MiniMax H3 Workflow: From AI Scriptwriting to Batch Video Generation

2026-11-08 · English · https://blog.ymcdirector.com

MiniMax H3 Workflow: From AI Scriptwriting to Batch Video Generation

The MiniMax H3 workflow in 墨川导演台 (MiniMax H3 Director) runs through four connected stages: AI scriptwriting → script library & storyboard cards → asset & audio prep → batch video generation in the project library. Each stage feeds the next, so a finished six-part script can become a queued batch of shot-level video jobs without leaving the workbench. Below is the practical, step-by-step version of that pipeline, plus the habits that keep characters and scenes consistent across a whole series.

---

What does the full MiniMax H3 workflow actually look like?

It's a linear pipeline with four stations, and every artifact you create stays in a library you can reuse. Think of it as a factory line rather than a single "generate" button.

Stage 1 — AI Scriptwriting. Open the AI screenwriting panel and describe your idea in conversation. The AI replies with 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.

Stage 2 — Script Library. Here you manage scripts, episodes, and individual storyboard cards. You can import an existing script or start a new one. This is where you edit shot descriptions before anything gets generated.

Stage 3 — Asset & Audio Prep. The asset library holds characters, scenes, props, and reference videos. The audio library handles voice cloning, lip-sync, background music, and sound effects. AI image generation (text-to-image and image-to-image) produces first frames, last frames, and character reference images.

Stage 4 — Project Library & Batch Generation. Projects organize generation tasks and versions. You can generate a single shot to test, then switch to batch generation for the rest of the episode.

A useful mental model: stages 1–3 are cheap and fast to iterate; stage 4 is where GPU time gets spent. Get the script and reference images right first, and batch generation becomes predictable instead of a slot machine.

---

How do I turn a conversation into a usable six-part script?

Talk to the AI screenwriter like a director briefing a writer — give it genre, runtime feel, cast size, and tone in the first message, then refine. The output lands in your script library as editable storyboard cards.

A briefing template that works well:

  • Logline — one sentence on the premise.
  • Format — short drama episode, vertical short video, or MV.
  • Cast — how many characters, and their one-line personalities.
  • Tone — e.g. cold sci-fi, warm slice-of-life, tense thriller.
  • Constraints — number of shots you can afford to generate, and whether dialogue is needed.

Then iterate in follow-ups: "make shot 3 a close-up," "split the opening into two shots," "rewrite the dialogue shorter for lip-sync." Because the AI writes back into the script library, every revision is a real edit to your project, not a lost chat message.

Two practical notes:

  • Write for the model, not just for readers. Keep shot descriptions concrete: subject, action, camera angle, lighting, environment. Vague poetic lines are hard to convert into a first frame.
  • Front-load consistency. Decide character looks and key locations in the script stage, because those decisions become your reference images later.

If you're scripting in a non-English language, the five-language support means you can draft in Spanish or Japanese and still keep the same storyboard structure — handy for multilingual channels.

---

How do I keep characters and scenes consistent across shots?

Consistency comes from reference images plus a disciplined asset library — not from hoping the model remembers. Upload character and scene reference images once, then reuse them across every shot in the project.

The steps:

  • Build the cast in the asset library. For each character, generate or upload a clean reference image (front-facing, neutral lighting, clear features). Use AI image generation's image-to-image mode to produce variations if you need multiple angles.
  • Lock the scenes. Create a reference image per location — the apartment, the rooftop, the lab. Store them alongside props so a recurring object looks the same in shot 2 and shot 40.
  • Generate first and last frames per shot. Use text-to-image or image-to-image to produce the opening and closing frame of each storyboard card. These frames anchor the motion between them.
  • Feed references into generation. The cloud ComfyUI workflow supports reference-image consistency and shot-to-shot continuity, so the same character reference carries through the batch.

The habit that saves the most rework: treat the asset library as a casting department. If a character's reference image is weak, every shot they appear in will drift. Fix the reference, not the individual outputs.

For dialogue-heavy scenes, pair this with the audio library's voice cloning and lip-sync so a character's face and voice stay matched across episodes.

