LoRA Scale & Save (CCN)

LoRA Scale & Save (CCN) is the "commit to this strength" node. You feed it a LoRA and a desired strength, and it writes a new LoRA file with that strength permanently baked into its weights - so when you load the new file at 1.0, you get the old file at the strength you specified. No slider to forget, no "which LoRAs had what strength" archaeology in your workflow JSON. It's the file-level equivalent of a preset, and it exists because every LoRA collection eventually has a file you only ever use at 0.6, and you're tired of typing 0.6.

It works by understanding the anatomy of a LoRA file: the weight tensors (lora_up / lora_down) and the alpha scalars that modulate them. The effective strength of a LoRA layer is a function of its alpha relative to rank. So the node gives you three modes for how to adjust it:

• scale_alpha (default) - multiplies every alpha value by value. If a layer has no alpha stored, it creates one from rank * value. This is the gentlest, most "correct" mode because it preserves the trained relationship between weights.
• set_alpha - writes every alpha to exactly value. This is the "make it precisely this strong, uniformly" mode.
• scale_weights - multiplies the actual weight tensors directly. Blunter: it rescales the LoRA's magnitude itself.

value is the number, and the tooltips in the node spell each mode out - trust them. Output is a single filepath string pointing at the newly written .safetensors, saved under your output folder with the filename_prefix you choose (default loras/CCN_scaled_lora, which lands in output/loras/...). It's an output node, so you'll see the path in the UI, and it auto-increments the filename to avoid overwriting.

One implementation detail worth knowing: it handles fp8 tensors safely, promoting them to fp32 for the multiply and converting back, so it won't crash on quantized LoRAs.

Install

ComfyUI Manager → "ComfyCollectorNodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/valkymaera/ComfyCollectorNodes

Restart ComfyUI; find it in the CCN/lora category as "LoRA Scale & Save (CCN)". Stock deps only, no model downloads, MIT.

Common issues

Where people get burned: you're writing a new file, not editing the old one. The original LoRA is untouched - check the output path and load the new file. Also, mode choice matters more than it looks: scale_weights multiplies the magnitude of the deltas themselves, which is not the same mathematical operation as adjusting alpha; for a given "I want 60%" you usually want scale_alpha first. And alpha isn't a universal key - some files (especially LoKr variants, which the KB notes now dominate new character LoRAs) store parameters differently, and the node's key-based alpha detection may not find what a non-kohya format expects. For those, scale_weights is the more reliable hammer. Test on a copy before you re-save your whole library.