Regularization and Lora's; using Lora strength during training

[oppie85]

You know how when generating an image with a Lora active, you can vary the strength of the Lora from 0 to 1 (and above)? What if we could control that strength during the training process, effectively eliminating the need for large folders of regularization images by just using the original model but with a Lora strength of 0?

Let's say you're trying to train your own likeness. It's tagged with the caption "an oppie85 man" (the caption is largely irrelevant for this thought experiment).
During the training process, the loss is calculated for the image as usual. However, next, you would generate a target latent using 0 as the Lora strength (thus using the untrained model) and then calculate the loss against the model with Lora strength 1 against a regularization caption (just "a man" in this case).
I hope this explanation makes sense.

I think the advantage of doing it this way would be:

• no need for regularization images
• regularization would happen against the original model which means greater accuracy
• more flexibility for regularization because you could use different regularization captions for each image

Is this viable? I've tried looking for where the Lora strength is applied in the script and whether or not it can be adjusted flexibly during training, but I feel like it must be possible. I'd love it if someone could help me find out how to implement and test this feature.