🎬 MiniMax H3 视频编辑公式

Editing a video with a prompt has one failure mode that beats all others: the model fixes the thing you asked for and quietly rewrites three things you didn't. This node exists to close that hole. It writes the kind of instruction sheet H3 actually follows - do this, don't touch that.

What it is, and where it fits

🎬 MiniMax H3 视频编辑公式 is a prompt-assembly node for MiniMax H3, the 33B omni-modal video model released mid-2026: 4–15 second clips, 24fps, up to 2K, with native stereo audio generated jointly with the picture rather than bolted on afterward. Video editing is the part that got it attention - #1 in Video Editing (With Audio) on Artificial Analysis, with V2V motion transfer singled out as "like Kling."

The node itself calls nothing. It has no video input and needs no key. It builds a structured text prompt; the source video and the API call happen in the separate MiniMax API node, where your clip is numbered @视频1 - the reference this node writes.

If you're in the US, EU, UK or South Korea, this API path isn't a convenience, it's the licensed one: H3's open weights ship under a Community License whose Applicable Territory explicitly excludes those regions. The hosted API is fine there. Worth knowing before you go hunting for GGUF quants.

How it builds the prompt

H3 responds to a four-part structure - 【参考素材说明】【核心创意】【画面过程描述】【整体要求补充】 - and this node assembles the editing-mode version of it from four pieces:

1. 【参考素材说明】 with your source-video description folded in: @视频1：视频编辑（一段雨夜街道的实拍视频…）.
2. 【编辑指令】 - your instructions, one per line. Lines you've already numbered are left alone; unnumbered lines get auto-numbered for you. This is the part H3 is good at, so being specific pays.
3. 【保持不变声明】 - skipped entirely if you leave it empty.
4. 【整体要求补充】 - a fixed consistency clause ("the rest of the source video stays completely identical, no visible edit artifacts") plus your 声音设计 and your BGM choice, each under a ▍ sub-heading. Empty sections drop out.

Two outputs: H3提示词 (the real one, into your API node's prompt field) and 技术说明, a plain-text readout of what mode you're in, the format, the source-video limits (single clip 2–15s, ≤50MB file) and, if your prompt passes H3's 7000-character ceiling, a warning appended. The second output is documentation for you, not something you wire anywhere.

The inputs that matter

Only two are required, and they're the two you should spend your time on:

• 源视频描述 - describe what's in the clip: subject, scene, action, colour grade. H3 needs to know what it's editing.
• 编辑指令 - one change per line, phrased as add/remove/change. The default shows the shape: 1. 将主角的雨伞换成红色 / 2. 删除画面左侧的路人 / 3. 增加一只黑猫从画面右侧跑过. You can run several in parallel.

Optional, and genuinely optional:

• 保持不变声明 - leave it on the default and you're declaring that camera movement, action timing and lighting stay 100% fixed. Empty means the paragraph isn't written at all.
• 声音设计 - only if you want the audio changed too (new line of dialogue, different voice). Empty drops the section.
• 背景音乐 - three choices: 按声音设计描述, 按剧情和画面自动适配 (default), or 不需要背景音乐. The last one is the one people need: H3 generates sound by default, so leaving the sound field blank is not silence. Choosing 不需要背景音乐 writes the officially recommended 非叙事性音乐：N/A phrasing for you.

Install

Manager → search prompt-formula, or:

cd ComfyUI/custom_nodes
git clone https://github.com/a63976659/ComfyUI-prompt-formula.git

Restart, then find it under 📕提示词公式 / MiniMax H3. Nothing to download - the pack's requirements.txt is a single torch>=2.0.0 line and there are no weights. The MiniMax API node is the thing that needs credentials and credits, and it's a separate install. Be clear-eyed about that cost: this whole node category is metered per call, and video jobs are where a session's spend runs away fastest.

Gotchas

There is no video input on this node. This is the number-one confusing thing about it. Your clip goes into the API node's reference-video slot, and this node just describes that clip in text so H3 knows what @视频1 refers to. If your edit comes back as if the model never saw your footage, that's a wiring problem upstream, not a prompt problem.

One instruction per line, and let it number them. The node detects a leading 1. / 1、 / 1) and won't double-number. Paste a paragraph and you'll get one numbered blob H3 has to untangle itself.

You can't turn off audio with an empty field. 声音设计 blank + BGM left on auto is the default, and the default produces sound. If you specifically want the original/mute, pick 不需要背景音乐.

Mind the character budget. 7000 characters is more than it sounds once you've listed five detailed edits plus descriptions, and the 技术说明 output is the only place you'll be told you blew past it.