Most people's first attempt at an AI short video goes like this: type a sentence into a box, wait, and get back four seconds of something that looks almost right but feels completely wrong. The character's face changes between shots, the camera drifts for no reason, and the story — if there was one — evaporates somewhere around the second cut.
That's not a model problem. It's a workflow problem. Generating a single clip is easy; stringing clips into something that reads as a *film* is a production discipline, and it's the part most tutorials skip.
This guide walks through the full pipeline using 墨川导演台 (MiniMax H3 AI Director's Console) — a workbench built specifically for AI short videos, AI short dramas, and AI film. The goal: take one prompt and carry it all the way to a finished sequence, without the usual chaos. 🎬
---
Why "Text to Video" Alone Won't Get You a Film
Think of a single text-to-video generation like a photograph. It can be beautiful. But a film is a *sequence of photographs that imply a story* — and the moment you need shot two to match shot one, you've left pure text-to-video territory and entered production territory.
The jump requires three things that a bare generation box doesn't give you:
Consistency. Your protagonist has to look like the same person in every shot. Without a reference image anchoring the character, the model reinvents their face each time.
Structure. A film has a shape — setup, turn, payoff. If you're generating shot by shot with no script, you're improvising a story you can't remember.
Continuity. Shot 3 should pick up roughly where shot 2 left off: same lighting logic, same wardrobe, same spatial relationship. This is where most solo AI attempts fall apart.
MiniMax H3 is the main model here, and it handles the generation itself well. But the console around it — script library, asset library, project library — is what turns generation into directing.
---
Step 1: Write the Script Before You Touch the Video Button
The instinct is to jump straight to visuals. Resist it. A six-shot film with a clear script will beat a twenty-shot film with none, every time.
Open the AI Scriptwriter and describe your idea in plain language. Something like:
*"A street food vendor in a rainy night market notices a stray cat watching him. He slides a piece of fish onto the counter. The cat hesitates, then jumps up. End on both of them under the awning, rain falling."*
The AI responds conversationally and drafts a six-segment script with storyboard cards. It writes in Chinese, English, Spanish, Arabic, or Japanese — pick whichever language you're most comfortable directing in.
Here's the part people miss: the AI scriptwriter doesn't just chat. It writes the draft into your script library. You now have actual storyboard cards you can edit, reorder, and generate from — not a text response you have to copy-paste somewhere else.
Treat this stage like a shot list on a real set. Six segments is a good default: it's short enough to finish, long enough to have an arc. You can edit any card before generating.
💡 Tip: Give the AI scriptwriter emotional beats, not just actions. "He hesitates" produces better shots than "he moves his hand."
---
Step 2: Lock Your Characters in the Asset Library
This is the step that separates a coherent film from a slideshow of strangers.
Go to the Asset Library and set up your cast: characters, scenes, props, and reference videos. Upload reference images for your protagonist, your setting, your key objects. These references are what the system uses to keep your character looking like *your character* across every shot.
The analogy: this is casting and costume, done before the camera rolls. If you skip it, you're hiring a different actor for every scene and hoping the audience won't notice.
For the night market film, you'd want:
- A reference image of the vendor (face, build, apron, hair)
- A reference image of the cat
- A reference for the market environment — the awning, the lighting, the rain
You can also generate these references in AI Image, which supports text-to-image and image-to-image. It's built for producing first frames, last frames, and character reference images for your storyboards — so if you don't have source photos, you can create the references from scratch.
💡 Tip: Generate a clean, front-facing character reference first, then use image-to-image to produce alternate angles. Consistency starts with a strong anchor image.
---
Step 3: Generate Shots in the Project Library
Now you generate. The Project Library organizes your generation tasks and versions by project — so all the shots for the night market film live together, with their version history.
Two modes matter here:
Single-shot generation — you generate one storyboard card at a time, review it, and iterate. This is your precision tool. Use it for hero shots, emotional beats, anything the film hinges on.
