How to Turn a Script into AI Video: MiniMax H3 Full Workflow
Turning a script into AI video means moving through five connected stages: write the script with AI screenwriting, break it into shot cards, lock character and scene consistency with reference images, generate each shot (first frame → video), then assemble and voice it. On MiniMax H3 running inside 墨川导演台 (MiniMax H3 Director), all five stages live in one workbench — script library, project library, asset library, audio library, and cloud ComfyUI inference — so you never have to stitch four different tools together. This guide walks the full script to video workflow step by step, with the exact order I use and the places where beginners usually get stuck.
---
What does the full script-to-video workflow actually look like?
The workflow is a straight line: AI screenwriting → script library → asset library → AI storyboard images → video generation → audio/voice → export. Skipping a stage usually costs you more time than it saves, because consistency problems compound downstream.
Here's the order I recommend, and why each step exists:
- AI screenwriting (对话式编剧). Talk to the AI screenwriter in natural language. It returns a six-part script structure plus a storyboard draft, and — importantly — it writes the draft into your script library rather than leaving it in the chat window. It supports Chinese, English, Spanish, Arabic, and Japanese.
- Script library (剧本库). Your script, episodes, and shot cards live here. You can import an existing script or create a new one. This is where you convert prose into individual shot cards.
- Asset library (资产库). Upload reference images for characters, scenes, and props. This is the single biggest lever on visual consistency in AI video — do not skip it.
- AI storyboard images (AI 制图). Generate first frames, last frames, and character reference images with text-to-image and image-to-image.
- Project library (项目库). Organize generation tasks and versions per project. Generate single shots first, then batch.
- Audio library (音频库). Voice cloning, lip-sync, background music, and sound effects.
- AI dubbing (AI 配音). IndexTTS 2.5 for multilingual voiceover with emotion control.
The logic is simple: text decisions are cheap, image decisions are medium, video decisions are expensive. Fix as much as possible in stages 1–4 before you spend GPU time on stage 5.
---
How do I write a script that AI can actually shoot?
You write it in shots, not paragraphs — every line of your script should map to one visual unit the model can render. AI screenwriting helps you get there fast, but you still need to enforce shot discipline.
Practical steps:
- Start with a one-paragraph premise, then ask the AI screenwriter to expand it into the six-part structure. Give it genre, target length, and language up front.
- Ask for shot cards, not prose. A usable shot card contains: shot number, duration intent, subject, action, camera framing, lighting/mood, and dialogue or VO line.
- Keep one action per shot. "She walks in, sits down, and opens the letter" is three shots. Models handle one clear action far better than three chained ones.
- Write dialogue separately from action. Dialogue goes to the audio stage; action goes to the video stage. Mixing them in one block makes both harder to control.
- Name your characters consistently. Use the exact same name string everywhere — script, asset library, prompts. Inconsistent naming is a silent consistency killer.
Once the draft looks right, confirm it landed in the script library and split it into episodes and shot cards there. From this point on, the script library is your source of truth, not the chat.
---
How do I keep characters and scenes consistent across shots?
Consistency comes from reference images plus a locked asset library, not from clever prompt wording. Prompts drift; references anchor.
My checklist:
- Build the asset library first. For each main character, upload a clean reference image — front-facing, neutral expression, even lighting. Add scene references and key props the same way.
- Generate character reference sheets with AI 制图. If you don't have photos, use text-to-image to create a character sheet, then reuse that image as the reference for every shot the character appears in.
- Use the same reference across a scene's shots. Cloud inference runs through ComfyUI workflows that support reference-image consistency and shot-to-shot continuity — feed the reference in, don't rely on description alone.
- Lock look before you batch. Generate one test shot per scene, check face/scene/prop match, and only then run the batch. Re-generating a whole batch because shot one drifted is the most common waste of GPU time.
- Version deliberately in the project library. When you change a reference, treat it as a new version. Keep the old one until the new batch is verified.
A useful rule: if two shots share a character, they should share a reference image. If they share a location, they should share a scene reference. Everything else is negotiable.
---
How do I generate the shots — first frame, last frame, then video?
You generate stills first, video second: create a first frame (and often a last frame) with AI 制图, approve it, then animate. This gives you a checkpoint before spending inference time.
