An SSH tunnel keeps each user's cloud ComfyUI inference isolated by mapping every remote ComfyUI port to a separate local port over an encrypted SSH connection, so two people can run jobs on the same kind of cloud GPU setup without ever touching each other's queue, models, or output. In practice, the local tunnel port is assigned and isolated per user, which means multiple users can work at the same time without cross-talk ("串台"). That single design decision is what makes cloud deploy practical for teams and small studios.
Below I'll break down how that works, why it matters, and how to set it up step by step.
Why does a cloud ComfyUI setup need port isolation at all?
Because ComfyUI is a web service bound to a port, and if everyone points at the same host and port, everyone lands in the same instance. Port isolation gives each user their own local entry point into their own remote instance.
Here's the problem in plain terms. When you rent a GPU server and run ComfyUI on it, ComfyUI listens on some port on that machine. Your browser then talks to that port. If you have one user, fine. If you have five people sharing a workflow, three things go wrong fast:
- Queue collisions. Everyone's jobs land in the same ComfyUI queue, so your 20-shot batch generation gets interleaved with someone else's test renders.
- Model and asset confusion. One person swaps a checkpoint or a LoRA, and everyone's output silently changes.
- Output overwrites. Files land in the same output directory, and naming collisions eat your renders.
Port isolation solves this by making the connection *per user* rather than *per server*. Each user gets their own tunnel, their own local port, and therefore their own isolated channel to their own inference environment. The remote GPU can be the same class of hardware; the sessions stay separate.
How does an SSH tunnel actually isolate each user?
An SSH tunnel forwards a port on your local machine to a port on the remote machine over an encrypted SSH connection. Because the local port is allocated per user, the forwarding paths never overlap.
Think of it as a private pipe. You run something like a local forward — your machine's port 8188 (or whatever the tool assigns) maps to the remote ComfyUI port over SSH. Traffic goes out encrypted, arrives at the remote ComfyUI, and comes back. Nobody else on the network sees an open ComfyUI port, because you never exposed one publicly.
The isolation comes from three layers stacking up:
- Separate SSH credentials. Each user authenticates to their own cloud instance with their own host, port, and password.
- Separate local tunnel ports. The tool assigns and isolates the local port automatically, so two users on the same office network don't fight over
8188. - Separate remote instances. Each user's ComfyUI runs in their own deployed instance, so queues, models, and outputs are physically separate.
That's the whole trick. It's not exotic — it's just disciplined port mapping plus per-user credentials. But it's the difference between a cloud deploy that scales to a team and one that only works for a solo creator.
How do you set up an isolated cloud ComfyUI tunnel step by step?
You connect your cloud GPU instance over SSH, hand the host, port, and password to your director workstation, verify the connection, then launch cloud ComfyUI — and the tunnel stays isolated per user throughout.
Here's the flow using the 仙宫云 (Xiangongyun) test path as a concrete example, since that's the documented route:
- Register on 仙宫云 with the invite code. Go to https://www.xiangongyun.com/register/B1H9M6 and register fresh, filling in invite code
B1H9M6. (You must re-register for this path; an existing account won't carry the invite.) - Complete verification, then send your account ID to support. After verification succeeds, send your 仙宫云 account ID (it looks like
ac401472-db02-5df3-xxxx-xxxxxxxxxxx) to customer service so the test account can be opened. - Deploy an instance from the director workstation image. Once the test account is active, the image becomes visible — deploy an instance using it.
- Grab the SSH details. With the instance running, click SSH on the right side of the instance panel. You'll get the connection host, port, and password.
- Fill them into the cloud deploy page. In the director workstation's 云部署 (cloud deploy) page, enter SSH Host, SSH Port, and SSH password.
- Click "验证 SSH" (Verify SSH). Wait for it to pass. This confirms the tunnel path is live before you commit any GPU time.
- Click "启动云端 ComfyUI" (Launch cloud ComfyUI). Wait for the green success indicator.
- Write a script and test a generation. From here you're generating video through your own isolated tunnel.
