfyUI, and start generating. The whole flow lives inside the director's 云部署 (Cloud Deploy) tab — no local GPU required, no manual ComfyUI wiring. This guide walks through each step in order, then explains what to do once the green "startup successful" message appears.
If you're searching for a Xiangong Cloud image that ships with everything pre-configured for MiniMax H3, this is that image. The rest of the article is the practical path from a blank account to your first finished clip.
---
Why deploy MiniMax H3 on Xiangong Cloud instead of local?
Because cloud deployment lets you skip the GPU hardware question entirely — you rent compute by the instance, connect over SSH, and let the director handle the tunneling. Local deployment is still supported (inference runs on your own GPU machine), but if you don't have a capable card or you want to iterate across multiple projects without tying up your workstation, cloud is the smoother route.
A quick comparison of the three run modes the director supports:
- Local deploy — inference on your own GPU PC. Best if you already have the hardware and want zero recurring cost.
- Cloud deploy — connects to a GPU server you rent yourself (Xiangong Cloud, AutoDL, Alibaba Cloud, Tencent Cloud, or any custom SSH server). Inference happens remotely through an SSH tunnel.
- API calls — integrate director capabilities as an interface, useful if you're building a pipeline rather than clicking through a UI.
The Xiangong Cloud path is the one this article covers. Note that the director's Xiangong image is a private image (非公开) — you won't find it in the public marketplace. You get access by having your account ID whitelisted, which is why the registration and support steps below matter.
One more thing worth knowing up front: the director isolates local tunnel ports automatically, so multiple users on the same machine or network don't collide. That's the detail that makes cloud deploy actually usable for teams rather than a single-person hack.
---
How do I register on Xiangong Cloud and unlock the director image?
Register through the invitation link, complete real-name verification, then send your account ID to support so they can open the test account that exposes the image. The image is private, so this whitelisting step is not optional — skip it and you'll deploy a bare instance with none of the director's ComfyUI workflows.
Step by step:
- Go to the Xiangong Cloud registration page: https://www.xiangongyun.com/register/B1H9M6 — you must register fresh through this link and fill in invite code B1H9M6.
- Complete registration and identity verification (认证).
- Copy your Xiangong Cloud account ID — it looks like
ac401472-db02-5df3-xxxx-xxxxxxxxxxx. - Send that ID to the director's customer service (客服微信: ymcandai).
- Wait for the test account to be activated. Once it's open, the director image appears in your Xiangong console.
A few practical notes:
- Use the invite code exactly as written — B1H9M6 — during signup. It's tied to the private image access flow.
- The account ID is a long UUID-style string, not your username or phone number. Copy it carefully; a truncated ID means support can't match you.
- If the image doesn't show up after activation, message ymcandai again with your account ID rather than re-registering. Re-registering creates a new ID and restarts the whitelist process.
---
How do I deploy an instance from the image and connect it to the director?
Deploy an instance using the director image, then copy the SSH credentials from the instance panel into the director's cloud deploy page and hit "验证 SSH" (Validate SSH). The connection is a standard SSH tunnel — host, port, password — so there's nothing exotic to configure.
The deployment sequence:
- In the Xiangong console, create a new instance and select the director image you now have access to.
- Launch the instance and wait for it to reach running state.
- On the right side of the instance panel, click SSH. This reveals three values: connection host, port, and password.
- Open 墨川导演台 and go to the 云部署 (Cloud Deploy) page.
- Paste the values into the matching fields: SSH Host, SSH Port, SSH Password.
- Click 验证 SSH (Validate SSH).
What "validate" actually does: it opens a tunnel from your local director app to the remote instance and confirms the credentials work. If it fails, the usual culprits are a stale port (instances sometimes reassign ports after restart), a typo in the password, or the instance not being fully booted yet. Re-copy all three values from the SSH panel rather than editing the old ones.
Because the director isolates tunnel ports locally, you can run more than one cloud instance — or share a machine with a colleague — without the two sessions stepping on each other. If you've ever fought with port 8188 conflicts in a shared ComfyUI setup, this is the part you'll appreciate.
