⭐ Star Minimax H3 LoRA Saver

The last stop in the ⭐StarNodes H3 LoRA pipeline: you merged two LoRAs, and now you need a .safetensors file you can actually load. That's this node's entire job, and it's the difference between a merge that exists only in memory and a merge you can reuse in any H3 workflow, on any machine.

What makes it more than a dumb file-writer is how it writes. The pack's saver streams the file tensor-by-tensor - it writes each tensor to disk as it goes and never holds more than one in RAM at a time. That matters more than you'd think. A full merge of two chunky LoRAs holds a fair amount of weight data in fp32 during the SVD step; a naive saver would keep the entire result in memory on top of that. This one doesn't, which is the difference between a smooth save and an OOM kill on a modest machine.

The inputs you'll set:

• minimax_h3_lora - the merged LoRA from ⭐ Star Minimax H3 LoRA Merge.
• location - where it lands: loras folder (default models/loras), output folder, or custom path.
• filename - defaults to minimax_h3_lora_merged.safetensors. For custom path, this field holds the full absolute path.
• overwrite - off by default. This is the one switch that'll bite you.

Output: saved_path, a STRING with the absolute destination - handy if you ever want to log or display where the file went.

Save into the loras folder and the node refreshes ComfyUI's LoRA file list afterward, so your freshly merged file shows up in the stock LoraLoader dropdown without a restart. The written file also carries metadata marking it as a StarNodes-merged LoRA and naming its source - nice when you've got ten merged files lying around and can't tell them apart.

The one error people hit is FileExistsError - the node refuses to overwrite an existing file unless you tick overwrite. It's deliberate, but it reads as a crash the first time you see it. Either enable overwrite or change the filename; that's the fix.

Everything else is the shared StarNodes routine: ComfyUI Manager → search Starnodes → Install → restart, or git clone https://github.com/Starnodes2024/ComfyUI_StarNodes into custom_nodes and pip install -r requirements.txt. The merge nodes don't download any model and don't need a GPU - the whole pipeline is CPU math on files you already own. Restart ComfyUI after installing or the nodes won't appear, and if a node is missing entirely, the console prints exactly which package to install.

A note on the bigger picture: the file you save here is only useful as a LoRA for the H3 base model, whose community license excludes the US, EU, UK and South Korea. Saving is legal everywhere; using the result isn't, in those regions.