Save LoRA Block Weight

The endpoint of the block-weight workflow. Once you've dialled in a block vector that keeps the parts of a LoRA you want and drops the rest, Save LoRA Block Weight writes that result to disk as a .lbw.safetensors file. From then on it's a normal LoRA you can load and reuse - no more carrying a block vector around in your head or copy-pasting it between graphs.

This matters because tuning block weights is real work. You sweep presets, compare grids, land on something like "just the style blocks of this character LoRA." That tuning is worth preserving. Baking it to a file means the next time you want that effect you load one file at strength 1.0, done - the block weighting is already inside it. It also makes the result shareable and reproducible across machines, which is the same instinct the rest of Dr.Lt.Data's Inspire Pack keeps reaching for.

How it works

It takes an LBW_MODEL - the block-weighted LoRA object produced by Make LoRA Block Weight - and serializes it to a safetensors file. This is an output node: it doesn't pass anything downstream, it just writes the file when the graph runs. The .lbw.safetensors it produces is designed to be read back by Load LoRA Block Weight.

The inputs and outputs that matter

• lbw_model (LBW_MODEL) - the thing you're saving. Comes from Make LoRA Block Weight.
• filename_prefix (default ComfyUI) - the name stub for the saved file.

There are no outputs - it's a terminal save node.

Installing it

ComfyUI Manager: search "Inspire Pack", install, restart. Manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Inspire-Pack

Then restart ComfyUI. The pack carries a requirements.txt (Manager installs it; manual users run pip install -r requirements.txt in ComfyUI's Python env).

Where people get tripped up

You can't feed a plain LoRA or a MODEL into this - it only accepts an LBW_MODEL, so there has to be a Make LoRA Block Weight node upstream producing one. If you tried to use the standard LoRA Loader (Block Weight), that one applies the weighting directly and never emits an LBW_MODEL, so there's nothing to save. The two-node path (MakeLBW → SaveLBW) is the only way to get a file out. Also worth remembering: what you save is a block-weighted derivative of one specific LoRA, still bound to that LoRA's architecture - it's not a universal preset you can point at other LoRAs.