---
How do I start ComfyUI on the cloud instance?
After SSH validation passes, click 启动云端 ComfyUI (Launch Cloud ComfyUI) and wait for the green success indicator. That green light means the remote ComfyUI workflow server is up and the director can queue jobs against it.
Things to keep in mind at this stage:
- Wait for green, don't rush. Clicking launch and immediately generating will fail. The remote ComfyUI needs to finish booting its workflows.
- The green indicator is your health check. If it never turns green, go back and re-validate SSH first — a broken tunnel is almost always the cause.
- Cloud inference runs ComfyUI workflows that support reference-image consistency, shot-to-shot continuity (分镜衔接), and batch queuing. These are the same workflows the director uses locally, just executed remotely.
- Watch the logs. The director's AI 日志 / 系统志 (AI Log / System Log) panel shows real-time generation logs and system status. When something stalls, this is where you look first — not at the Xiangong console.
If you're running a multi-shot project, this is also the moment to think about batching. The project library (项目库) organizes generation tasks and versions per project, and supports both single-shot and batch video generation. Queue a batch overnight on a cloud instance and you wake up to a folder of takes.
---
What's the fastest path to my first generated video?
Write a short script in the AI screenwriter, let it produce a six-段 (six-segment) script and storyboard draft, generate first and last frames for a shot, then queue the video. The director is built so that script → storyboard → frames → video is one continuous flow rather than four disconnected tools.
A minimal first run:
- Write the script. Open the AI 编剧 (AI Screenwriter) and describe your idea in conversation. It replies and writes a six-segment script plus storyboard draft directly into your 剧本库 (Script Library). It handles Chinese, English, Spanish, Arabic, and Japanese.
- Organize the project. Create or open a project in the 项目库 (Project Library) so your generation tasks and versions stay grouped.
- Set up consistency. In the 资产库 (Asset Library), add characters, scenes, and props, and upload reference images. Reference-image consistency is what keeps a character looking like the same person across shots — this is the single highest-leverage step for short drama work.
- Generate frames. Use AI 制图 (AI Image) — text-to-image or image-to-image — to produce first frames, last frames, and character reference images for your storyboard cards.
- Queue the shot. Generate a single shot first to confirm everything looks right, then move to batch generation once you're happy with the look.
- Add audio. Use the 音频库 (Audio Library) for voice cloning, lip-sync, background music, and sound effects, and AI 配音 (AI Voiceover) with IndexTTS 2.5 for multilingual dubbing with emotion control.
For a first test, keep it to one shot and one character. Confirm the pipeline end-to-end before scaling up — debugging a batch of twenty shots when your reference image was wrong is a bad afternoon.
If you're building toward music-driven content, AI MV generates music videos directly. And if you'd rather wire the director into your own tooling, the API mode covers that.
---
FAQ
Do I need my own GPU to use the Xiangong Cloud image?
No. Cloud deploy rents the GPU on Xiangong's side; your local machine only runs the director app and the SSH tunnel. Local deployment remains available if you do have a GPU and prefer on-device inference.
What if the director image doesn't appear after I send my account ID?
Message 客服微信 ymcandai with your account ID and confirm you registered through the invite link with code B1H9M6. The image is private, so it only appears once your account is whitelisted — re-registering won't speed this up and creates a new ID.
Can I use a different cloud provider?
Yes. The director supports AutoDL, Alibaba Cloud, Tencent Cloud, Xiangong Cloud, and custom SSH servers. The Xiangong path is documented here because it has a dedicated private image; other providers work through the same SSH host/port/password flow.
---
Ready to go from signup to first video? Register at Xiangong Cloud with invite code B1H9M6, send your account ID to ymcandai, and follow the seven steps above. For the full feature set, visit www.ymcdirector.com; if you prefer buying through a storefront, search 「杨墨川AI」 on Taobao. Support is also reachable at info@ymcdirector.com, and tutorials live on the blog at blog.ymcdirector.com.