MiniMax H3 (BudgetPixel)

MiniMax H3 is the omni-modal video model that treats text, images, video and audio as one input context, and this node is the budgetpixel route into it. That matters more than it sounds, because H3's local weights ship under the MiniMax H3 Community License, which excludes the EU, UK, South Korea and the US from its territory. If you live in one of those places, running the weights on your own machine isn't something you're licensed to do - but nothing stops you from calling the model through an API, which is exactly what this node does. The hosted path quietly sidesteps the paperwork problem, at a price.

It's also just a genuinely capable video node. Text-to-video, image-to-video with start and end frames, and a reference-to-video mode that mixes up to five reference images, three reference video clips, and three reference audio clips into one 2K clip with native audio. Length runs 4–15 seconds. It's billed at 160 credits per second of output, which puts a 5-second render around 800 credits - this is the most expensive video node in this pack's lineup, so it's not your draft-machine. Use it when you want the quality ceiling, not for iterating on ideas.

Inputs that matter

• prompt (required) - the text description.
• image + end_image - start frame and optional end frame for interpolation. These can't be combined with any reference media.
• reference_images (≤5), reference_videos (≤3), reference_audios (≤3) - the multimodal mixing. Reference videos are billed per second of input on top of the output rate, so a reference clip isn't free context. Reference audio is free. None of these can pair with start/end frames.
• length_seconds - 4–15, default 5.
• aspect_ratio - six options (21:9 through 3:4); ignored/adaptive once any image input sets the frame.

Output is one video socket with a core ComfyUI VIDEO - wire it to Save Video.

The gotchas

First, the input billing. reference_videos carries its own per-second charge on top of the output - the tooltip says estimate via POST /v1/cost, which in the UI means: run the BudgetPixel Cost Estimate node first. A 5-second output with a 10-second reference clip isn't 5 seconds of billing; it's 5 output + 10 input.

Second, the mode exclusivity: reference media and start/end frames are two separate ways to use the node, and the API rejects mixing them. If your graph feeds both image and reference_images, one of them isn't doing anything.

Third, the shared pack rule: interrupting a run stops the local wait but the server-side job keeps going and keeps billing. On a model this expensive, that's a real credit leak if you make a habit of stopping runs.

Install

Standard for this pack - ComfyUI Manager (search "BudgetPixel"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/budgetpixel-ai/comfyui-budgetpixel.git
pip install -r comfyui-budgetpixel/requirements.txt

Restart, find it under Add Node → BudgetPixel → Video, and set your API key via the BUDGETPIXEL_API_KEY environment variable or a budgetpixel.json file. A 2025+ ComfyUI is required for the core VIDEO type. And if you're going to spend H3 money, the Cost Estimate node is less of a nice-to-have and more of a sanity check.