How to Batch Generate Storyboard Shots for AI Short Dramas with MiniMax H3
Batch generating storyboard shots for an AI short drama means turning one script into a queue of shot cards, then letting MiniMax H3 render them in sequence inside 墨川导演台 (MiniMax H3 导演台). The short answer: write or import your script into the 剧本库, split it into 分镜卡片, lock your characters and scenes in the 资产库, then push the whole episode into the 项目库 and run 批量生成. The rest of this guide walks through that workflow step by step so you can repeat it on every episode.
If you're searching for minimaxh3 batch workflows, this is the practical version — no theory, just the order of operations that keeps consistency from drifting across 20+ shots.
---
What does a batch storyboard workflow actually look like in MiniMax H3?
It's a five-stage pipeline: script → storyboard cards → assets → project queue → batch render. Each stage feeds the next, and skipping one is the most common reason batch output looks inconsistent.
Here's the full chain inside the 导演台:
- AI 编剧 — Generate a six-part script and storyboard draft through conversation.
- 剧本库 — The draft gets written into your script library as episodes and 分镜卡片.
- 资产库 — Upload reference images for characters, scenes, and props.
- AI 制图 — Generate first frames, last frames, and character reference images per shot.
- 项目库 — Group the shots into a project, then run 单镜生成 or 批量生成.
The key insight: batch generation is not "select all and hit go." It's a queue that inherits whatever consistency anchors you've already set. If your 资产库 is empty, the batch will drift. If your 分镜卡片 are vague, the batch will drift faster.
---
How do I turn a script into storyboard cards fast?
Use the AI 编剧 in conversational mode, then let it write directly into the 剧本库 — don't copy-paste by hand. The 编剧 replies in chat *and* commits the draft to your library, so the cards are ready to edit immediately.
Step by step:
- Open AI 编剧 and describe your short drama: genre, episode length, number of characters, tone.
- Ask for a six-part structure — this gives you a natural episode skeleton you can split into shots.
- Specify language. The 编剧 supports Chinese, English, Spanish, Arabic, and Japanese — pick the one your final dialogue will be in.
- Request the storyboard draft explicitly ("give me the 分镜 breakdown for each beat").
- Confirm the write. Once the AI commits the draft, open 剧本库 and review the 分镜卡片.
A prompt pattern that works well:
"Write a six-part short drama script, 8–12 shots total, modern urban setting, two leads. For each beat give me: shot description, camera move, character present, and one line of dialogue. Then write it to my script library."
Then in 剧本库, edit each card down to a single visual action. One card = one shot. If a card describes two actions, split it. Batch rendering punishes compound shots.
You can also import an existing script if you already wrote it elsewhere — the library accepts imports and new scripts alike.
---
How do I keep characters consistent across a batch?
Lock them in the 资产库 with reference images before you generate a single frame. Consistency in batch mode comes from references, not from prompt wording.
The 资产库 manages four asset types:
- 角色 (characters) — upload a clear reference image per character
- 场景 (scenes) — upload location references
- 道具 (props) — anything that recurs across shots
- 参考视频 — motion or style references where useful
Then use AI 制图 to produce:
- 首帧 (first frames) — the opening image of each shot
- 尾帧 (last frames) — the closing image, which helps the next shot connect
- 角色参考图 — clean character sheets you reuse across the episode
Why this matters for batch work: the cloud inference runs through ComfyUI workflows that support reference-image consistency and 分镜衔接. Your uploaded references are what those workflows anchor to. Empty asset library = no anchor = every shot invents a new face.
Practical tip: name your assets exactly as they appear in your script ("Lin — red coat", "Rooftop night"). When you're scanning 20 cards in the project queue, matching names save you from mis-assigning references.
---
How do I set up and run the batch in the 项目库?
Group the shots into a project, verify one shot renders correctly, then release the batch. The 项目库 organizes generation tasks and versions by project, and supports both 单镜生成 and 批量生成.
