MiniMax H3 Image LoRA · Text to Image

What you get

MiniMax's H3 image model with a LoRA slot: one prompt and up to three of your own H3-trained weights, one still image back in about one or four megapixels, in fifteen aspect ratios, as JPEG, PNG or WebP. Around $0.042 an image.

The family context is worth thirty seconds, because H3 is the model that made MiniMax interesting in 2026: a 33B omni-modal system that treats text, images, video and audio as one context, which is why H3's video side generates stereo sound in the same pass instead of bolting on an audio model. Its open weights carry a habit nobody expected - the MiniMax H3 Community License scopes its grant to an "Applicable Territory" that excludes the US, the EU, the UK and South Korea, outputs included. That's a restriction on the downloadable files, not on hosted endpoints, which is why a node like this one is a legitimate way to reach H3 from a region the weights can't legally be run in.

Inputs

• prompt - the subject, composition and lighting. Nothing exotic; the H3 line takes long natural-language prompts, and this endpoint's description notes that there's no filter on the prompt, so treat the wording as simply a description rather than something you have to launder.
• aspect_ratio - 1:1 by default, fifteen shapes in all. Fifteen is unusually generous, and it's the reason this node works for layouts a five-shape model would force you to crop.
• resolution - 1k by default, which the tooltip translates as about one megapixel, with 2k at about four. Read that carefully before assuming "2k" means 2048 pixels wide: it's a megapixel budget, not a width, so the actual pixel dimensions depend on the aspect ratio you picked. Both tiers price the same per image here.
• output_format - JPEG by default, with PNG and WebP available. If the image is heading into a compositing chain, PNG is the one that won't compound compression artifacts on the second pass.
• seed - real and repeatable. Same seed and settings, same image, which makes strength tests meaningful.
• loras - a SPICY_LORAS socket. You fill it from a SpicyAPI LoRA node: a direct .safetensors link that opens without a login plus a strength from 0 to 4, chained up to three deep. The tooltip is explicit that the weights must be trained for MiniMax H3 - a LoRA from another base model won't transfer, and the symptom is mush, not an error message.

Outputs are images, a standard IMAGE batch for Save Image or anything downstream, and task id, the string naming this exact request on SpicyAPI.

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 rebuilt from your account. Servers can set SPICY_API_KEY instead; the key is stored in the ComfyUI user folder, never in a workflow or in an image's metadata, which is the part of API-node usage people are right to ask about. No pip install, no weights, no GPU. ComfyUI Manager users can search the pack title, SpicyAPI. ComfyUI with the V3 node API (the pack is tested on 0.39), Python 3.10+.

Things worth knowing

Training H3 LoRAs is early days. H3 shipped as a 33B model with no published consumer hardware floor, so the ecosystem of community weights is thin compared with something like LTX or Krea 2. The old training rule still applies and is worth repeating: architecture lock-in is absolute, and twenty well-chosen images beat two hundred careless ones. The LoRA socket here is a convenience for people who already have matching weights, not a reason to start training on a card that can't hold the model.

Four cents an image adds up across a sweep. There's no seed-less randomness here, so a strength comparison is cheap in theory - hold the seed, change one weight - but each variant is still a paid run. Set Max cost per run in the settings panel; an over-budget quote is refused before any charge.

Refusals belong to the endpoint. The provider describes this one as unfiltered at the prompt level, but the general rule for every API node stands: whatever a hosted provider refuses, the node can't work around, because there's nothing local to patch.

Billing mechanics. Cancelling stops your wait, not the job; the task keeps running on SpicyAPI, is charged if it succeeds, and stays in your console for up to 14 days. And if a model you can see on spicyapi.ai isn't in your node list, restart ComfyUI - the list is pulled from your account at startup.