MiniMax H3 LoRA Stack

This node has no output image, calls no API, costs nothing, and takes no seed at all. It exists to change which endpoint the H3 image nodes call - and if you don't notice that, you'll spend a while wondering why your LoRA isn't showing up.

Here's the trick, and it's the whole article: connecting this node's loras output to the MiniMax H3 Text-to-Image or Image Edit node switches that node's request from /wavespeed-ai/minimax-h3/text-to-image to /wavespeed-ai/minimax-h3/text-to-image-lora - the -lora twin of the same endpoint. The LoRA isn't merged, loaded into a local pipeline, or applied to a checkpoint. It's a different API URL, with your weights attached to the payload.

What goes in

Three {path, scale} pairs, and the paths are URLs:

• lora_1_path - required, "URL of the first LoRA file"
• lora_1_scale - default 1.0, clamped 0.0–4.0 in 0.1 steps
• lora_2_path / lora_2_scale - optional second slot
• lora_3_path / lora_3_scale - optional third slot

That cap of three is the endpoint's, not the node's. Paths with only whitespace are dropped, and if all three are blank the node raises At least one LoRA path is required - worth knowing, since the 2/3 slots being "optional" makes it look like you can leave the node empty and harmless. You can't.

The scale range deserves a note, because the author's own tooltip admits it: WaveSpeed documents no range for H3, so 0.0–4.0 is borrowed from the Qwen LoRA nodes in the same pack. Treat it as a UI guardrail rather than a spec, and start at 1.0 or below.

One output: loras, of the custom MINIMAX_H3_LORAS type. It only fits sockets expecting that type - the loras input on the two H3 image nodes.

The catch nobody mentions first

Your local LoRA folder is irrelevant here. lora_1_path wants a URL that WaveSpeed's servers can fetch - a HuggingFace resolve link, an S3 URL, a CDN file, not models/loras/mystyle.safetensors. The node never reads your disk. If you trained a character LoRA locally and want to use it on H3 through this route, you have to host the file somewhere public first.

That also means: no .safetensors in ComfyUI/models/loras will ever be picked up by this node, and no local LoRA is "already loaded" for H3 the way it is for a local sampling chain.

Which nodes accept it

Text-to-Image and Image Edit - the two H3 still-image nodes in this pack. The H3 video nodes (Video Edit, Video Extend, Text-to-Video, Image-to-Video, Reference-to-Video) have no loras socket at all, so this stack can't steer them. If you came here hoping to LoRA-ify an H3 video, that's not exposed.

When a stack is attached, expect +$0.015 per image on top of the base call ($0.02/$0.06 at 1k/2k for Text-to-Image, $0.03/$0.09 for Image Edit).

Installing

The stack ships inside the ERPK Collection, so it arrives with the pack:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk && pip install -r requirements.txt

Restart ComfyUI, then add it from ERPK → WaveSpeedAI → MiniMax H3 LoRA Stack. The root requirements.txt installs the Claude, Gemini, OpenAI and xAI SDKs as well - for the WaveSpeed nodes alone, requests, pydantic and Pillow are the ones that matter. You still need a WaveSpeed key configured (right-click canvas → ERPK Settings, or Settings → ERPK → API Keys), because the node feeding off this one makes the call.

Because it's a pure config node with no seed, ComfyUI caches it like any other input: it won't re-bill anything on its own, but changing a path changes the downstream call.

Where people get stuck

A page URL instead of a file URL. Pasting a HuggingFace model page link gets you an API error from WaveSpeed (API Error: ...), not a helpful "that's HTML, not a safetensors file". Grab the direct download link.

Wrong base model. H3 wants H3-compatible LoRAs. A Flux or SDXL LoRA will either error out or quietly do nothing useful - the endpoint does the loading, so you get its error message and no local diagnostics.

Scale doing nothing visible. At low scales on a busy prompt, a LoRA barely registers. Try 1.0 first, confirm it's clearly taking effect, then dial down - otherwise you can't tell "too weak" from "not wired up".

Forgetting the extra cost. The +$0.015 compounds across a batch, and it's added whether the LoRA helps or not.