⭐ Star Minimax H3 LoRA Loader

This node does one thing, and it does it so you can do something else. MiniMax H3 is a 33B open-weight video model, and like every living model ecosystem it's already growing a LoRA library. LoRAs are the unit of currency there - small adapters that steer a big model. But ComfyUI's stock LoraLoader only knows how to apply a LoRA to a model on the fly. If you want to merge two H3 LoRAs into one file - old SD "supermerger" habits ported to video - you need the raw up/down weight pairs, not an applied model. That's what the ⭐ Star Minimax H3 LoRA Loader hands you.

It's the entry point of a three-node mini-pipeline in the ⭐StarNodes pack: Loader → Merge → Saver. The Loader reads a .safetensors LoRA out of your models/loras folder and wraps it into a MINIMAX_H3_LORA object that the merge node can actually chew on. That's the whole job, and it does it without a GPU, without loading the 42GB H3 base model, and without calling any API.

How it works. The node loads the file with safetensors.load_file on CPU and splits the state dict into LoRA pairs - lora_up/lora_down/alpha, plus the lora_A/lora_B aliases some trainers emit. Incomplete pairs (a dangling lora_up with no matching lora_down) are set aside rather than silently corrupting a merge, and everything it found goes into the report output so you can see what happened without guessing.

The one input that matters: lora_name, a dropdown of every .safetensors file currently sitting in models/loras. If your H3 LoRA isn't in that list, it's not in that folder - that's the whole failure surface of this node.

Outputs:

• minimax_h3_lora - the loaded LoRA as a MINIMAX_H3_LORA object. Wire this into a ⭐ Star Minimax H3 LoRA Merge.
• report - a STRING saying how many pairs it found, how many extra tensors, and whether any dangling pairs were ignored. Glance at it once; it tells you if the merge later is working with a full file or a partial one.

Installing. It ships inside StarNodes, so install the pack once: ComfyUI Manager → search Starnodes → Install → restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Starnodes2024/ComfyUI_StarNodes
cd ComfyUI_StarNodes
pip install -r requirements.txt

No model downloads for this one - it only reads LoRA files you already have. The pack's requirements.txt lists heavier deps for other nodes (OpenCV, psd-tools, ollama…), but the pack loads gracefully when one is missing: only the affected node is skipped and the console tells you what to install. A missing node usually means a missed restart.

One honest caveat while you're here: the LoRA merge pipeline itself touches no model and runs anywhere, but the H3 weights these adapters patch are under a community license that excludes the US, EU, UK and South Korea. If you can't legally run the base model, merging its LoRAs is putting the cart before the horse.