LoRA Scale Weights (Yogurt Nodes)

Some LoRAs are just weak. Trained conservatively, underfit, or delivered by someone who tuned the alpha weirdly - you load it, crank the slider to 1.0, and it barely shows. YogurtLoraScaleWeights is the brute-force fix: it multiplies the LoRA's actual tensor weights by a scale factor, permanently changing how strong the file is at any strength setting.

It's part of ComfyUI-YogurtNodes, yogurt7771's 150+ node all-in-one pack - obscure, MIT-licensed, auto-generated README - and alongside YogurtLoraScaleAlpha it's one of two ways this pack lets you tune a LoRA's presence before it ever hits a sampler.

How it works

Where LoraScaleAlpha changes the dial (metadata), this one changes the engine (weights). A LoRA's delta is A·B, and scaling those tensors scales the whole delta - so the LoRA's effective contribution at any strength setting shifts by the same factor. Set scale_factor to 2.0 and the LoRA is twice as strong at strength 1.0, at 0.5, at anything.

Inputs:

• lora - the pack's LORA object, from YogurtLoadLoraOnly.
• scale_factor - FLOAT, default 1, range -10 to 10. Global multiplier for the selected tensor keys. Negative values invert the LoRA's effect, which is a legitimate trick for subtracting a style.

Optional:

• key_pattern - regex (default .*) to scale only matching keys.
• scale_alpha - BOOLEAN, default false, and the node's own tooltip says "Usually keep this off." It scales .alpha tensor keys too - and if you've read the alpha article, you can see why that's dangerous: it would multiply the effect twice.

Outputs: scaled_lora and scaled_keys_count (how many keys were scaled).

Where you'll use it

• Rescuing a weak LoRA so it works at a sane strength instead of maxing the slider.
• Damping a too-strong one down (0.5) so the default 1.0 is usable.
• Substituting styles: scale to -1 and you get "the opposite of this LoRA" as a blend ingredient.

The honest caveats

Scaling weights is linear and global - it changes everything proportionally, so it can't fix a LoRA that's weak because it's bad, only one that's weak because it's quiet. And it's a permanent change: after you save, that's the new reality of the file, so save to a new filename and keep the original. As always in this pack, check with YogurtLoraStatViewer first if you're unsure about the key layout you're scaling.

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; tensor math only, so dependencies stay light.