Batch generation — you queue multiple shots and let them run. This is your efficiency tool. Use it for coverage, transitions, and shots that don't need frame-by-frame scrutiny.
Under the hood, cloud inference runs through ComfyUI workflows, which is what enables reference-image consistency, shot-to-shot continuity, and batch queuing. You don't have to build the workflows yourself — the console handles that — but it's why the consistency actually holds.
💡 Tip: Generate your hardest shot first. If the emotional climax doesn't work, you want to know before you've generated fourteen supporting shots around it.
---
Step 4: Sound, Voice, and the Final Assembly
A silent AI short video reads as a tech demo. Sound is what makes it read as a film.
AI Voiceover uses IndexTTS 2.5 for multilingual dubbing with emotion control — so your vendor's line can land tired, or warm, or wary, rather than flat.
Audio Library covers voice cloning, lip sync, background music, and sound effects. For the night market scene, that's rain ambience, distant crowd noise, the sizzle of the griddle. These small layers do enormous work.
If your project is music-driven rather than narrative, AI MV handles AI music video generation as its own mode.
Throughout all of this, AI Logs / System Monitor shows you real-time generation logs and system status — useful when a batch is running and you want to know what's actually happening rather than guessing.
---
Where to Run It: Local, Cloud, or API
The console supports three real deployment modes, and picking the right one early saves headaches.
Local deployment runs inference on your own GPU machine. Good if you already have the hardware and want everything on-premises.
Cloud deployment connects to a GPU server you rent yourself, with inference running over an SSH tunnel to cloud ComfyUI. Supported providers include AutoDL, Alibaba Cloud, Tencent Cloud, 仙宫云 (Xiangongyun), and custom SSH servers. The console isolates local tunnel ports automatically, so multiple users don't collide on the same machine.
API access lets you wire the console's capabilities into your own pipeline as an interface.
The 仙宫云 Test Path
If you want to try cloud deployment, the 仙宫云 route works like this:
- Register at 仙宫云 using the invite code B1H9M6 — you must register fresh with this code: https://www.xiangongyun.com/register/B1H9M6
- Complete verification, then send your 仙宫云 account ID to customer service (it looks like
ac401472-db02-5df3-xxxx-xxxxxxxxxxx). - Once your test account is opened, you'll see the director's console image — deploy an instance with it.
- When the instance is running, click SSH on the right side of the instance to get the host, port, and password.
- Enter those into the Cloud Deployment page in the console (SSH Host / SSH Port / SSH password).
- Click Verify SSH. Once it passes, click Start Cloud ComfyUI and wait for the green success indicator.
- Write your script and start generating.
Specific hardware requirements and pricing depend on your chosen provider — confirm those with 仙宫云 directly, or ask via the customer service WeChat ymcandai.
---
Practical Tips That Save Real Time
A few things worth internalizing before you generate your first hundred clips:
Write six shots, not sixty. Finish a short film. The lessons from a completed six-shot piece are worth more than a folder of abandoned fragments.
Build your asset library before you build your shots. Character references are cheap to make and expensive to skip.
Use single-shot generation for anything that carries emotion, batch for everything else. Your time is the scarce resource.
Iterate on the script, not the render. If a shot feels wrong, the fix is usually one line earlier in the storyboard card.
Watch the logs. When a batch behaves unexpectedly, AI Logs tells you why faster than re-generating will.
---
Getting Started Today
The pipeline is genuinely simple once you see it as a pipeline: script → assets → shots → sound → assembly. The AI Scriptwriter drafts it, the Asset Library locks your cast, the Project Library generates and versions it, and the Audio Library finishes it.
Start with one short idea — something with a beginning, a turn, and an end. Six segments. Two characters. One location. Generate it all the way through before you scale up.
If you get stuck, the official site is www.ymcdirector.com, customer service WeChat is ymcandai, and there's a Taobao store under 「杨墨川AI」. The blog at blog.ymcdirector.com has more walkthroughs.
Your first film won't be perfect. Make it anyway — a finished six-shot short teaches you more than a year of planning one. 🎥