Recommended sequence:
- Create a project for the episode (e.g., "Ep03 — Rooftop Confrontation").
- Pull in the 分镜卡片 from 剧本库.
- Attach assets — assign the right character, scene, and prop references to each card.
- Generate one shot first. Pick the most complex card (usually the one with the most characters). Fix prompts, references, or framing here.
- Only then run 批量生成. The queue inherits the settings you validated.
- Watch the AI 日志 / 系统志 while it runs — you get real-time generation logs and system status, so you can spot a stalled job instead of guessing.
Version control matters here. The 项目库 keeps versions per project, so when a batch half-fails you re-run only the broken shots rather than the whole episode. Don't overwrite a good batch with a speculative one — keep the version.
If you're running cloud inference, confirm 云端 ComfyUI is started and green before you queue. A batch that fires against a cold backend just wastes queue slots.
---
Should I run batch storyboards locally, in the cloud, or via API?
Pick based on volume and hardware, not preference. All three modes are real and supported.
- 本地部署 — Runs on your own GPU machine. Inference stays local. Good for small batches and privacy-sensitive scripts.
- 云部署 — Connects to a GPU server you rent yourself, inferring over an SSH tunnel. Good for large batches and long episodes.
- API 调用 — Access 导演台 capabilities as an interface. Good if you're wiring generation into your own pipeline.
For cloud, the 导演台 supports general cloud GPU deployment across AutoDL, 阿里云, 腾讯云, 仙宫云, and custom SSH servers. It connects to cloud ComfyUI over SSH, with local tunnel ports isolated automatically — so multiple users on the same machine don't collide. That port isolation is genuinely useful if you share a workstation with an editor.
If you're testing on 仙宫云, the flow is:
- Register at the 仙宫云 site with invitation code B1H9M6 (you must re-register and fill it in).
- Complete verification, then send your 仙宫云 account ID (the long UUID) to customer service.
- Once your test account is opened, the 导演台 private image becomes visible.
- Deploy an instance using that image.
- Click SSH on the instance panel and copy the host, port, and password.
- Paste them into the 云部署 page's SSH Host / SSH Port / SSH password fields.
- Click 验证 SSH, then 启动云端 ComfyUI, and wait for the green success indicator.
- Write a script and test a generation.
Exact GPU specs, rental pricing, and generation times vary by provider — check with 客服微信 ymcandai or the official site for current details rather than relying on stale forum posts.
---
What else can I batch alongside the video shots?
Audio and music, using the same project structure. The 音频库 handles voice cloning, lip-sync, background music, and sound effects, and AI 配音 runs on IndexTTS 2.5 with multilingual output and emotion control.
For a short drama batch, this means:
- Voice — clone or assign a voice per character, generate dialogue lines per shot.
- Lip-sync — align mouth movement to the generated dialogue.
- Music/SFX — drop in background beds and effects per scene.
- AI MV — if your short drama has a musical segment, the AI MV generator covers it.
Generate audio after the video batch is locked, not in parallel. If you re-render a shot, its dialogue timing shifts, and you'll be re-syncing audio you already finished.
---
FAQ
Can I batch generate without uploading reference images?
Yes, technically — but expect character drift across shots. The 资产库 references are what the ComfyUI workflows anchor consistency to. For anything longer than 3–4 shots, upload references first.
How many shots can I queue at once?
Queue limits depend on your deployment (local GPU vs. cloud instance) and are not a fixed published number. Check with 客服微信 ymcandai for your setup.
Does batch generation work in languages other than Chinese?
Yes. AI 编剧 supports Chinese, English, Spanish, Arabic, and Japanese, and AI 配音 via IndexTTS 2.5 is multilingual with emotion control.
---
Where to go next: start with a 3-shot test batch — one character, one scene, one prop — and confirm consistency before scaling to a full episode. Browse the tool at www.ymcdirector.com, add 客服微信 ymcandai for deployment questions, or search 淘宝 for 「杨墨川AI」. More workflow write-ups live on the blog at blog.ymcdirector.com.