MiniMax H3 LoRA · Image to Video

What this node is

MiniMax H3 - the 33B omni-modal model that generates picture and stereo sound in one pass - with a LoRA slot on the front. Your opening frame goes in, a 3-to-15 second clip comes out, optionally landing on a closing frame you supply, and up to three of your own H3-trained LoRA files ride along. From $0.05 per second, which is twice the price of the plain H3 image-to-video node; you're paying for the ability to apply weights.

The licence situation is worth one paragraph, because it's the strangest part of H3. The open weights ship under the MiniMax H3 Community License, whose grant is scoped to an "Applicable Territory" that excludes the United States, the EU, the UK and South Korea, outputs included. MiniMax's hosted service was never geofenced, and this node is a hosted call - so if you're in one of those regions, this is the licensed route to the model while the download isn't. Whether that's a technicality or a real distinction depends on your counsel, not on your GPU.

Wiring the LoRA

You don't type a URL into this node. Add a SpicyAPI LoRA node, put a direct .safetensors link and a strength from 0 to 4 on it, and connect its loras output here. Chain LoRA nodes to stack up to three. The link must open without a login - on Hugging Face that means the resolve/main/... file URL rather than the repo page - and the weights must be trained for MiniMax H3, not for some other base; architecture lock-in applies to video exactly as it does to image LoRAs.

The rest of the inputs:

• prompt - action, camera and sound. H3 generates audio alongside the picture, so describe what you want to hear as well as see.
• image_url - the opening frame, and the source of the clip's aspect ratio.
• resolution - 480p by default, up to 1080p. Higher tiers cost more per second.
• duration_seconds - 3 to 15 whole seconds.
• seed - a genuine, repeatable one. This is the input that makes the LoRA slot useful: hold the seed, change the strength, and you can see what the weight actually did.
• last_image_url - optional closing frame for the clip to resolve to.

Outputs are video and task id; the latter is the string that names the request on SpicyAPI, which is what support will ask for.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/spicyapi-ai/comfyui-spicyapi

Restart ComfyUI, then Settings > SpicyAPI → paste a key from spicyapi.ai/console/keys → restart once so the node list is built from your account rather than the bundled snapshot. Headless boxes can export SPICY_API_KEY; either way the key is stored in the ComfyUI user folder and never written into a workflow or a saved image's metadata. No dependencies, no weights, no GPU needed. ComfyUI Manager: search the pack title, SpicyAPI. Recent ComfyUI (V3 node API, tested on 0.39), Python 3.10+.

Practical notes

Train on H3 or don't bother. The KB's rule for LoRA training holds here unchanged: a weight trained for one architecture does not transfer, and the failure is mush rather than an error. H3 is a 33B model that shipped without a published consumer hardware floor, so community training recipes are thin - expect to be early, or to be using somebody else's published weights.

Fifty cents per ten seconds, before the LoRA is even good. Set Max cost per run in the settings panel so a strength sweep can't quietly turn into real money, and use a fixed seed across the sweep so you're comparing the weight rather than the roll. Seeds default to fixed for exactly this reason - a repeated Queue press with nothing changed reuses the cached clip instead of buying another.

The upload path has a ceiling. Your opening frame is re-encoded to PNG and falls back to JPEG only if it exceeds 10 MiB, after which it fails with a "scale it down first" message.

Cancelling stops you waiting, not the job. The task keeps running on SpicyAPI, is charged if it succeeds, and stays in your console for up to 14 days.

Audio is the H3 selling point, and it's in the prompt. H3's whole pitch is that text, image, video and audio are one context rather than an audio model bolted to a silent generator. Use it - a prompt with no sound description wastes the main thing you're paying for.