Step-by-step:
- Generate the first frame for each shot in AI 制图, using your asset library references. Text-to-image for new compositions, image-to-image when you're refining an approved look.
- Generate a last frame for shots with meaningful movement or a specific end state. First-frame + last-frame pairing is how you control where a shot lands, not just where it starts.
- Write the motion prompt to describe only what changes between the frames — subject movement, camera move, pacing. Don't re-describe what's already visible in the frame.
- Test single shots in the project library. One shot, one version, review, adjust. Single-shot generation is your debugging loop.
- Batch generate once a scene's shots are individually approved. Batch queuing is supported, and the AI log / system log (AI 日志 / 系统志) lets you watch generation progress and system status in real time.
- Track versions per project. Keep the approved take marked so you don't mix versions during assembly.
Where you run inference matters here. Local deployment runs on your own GPU machine. Cloud deployment connects to your own rented GPU server over SSH, with cloud ComfyUI doing the inference. Either works — pick based on how much batch volume you're pushing, and confirm hardware guidance with the official site or support rather than guessing.
---
How do I add voice, music, and lip-sync?
Audio is its own stage: dub with IndexTTS 2.5, add music and SFX from the audio library, then lip-sync where the shot needs it. Doing audio after picture lock keeps you from re-dubbing every time a shot changes.
- AI dubbing (AI 配音) uses IndexTTS 2.5 with multilingual output and emotion control — match the emotional register to the shot, not the whole scene.
- Voice cloning in the audio library lets you keep a consistent voice across episodes, which matters enormously for series work.
- Lip-sync for any shot where a character's mouth is visible and speaking. Skip it for wide shots and VO-over-B-roll — it's wasted effort.
- Background music and SFX are managed in the same library, so your audio assets stay with the project rather than scattered in folders.
- AI MV is worth knowing about if your project is music-driven — it's a separate path for AI music video generation.
If you're producing a vertical short drama, dub first, then lip-sync only the dialogue close-ups. That's usually 20–30% of your shots, and it's where the audience actually looks.
---
How do I set up cloud deployment if my local machine can't keep up?
You connect the director台 to your own rented GPU cloud server over SSH, and cloud ComfyUI handles inference through an isolated local tunnel. Setup is a one-time configuration, and multi-user tunnel port isolation means accounts don't collide.
The tested path (using 仙宫云 as the example):
- Connect to the 仙宫云 private image (non-public).
- Register at https://www.xiangongyun.com/register/B1H9M6 — you must register fresh and enter invite code B1H9M6.
- After registration and verification, send your 仙宫云 account ID to customer support (format example:
ac401472-db02-5df3-xxxx-xxxxxxxxxxx). - Once your test account is activated, the director台 image becomes visible — deploy an instance with it.
- When the instance is running, click SSH on the right side of the instance panel and copy the host, port, and password.
- Paste them into the Cloud Deployment (云部署) page fields: SSH Host / SSH Port / SSH password.
- Click Verify SSH (验证 SSH). Once it passes, click Start Cloud ComfyUI (启动云端 ComfyUI) and wait for the green success indicator.
- Write a script and test a video generation.
Other supported targets include AutoDL, Alibaba Cloud, Tencent Cloud, and custom SSH servers. If you hit a wall, the fastest route is the support WeChat ymcandai — send your setup details rather than a vague "it doesn't work."
---
FAQ
Do I need a GPU to use MiniMax H3 Director?
Not necessarily on your own machine. You can run local deployment on your own GPU computer, cloud deployment against a rented GPU server over SSH, or use API calls. Which fits you depends on your volume — check with the official site or WeChat ymcandai for current guidance.
Can I import a script I already wrote?
Yes. The script library supports importing existing scripts as well as creating new ones, and you can organize them into episodes and shot cards from there.
What's the single biggest mistake beginners make?
Generating video before locking references. Build the asset library and approve first frames first — it saves more GPU time than any prompt trick.
---
Ready to run the whole pipeline in one place? Start at www.ymcdirector.com, add support on WeChat ymcandai for deployment questions, and search 「杨墨川AI」 on Taobao if you prefer to set up that way. For workflow notes and updates, the blog lives at blog.ymcdirector.com — and if you're searching for the tool itself, look for minimaxh3导演台 or MiniMax H3 导演台.