---

How does batch generation work in the project library?

Batch generation lets you queue many shots at once instead of clicking generate shot by shot — but you should always test one shot first. Projects group tasks and versions, so you can compare takes and roll back.

A reliable sequence:

  • Create a project and attach the script/episode you're producing.
  • Generate one hero shot. Pick the most difficult shot — the one with the most motion or the tightest consistency requirement. If it works, the rest usually will.
  • Review and adjust. Check framing, character likeness, and continuity against the neighboring cards.
  • Queue the batch. Send the remaining shots to batch generation. The cloud ComfyUI workflow handles queuing, so jobs process in order rather than you babysitting each one.
  • Track progress in AI logs / system log. The real-time generation log and system status panel show what's running and what failed — use it to catch a bad prompt before it burns the whole queue.
  • Version and compare. Keep the takes you like; the project library retains versions so you can revisit an earlier pass.

Where you run this matters. The director desk supports local deployment (inference on your own GPU machine), cloud deployment (connecting to a GPU server you rent, via SSH tunnel), and API calls. For batch work, cloud deployment is usually the practical choice because long queues don't tie up your workstation.

Cloud deployment in practice (Xiangongyun test flow)

If you're testing on 仙宫云, the flow is:

  • Register at the Xiangongyun site via the invitation link and use invitation code B1H9M6 — you need a fresh registration with the code.
  • Complete verification, then send your Xiangongyun account ID (the long UUID-style string) to customer service.
  • Once your test account is opened, the director desk image becomes visible — deploy an instance from it.
  • When the instance is running, click SSH on the right side of the instance to get host, port, and password.
  • Enter those into the Cloud Deployment page in the director desk (SSH Host / SSH Port / SSH password), click Verify SSH.
  • After it passes, click Start Cloud ComfyUI and wait for the green success indicator.
  • Now write your script and test video generation.

A detail worth knowing: local tunnel ports are automatically isolated, so multiple users on the same setup don't collide. The same cloud deployment page also supports AutoDL, Alibaba Cloud, Tencent Cloud, and custom SSH servers.

---

Which settings and habits make batch output more reliable?

Reliability comes from small, boring decisions made before you queue: shorter shots, concrete prompts, and a tested hero shot. There's no magic parameter — but there are habits that consistently reduce failures.

  • Keep shots short and single-action. One subject, one movement, one camera idea per card. Compound actions are where continuity breaks.
  • Anchor with first and last frames. When both ends of a shot are pinned by generated frames, the model has less room to drift.
  • Reuse references aggressively. Same character reference, same scene reference, same prop reference — every time.
  • Test before you queue. One hero shot per scene type beats a 40-shot batch you have to throw away.
  • Watch the logs. The AI log / system log is your early-warning system during a long batch.
  • Version deliberately. Name or track versions so you know which batch was the "good" one.

For AI MV work, the same pipeline applies with the music video generator — script the visual beats, prep references, then batch the shots against your track.

Specific performance figures, GPU requirements, and pricing vary by setup — check the official site or ask customer service (WeChat: ymcandai) for current details rather than relying on secondhand numbers.

---

FAQ

Do I need my own GPU to use this workflow?

No — you can run locally on your own GPU machine, connect to a rented GPU cloud server via SSH, or use API calls. Cloud deployment through Xiangongyun, AutoDL, Alibaba Cloud, Tencent Cloud, or a custom SSH server is supported.

Can the AI screenwriter work in languages other than English?

Yes — it supports Chinese, English, Spanish, Arabic, and Japanese, and writes drafts directly into your script library.

What if my characters look different in every shot?

Almost always a reference-image problem. Rebuild the character's reference image in the asset library, regenerate first/last frames from it, and re-run the shot. Fix the casting, not the take.

---

Ready to run the whole pipeline yourself? Start at the official site, and if you hit a snag with cloud deployment or reference consistency, the fastest route is WeChat ymcandai — or search 「杨墨川AI」 on Taobao for the shop.

MiniMax H3 Director · Official Website →

WeChat QR

WeChat: ymcandai · Taobao: 杨墨川AI

Articles