MiniMax H3 Batch Video Generation Tutorial: Project Queues and Single Shots
Batch video generation in MiniMax H3 works best when you treat the Project Library as your production queue: you generate one shot first to lock the look, then push the rest of the storyboard through as a batch, and manage every version inside the same project. In 墨川导演台 (the MiniMax H3 AI Director's Desk), single-shot generation and batch generation live in the same place — the Project Library — so you never have to rebuild context between a test render and a full episode run. This tutorial walks through the exact workflow, from script to queued batch, plus the habits that keep multi-shot AI short dramas consistent.
---
What's the difference between single-shot and batch video generation in MiniMax H3?
Single-shot generation renders one storyboard card at a time so you can iterate on prompt, first frame, and reference image; batch generation queues multiple shots from the same project and runs them through the cloud ComfyUI workflow in sequence. The practical difference is *when* you spend your attention.
Single shot is for decisions. Use it when:
- You're testing a new character reference image and need to confirm the face holds.
- The first frame or last frame of a shot is still being tuned in AI 制图.
- You want to check how a specific camera move reads before committing an entire scene to it.
Batch is for execution. Use it when:
- The look of a scene is locked and 5–20 storyboard cards are ready.
- You're producing a full short drama episode and need every shot rendered in one pass.
- You want versioned outputs sitting side by side in the Project Library for review.
The key point: both modes pull from the same project, the same asset library, and the same ComfyUI workflow on your cloud GPU. A batch is not a separate pipeline — it's the same pipeline with the queue turned on. That's what makes minimaxh3 practical for episodic work rather than one-off clips.
---
How do I set up a project queue step by step?
Set up the queue by creating a project, importing or writing the script, generating storyboard cards, attaching assets, then triggering batch generation from the project view. Here's the full sequence.
- Write or import the script. Use AI 编剧 to generate a six-part script and storyboard draft through conversation — it supports Chinese, English, Spanish, Arabic, and Japanese, and the AI actually writes the draft into your Script Library rather than just chatting about it. You can also import an existing script.
- Break it into storyboard cards. In the Script Library, each episode's shots become individual cards. This card list *is* your future batch queue, so fix the shot order here before generating anything.
- Create the project. The Project Library organizes generation tasks and versions by project. One project = one episode or one short film. Don't mix episodes.
- Load the Asset Library. Upload character reference images, scene references, prop references, and any reference video. This is the single biggest lever on batch consistency — every shot in the queue will draw from these.
- Generate one hero shot first. Pick the shot with the most screen time or the most difficult character angle. Run it as a single shot. If the face drifts or the lighting is wrong, fix the reference image now, not after 15 renders.
- Queue the batch. Once the hero shot passes, trigger batch generation for the remaining cards in the project. The cloud ComfyUI workflow handles reference-image consistency, shot-to-shot continuity, and queueing.
- Review versions in the project. Each run lands as a version inside the project, so you can compare takes without hunting through folders.
If you haven't connected a GPU yet, do that before any of this — see the next section.
---
How do I connect a cloud GPU so batch generation can actually run?
You connect a cloud GPU by deploying the director's desk image on a rented server (仙宫云, AutoDL, Alibaba Cloud, Tencent Cloud, or a custom SSH server), then filling the SSH details into the 云部署 page and starting cloud ComfyUI. Local deployment on your own GPU machine also works if you have the hardware.
The 仙宫云 route, step by step:
- Register at 仙宫云 using the invite link and fill in invite code B1H9M6 — you must register fresh with the code for the private image to appear.
- Complete verification, then send your 仙宫云 account ID (the long UUID-style string) to customer service.
- Once your test account is activated, the director's desk image becomes visible in your console. Deploy an instance from that image.
- When the instance is running, click SSH on the right side of the instance panel and copy the host, port, and password.
- In the director's desk, open the 云部署 page and paste them into SSH Host / SSH Port / SSH password.
- Click 验证 SSH. Once it passes, click 启动云端 ComfyUI and wait for the green success indicator.
- Now write a script and test a single shot before queueing a batch.
Two things worth knowing: the local tunnel port is isolated automatically, so multiple users on the same machine won't collide — useful if you're running a small team. And if you'd rather keep inference on your own hardware, local deployment runs everything on your GPU with no tunnel at all. For API-style integration, the director's desk also exposes its capabilities as an API.
---
How do I keep characters and scenes consistent across a batch?
Keep batch consistency by locking the Asset Library *before* you queue, and by using first-frame and last-frame images generated in AI 制图 as anchors for every shot. Consistency in AI video is mostly a pre-production problem, not a prompt problem.
A checklist that works:
- One character, one reference. Don't upload three different photos of the same character and hope the model picks the best. Pick the reference that matches your intended look and stick with it for the whole project.
- Generate first and last frames in AI 制图. Text-to-image and image-to-image both live there. A strong first frame does more for continuity than any adjective in your prompt.
- Write prompts that describe motion, not appearance. Appearance is already handled by the reference image. Spend your prompt words on what happens in the shot.
- Batch in scene order. Shots in the same location benefit from being queued together, because the scene reference stays fresh in the workflow.
- Use the AI 日志 / 系统志 panel. If one shot in a batch comes back wrong, the log tells you whether it was a queue failure or a content failure. Those need different fixes.
For dialogue-heavy shots, pair the batch with AI 配音 — IndexTTS 2.5 handles multilingual voiceover with emotion control — and use the audio library for voice cloning, lip sync, background music, and sound effects. Render picture and sound as separate passes; trying to solve both at once makes debugging a batch miserable.
---
When should I use AI MV or the audio library instead of a straight batch?
Use AI MV when the deliverable is music-driven rather than narrative-driven, and use the audio library whenever a batch needs voice, lip sync, or sound design layered on top. A straight batch produces picture; these tools finish it.
- AI MV is for music videos and rhythmic short-form content where cuts follow a beat rather than a script.
- AI 配音 (IndexTTS 2.5) covers multilingual voiceover with emotion control — good for short dramas that need to ship in more than one language.
- Audio library handles voice cloning, lip sync, BGM, and SFX, so your batch output becomes a finished cut rather than a silent assembly.
A realistic division of labor for a short drama episode: batch-generate all picture in the Project Library, generate all voice in the audio tools, then assemble. Keep the two pipelines separate and your re-render costs stay low.
---
FAQ
Can I run batch generation on my own GPU instead of renting cloud hardware?
Yes. Local deployment runs inference on your own GPU machine. Cloud deployment connects to a server you rent yourself via SSH. Which one fits you depends on your hardware and how many shots you queue at once — check with the team for current guidance rather than guessing.
Does batch generation guarantee identical characters across every shot?
No tool guarantees that. What MiniMax H3 gives you is reference-image consistency plus shot-to-shot continuity inside the ComfyUI workflow, which gets you most of the way there *if* your Asset Library is locked before you queue. Generate a hero shot first, confirm it, then batch.
Where do I get help if my SSH validation fails?
Add customer service on WeChat at ymcandai, or find the shop by searching 杨墨川AI on Taobao. Full product details are at www.ymcdirector.com, and longer guides live on the blog. Pricing, hardware requirements, and account activation specifics should be confirmed with the team — those change, and the official channels are the source of truth.
---
Action plan: write one six-part script in AI 编剧 → lock your character and scene references in the Asset Library → generate one hero shot as a single shot → queue the rest as a batch → layer voice and sound in the audio tools. That's the whole minimaxh3导演台 production loop, and it scales from a 15-second clip to a full episode.