Deploy the Director Studio on Xiangongyun Private Image with Invite Code B1H9M6
Renting a GPU cloud box is a lot like renting a workshop: the space is yours, but it's empty until someone hands you the keys to the machines. For AI filmmakers running MiniMax H3 workflows, that "key" is a private image — a pre-configured snapshot that already has the whole environment baked in. No wrestling with CUDA versions at 2 a.m., no hunting for missing nodes, no wondering why ComfyUI refuses to see your GPU.
This guide walks through deploying 墨川导演台 (MiniMax H3 AI Director Studio) on Xiangongyun using its private image, then wiring it into the studio's cloud deploy panel over SSH. It's the fastest route from "I have a rented GPU" to "I'm generating video from a script."
Why Bother With a Private Image at All?
Public images are the equivalent of a rental car with 200,000 km on it — technically functional, but you never quite know what's under the hood. A private image, by contrast, is a locked-down, purpose-built environment maintained by the studio team. It ships with the inference stack already aligned for MiniMax H3, the ComfyUI workflows that handle reference-image consistency and shot-to-shot continuity, and the batch queueing logic that keeps long renders from falling over.
That matters because AI video is not a single-model problem. A finished shot involves text-to-image for the first and last frames, reference images for character consistency, upscaling, interpolation, and audio — voice cloning, lip sync, background music, and effects. If any one piece is misconfigured, the whole chain stalls. The private image removes that class of failure entirely.
There's also a subtler benefit: reproducibility. When you and a collaborator both deploy the same image, you're working in the same environment. A prompt that renders one way for you renders the same way for them. In a pipeline where you're managing 剧本库 (script library), 项目库 (project library), 资产库 (asset library), and 音频库 (audio library) across multiple episodes, that consistency is worth more than it sounds.
Step-by-Step: Getting the Image and Launching an Instance
The private image isn't publicly listed — you get access through an invite-code registration flow. Here's the whole sequence.
1. Register on Xiangongyun with the invite code.
Go to https://www.xiangongyun.com/register/B1H9M6 and sign up. The critical detail: you must register fresh through that link and enter the invite code B1H9M6. If you already have an account, create a new one — the code is what unlocks the private image, and it won't attach retroactively.
2. Complete identity verification.
Xiangongyun requires real-name verification before it will let you rent GPU instances. Finish this step properly; it's the gate for everything downstream.
3. Send your Xiangongyun account ID to support.
Once verified, copy your account ID — it looks like a UUID, e.g. ac401472-db02-5df3-xxxx-xxxxxxxxxxx — and send it to the studio's support contact. On WeChat that's ymcandai. This is how the team knows which account to grant image access to.
4. Wait for the test account to be activated.
After activation, the Director Studio image appears in your Xiangongyun console. Until then, you simply won't see it — which is normal, not a bug.
5. Deploy an instance from the image.
Pick a GPU spec that suits your workload and launch an instance using the Director Studio image. If you're unsure which tier fits your needs, ask via WeChat ymcandai — hardware requirements depend on resolution, clip length, and how many jobs you queue at once, so there's no universal answer.
6. Grab the SSH credentials.
Once 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 — you'll need them in the next section.
A quick tip here: don't close that SSH panel. It's your only source for the password, and regenerating it mid-setup is annoying.
Wiring the Studio to Your Cloud GPU
Now the interesting part. The Director Studio supports three run modes — local deployment on your own GPU machine, cloud deployment to a rented GPU server over an SSH tunnel, and API access. We're doing the middle one.
7. Fill in the cloud deploy panel.
Open the studio and go to the 云部署 (Cloud Deploy) page. Enter the three values from the SSH panel into the corresponding fields: SSH Host, SSH Port, and SSH Password.
8. Verify the SSH connection.
Click 验证 SSH (Verify SSH). This confirms the studio can actually reach your instance. If it fails, the usual culprits are a typo in the port, a stopped instance, or a firewall rule — check those before assuming something deeper is broken.
9. Launch cloud ComfyUI.
Once verification passes, click 启动云端 ComfyUI (Launch Cloud ComfyUI). Wait for the green success indicator. That green light means the ComfyUI backend is live and the studio is talking to it.
One thing worth appreciating here: the studio isolates local tunnel ports per user, so if several people share a machine or a network, sessions don't bleed into each other. In practice this means you won't accidentally hijack a colleague's render queue — a small feature that saves a lot of confusion on team projects.
10. Write a script and test a generation.
With the connection green, you're ready. Use the AI 编剧 (AI Screenwriter) to generate a six-part script and storyboard draft through conversation — it supports Chinese, English, Spanish, Arabic, and Japanese, and the AI writes the draft directly into your script library rather than dumping text into a chat window you then have to copy by hand. From there, move into the project library and run a single-shot generation first. Test one shot before batching twenty.
Practical Tips for a Smooth First Run
Test small, then scale. A single shot tells you whether the whole chain works — script, storyboard, first frame, motion, audio. Batch generation is a multiplier on whatever you've already proven, including mistakes.
Build your asset library early. Character and scene consistency is the hardest part of AI short drama. Upload reference images into the 资产库 before you start generating episodes, not after. Retrofitting consistency onto 30 finished shots is painful.
Use the logs. The AI 日志 / 系统志 panel shows generation logs and system status in real time. When a job stalls, that's where you find out whether it's the model, the queue, or the connection.
Keep the LLM key separate. The screenwriting and dialogue features can run on your own LLM API key. That gives you control over cost and model choice independent of the video pipeline.
Know your support channels. Beyond WeChat ymcandai, you can find the studio on Taobao by searching 「杨墨川AI」, and on Douyin, Xiaohongshu, and WeChat Channels by searching 「杨墨川」. Email works too: info@ymcdirector.com. The official site is https://www.ymcdirector.com and the blog lives at https://blog.ymcdirector.com.
The Takeaway
Deploying on Xiangongyun with the private image isn't just about convenience — it's about removing variables. You rent the GPU, the image handles the environment, and the studio handles the orchestration. What's left for you is the part that actually matters: the story, the shots, and the pacing.
If you're just starting, do this in order. Register at https://www.xiangongyun.com/register/B1H9M6 with invite code B1H9M6, verify, send your account ID to ymcandai, launch the instance, wire up SSH, and generate one test shot. Don't over-plan the first episode — get one clip out the door, then decide what to build next.