A few practical notes that save headaches:
- Verify SSH before launching ComfyUI. If the tunnel is broken, you want to know before a long model load, not after.
- One tunnel per person, not per team. Sharing a single tunnel defeats the isolation — you're back to a shared queue.
- Keep credentials out of shared docs. Each user should hold their own SSH password; that's the first isolation layer.
- Watch the AI log / system log. The real-time generation log and system status view tell you whether a stall is your job or the connection.
If you're running this on AutoDL, Alibaba Cloud, Tencent Cloud, or your own custom SSH server, the pattern is identical: the tool supports general-purpose cloud GPU deployment across those providers, and the SSH host/port/password fields are the same handoff point.
What does port isolation buy a multi-user AI video team?
It buys parallel work without coordination overhead — everyone generates when they want, and nobody's batch job blocks anyone else's.
For AI short drama and AI film production specifically, this matters more than it sounds. A single episode might involve dozens of shots, each needing a first frame, an end frame, and a reference image for character consistency. If you're doing 单镜生成 (single-shot generation) alongside 批量生成 (batch generation), a shared queue turns a 30-minute batch into an all-afternoon negotiation.
With isolated tunnels you get:
- Independent batch queues. Your 20-shot batch runs to completion while a colleague tests a new prompt.
- Stable reference-image consistency. Nobody swaps the character reference mid-batch, because nobody shares your instance.
- Clean version history. Projects and versions stay organized per project, so you can compare takes without digging through someone else's output folder.
- Predictable debugging. When something breaks, the AI log is yours alone — no guessing whose job caused the error.
This is the same reason a professional MiniMax H3 导演台 workflow separates 项目库 (project library), 资产库 (asset library), and 音频库 (audio library) instead of dumping everything into one folder. Isolation isn't a technical nicety; it's what lets a small team behave like a small studio.
Do you need to be technical to run this?
No — the hard part (port allocation, tunnel management, SSH handshakes) is handled by the tool, so your job is basically copy-paste three fields and click two buttons.
The reason people assume cloud ComfyUI is painful is that doing it manually *is* painful: you're writing SSH commands, managing -L forwards, remembering which port maps where, and re-doing it every time an instance restarts. A director workstation abstracts that into a form. You supply host, port, password; it handles the tunnel and the port isolation.
What you should still understand, so you can debug when something's off:
- The tunnel is the lifeline. If Verify SSH fails, nothing downstream works. Check the host, port, and password first, then whether the instance is actually running.
- Local port conflicts are a symptom, not a cause. If a port is busy, the tool's per-user allocation should route around it — if it doesn't, that's worth flagging to support.
- Cloud ComfyUI startup takes time. The green success indicator is your signal to wait for, not to refresh past.
For setup questions, the fastest route is usually the customer service WeChat: ymcandai. You can also find the team by searching 杨墨川AI on Taobao, or 杨墨川 on Douyin, Xiaohongshu, and WeChat Channels. Full details live at www.ymcdirector.com, and there's a blog at blog.ymcdirector.com if you want to read more before committing GPU hours.
FAQ
Q: Can two users share one cloud GPU instance and still stay isolated?
A: Isolation in this setup comes from separate SSH credentials, separate local tunnel ports, and separate deployed instances. If two people point at the same instance with the same credentials, they share a queue and an output directory — that's the cross-talk the design avoids. Keep one instance per user.
Q: What should I check first if Verify SSH keeps failing?
A: Confirm the instance is actually running, then re-copy the SSH host, port, and password from the instance's SSH panel — a single wrong character in the port is the most common cause. If it still fails after re-entering, contact support via WeChat ymcandai with your instance details.
Q: Does this work with cloud providers other than 仙宫云?
A: Yes. The cloud deploy page supports general-purpose cloud GPU deployment including AutoDL, Alibaba Cloud, Tencent Cloud, 仙宫云, and custom SSH servers. The handoff is always the same three fields: SSH Host, SSH Port, SSH password. Exact provider-side steps vary, so check the official site or ask customer service for your specific provider.
Ready to stop queuing behind your own team? Set up an isolated tunnel, verify it, and get back to directing.