LoRA Scale Alpha (Yogurt Nodes)

Here's a workflow fact everyone learns the hard way: the default strength of 1.0 is often too strong, and the sweet spot for most LoRAs lives around 0.5–0.8. So every run, you drag the slider. YogurtLoraScaleAlpha is the "stop dragging the slider" node - it scales the LoRA's alpha metadata so that the file itself carries your preferred strength, and you can load it at 1.0 forever.

It's part of ComfyUI-YogurtNodes, yogurt7771's 150+ node all-in-one pack - obscure, MIT-licensed, auto-generated README - and it's one of the pack's small, sharp LoRA tools.

How it works

The key insight is what alpha is. A LoRA's effective contribution to the model is scaled by alpha/rank - that ratio is how the training encodes strength. The weights encode what the LoRA does; alpha is the dial that sets how much of it shows. YogurtLoraScaleAlpha only touches that dial.

Two inputs:

• lora - the pack's LORA object, from YogurtLoadLoraOnly.
• alpha_scale - FLOAT, default 1, range -10 to 10. A multiplier applied only to alpha / network_alpha keys.

And one optional: key_pattern - a regex (default .*) to limit which alpha keys get scaled, if you only want to affect part of the LoRA.

Outputs: scaled_lora and scaled_alpha_count (how many alpha keys were touched). Nothing else changes - the actual tensor weights stay exactly as they were.

Where you'll use it

• Bake in a permanent strength preference: a LoRA you always run at 0.7 gets alpha_scale = 0.7, saved, and loaded at strength 1.0 from then on.
• Fix a LoRA whose alpha metadata is off (trained at an unusual alpha/rank ratio) so it behaves at 1.0 like it did at the trainer's intended strength.
• Nudge a specific block's influence via key_pattern without touching the rest.

The caveat that matters

Because it only changes metadata, the real-world effect depends on how the loader computes scale - ComfyUI's LoraLoader uses the alpha/rank ratio, which is exactly why this works. But remember: this is the complement to YogurtLoraScaleWeights, which scales the tensors themselves. Scale both and you've multiplied the effect twice, which is a great way to invent a LoRA that's dramatically too strong. Pick one path - metadata or weights - and stick with it.

Install

ComfyUI Manager, search ComfyUI-YogurtNodes, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/yogurt7771/ComfyUI-YogurtNodes
cd ComfyUI-YogurtNodes
pip install -r requirements.txt

Restart, look under "Yogurt Nodes". No model downloads; the LoRA is loaded by YogurtLoadLoraOnly and saved with YogurtSaveLora to a new filename so your original stays intact.