MiniMax H3 Fast Generate Video

MiniMax H3 Fast Generate Video targets a model called MiniMax-H3-Fast, and the honest summary is in the name: 480P only, 4 to 15 seconds, and the cheapest per-second rate in the pack at ¥0.30. It's the node you use to find out whether your prompt, framing and timing work before you spend 2K money on the same idea.

The one thing that separates it from a plain "low quality setting" is that Fast isn't just H3 with the resolution turned down. It's a different model with its own request shape - the pack injects the required style1 subtype into every request automatically, which you'd otherwise have to know to send by hand. There's no style dropdown and there shouldn't be; the API wants that one subtype and the node applies it for you.

What it supports

Text-to-video, first/last-frame image-to-video, and multimodal reference generation all work here, which is more than the Max Turbo tier gives you - references survive on Fast. What doesn't survive is the 2K tier and the H3-Context-IR prompt-enhancement path, neither of which Fast supports, and there's no prompt_expansion_mode or seed input either. If you're moving a workflow between models, that's the shape of what breaks.

Inputs and outputs

Required: content (from Content Builder), resolution, duration and ratio. The resolution menu has exactly one entry, 480P, so it's really a label rather than a choice. duration is the usual string list from 4 to 15, and ratio covers adaptive, 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16.

Same two ratio rules as the rest of the pack, since the validation is shared: text-only content can't use adaptive (there's no frame to adapt to), and the moment a first or last frame is connected the ratio is rewritten to adaptive regardless of your selection.

Outputs are the standard five: video_url, task_id, status, task_json and request_json. Send video_url into Preview Video for a downloaded MP4, or into Query Task later if you want the metadata again. request_json is the literal body that was posted - for Fast it's only good for your own records, since the 2K regeneration path won't take it.

Cost

At ¥0.30 per generated second, a 5-second clip is about ¥1.50, and the node prints its own estimate in the UI after each run, credits included. Fifteen seconds will run you ¥4.50 - still less than a 6-second 2K render on the base model. That's the argument for iterating here: you can afford to throw away six takes to find the timing that works, then rebuild the same prompt at 768P or 2K and actually use it.

But be clear-eyed about what you're buying: 480P is a draft resolution. It's fine for judging motion, composition and whether the model understood your instruction. It is not fine for delivering anything.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/magicwang1111/ComfyUI-MiniMax-H3-API

Restart ComfyUI, then in the pack directory:

cp local.example.json local.json

Add your MiniMax API key and set base_url to match the key's region - https://api.minimax.cn for a China account, https://api.minimax.io for an overseas one. The pack installs requests and Pillow and nothing else; no checkpoints, no CUDA work, no VRAM. ComfyUI Manager users: search ComfyUI-MiniMax-H3-API.

Traps

Because Fast shares its input validation with the base node, the errors you'll see are the same as everywhere else - but the capability defaults are not. Workflows copied from a Max example will bring a prompt_expansion_mode widget or a 2K resolution setting with them, and this node has neither. Expect to re-wire rather than copy-paste between tiers.

And if the video comes back with something that looks like a compression artifact rather than a generation failure, don't chase it - that's your first clue you're looking at 480P output and should have run the 768P tier.