Fix DoRA alpha

[ntc-ai]

This change improves the DoRA implementation by moving the alpha application to after the weight_norm.

This aligns with A1111 allowing DoRA authors to target both platforms.

Example DoRA:
extremely_detailed_v2.zip

Base image (no DoRA)

Before patch (attached DoRA same seed)

(identity changed)

After patch

purrfect

[comfyanonymous]

The A1111 implementation is wrong: https://github.com/KohakuBlueleaf/LyCORIS/blob/main/lycoris/modules/locon.py#L298

[ntc-ai]

Thanks for the link, I just opened up an issue here: KohakuBlueleaf/LyCORIS#196

I currently use alpha to adjust strength to the -1 / 1 range which makes my DoRAs more usable / more easily merged.

My goal is to create DoRAs that work across all the different platforms as well as LoRAs do.