Making AI video alone is one thing. Making it with a team — a writer, a storyboard artist, a voice person, an editor — is a different sport entirely. Suddenly the question isn't "can I generate this shot?" but "whose version is the latest, where did that reference image go, and why does the character look different in scene 4?" 🎬
That's where a shared workspace matters. In this walkthrough, I'll show you how to set up a small AI video team inside 墨川导演台 (MiniMax H3 AI Director Studio) — who does what, how to share assets so everyone pulls from the same source, and how to run a review loop that doesn't turn into a group chat of conflicting opinions.
Everything below maps to real features in the tool: script library, project library, asset library, audio library, cloud deployment, and the generation logs. No invented buttons, no imaginary pricing — if something depends on your setup, check with the team (contact details at the end).
---
Step 1: Assign roles that match the actual modules
Before touching a single prompt, agree on who owns which part of the studio. The tool's structure makes this easy because each module is a natural "department."
The writer / showrunner owns the AI Scriptwriter and the Script Library. They talk to the AI in chat to produce a six-segment script and storyboard draft — in Chinese, English, Spanish, Arabic, or Japanese — and the AI writes the draft directly into the script library. From there they manage scripts, episodes, and storyboard cards. Think of this person as the one holding the blueprint; everyone else builds to it.
The visual lead owns AI Image Generation and the Asset Library. They generate first frames, last frames, and character reference images (text-to-image and image-to-image), then upload reference images so characters and scenes stay consistent across shots. This is the person who decides "this is what our protagonist looks like, forever."
The audio lead owns the Audio Library — voice cloning, lip-sync, background music, and sound effects — plus AI Voiceover with IndexTTS 2.5 for multilingual dubbing with emotion control. If you're doing a music piece, they also drive AI MV.
The editor / producer owns the Project Library and the AI Log / System Log. They organize generation tasks and versions by project, run single-shot and batch generation, and watch the logs to see what actually happened versus what everyone assumed happened.
A practical tip: write these roles down in a shared doc on day one. Most team friction in AI video isn't creative — it's someone editing the wrong version because nobody said who owns what.
---
Step 2: Build one shared asset library (the single source of truth)
Here's the analogy I use with new collaborators: the Asset Library is the studio's prop room. If everyone brings their own fake mustache, your actor will have three different mustaches across five scenes. One prop room, one mustache. 🎭
In practice:
- The visual lead uploads first. Character references, scene references, props, and reference videos all go into the Asset Library before mass generation begins.
- Everyone generates from those references. When the project library runs single-shot or batch generation, the consistency comes from the reference images — so the whole team must point at the same files, not their local downloads.
- Name things like a librarian. Use a simple convention:
CHAR_lead_front,SCENE_rooftop_night,PROP_phone_v2. Future-you will thank present-you. - Version deliberately. When a character design changes, upload the new reference and note it. Don't silently overwrite — the editor needs to know which shots were made with the old face.
This is also where cloud deployment earns its keep for teams. The Director Studio supports connecting to GPU cloud servers via SSH — AutoDL, Alibaba Cloud, Tencent Cloud, 仙宫云, or a custom SSH server — with cloud ComfyUI workflows, local tunnel ports automatically isolated so multiple users don't collide. In plain terms: several teammates can work against the same cloud backend without stepping on each other's sessions. (Setup specifics and any account requirements are best confirmed with the team — 客服微信 ymcandai — rather than guessed.)
---
Step 3: Run the review loop through projects and versions, not chat threads
The most common team failure mode: feedback lives in five different chat windows, and nobody knows which note was addressed. Fix it by making the Project Library the review surface.
A workable loop:
- Draft pass. The writer finalizes the six-segment script and storyboard cards in the Script Library. No generation yet — review the blueprint first. Cheaper to fix a sentence than a shot.
- Frame pass. The visual lead generates first and last frames for each storyboard card. Review these as stills. Approve or reject before video generation.
- Motion pass. The editor runs single-shot generation for risky shots, then batch generation for the rest, organized by project and version. Batch queuing means the team can review in parallel instead of waiting in a queue of one.
- Audio pass. The audio lead adds voiceover, lip-sync, music, and effects. If the script is multilingual, this is where IndexTTS 2.5's language coverage and emotion control get exercised.
- Log check. Before calling anything "done," the editor opens the AI Log / System Log to confirm what generated, what failed, and what's still queued. Logs are the referee; opinions are not.
Three review rules that save hours:
- One approver per stage. The visual lead approves frames; the editor approves final cuts. Everyone can comment, but one person signs off.
- Timestamp your notes. "Scene 3, frame B, character's jacket color" beats "the jacket looks weird."
- Batch by risk. Generate the shots most likely to fail first, alone. Once the reference and prompt are proven, batch the rest.
If your team uses an LLM for the scriptwriter chat, note that the AI Scriptwriter and dialogue features can connect to your own LLM API key — useful if your team already has a preferred model for drafting.
---
Step 4: Pick the deployment mode your team can actually support
Collaboration breaks down when the infrastructure doesn't match the team. Three real options:
- Local deployment — runs on your own GPU machine, inference stays local. Great for a solo creator or a two-person team in one room. Less great when a remote collaborator needs to generate.
- Cloud deployment — connects to your own rented GPU cloud server over SSH. Best fit for distributed teams, since everyone works against the same backend. The 仙宫云 test flow is documented: register with the invite code, get your account ID to support, deploy the Director Studio image, then paste the SSH host, port, and password into the Cloud Deployment page, click "Verify SSH," then "Start Cloud ComfyUI" and wait for the green success message.
- API access — for teams that want to wire Director Studio capabilities into their own pipeline or internal tools.
Match the mode to the team, not the other way around. A three-person remote team on local-only deployment will spend more time sending files than making videos.
---
A quick starter checklist ✅
- [ ] Write down roles: writer, visual lead, audio lead, editor.
- [ ] Visual lead populates the Asset Library with character, scene, and prop references before any batch generation.
- [ ] Agree on a file-naming convention and a versioning habit.
- [ ] Set one approver per stage; keep notes specific and timestamped.
- [ ] Choose local, cloud, or API deployment based on where your teammates actually sit.
- [ ] Check the AI Log before declaring a project finished.
None of this is glamorous. But the teams that ship consistent AI short dramas aren't the ones with the fanciest prompts — they're the ones with a shared prop room, a clear sign-off chain, and a log they actually read.
For setup questions, deployment details, or anything that varies by configuration, reach the team directly: official site www.ymcdirector.com, customer service WeChat ymcandai, or search 「杨墨川AI」 on Taobao. You can also find 杨墨川 on Douyin, Xiaohongshu, and Video Accounts, or email info@ymcdirector.com. Anything pricing- or spec-related is best confirmed there rather than assumed.