search

Karonar1 / lora_viewer

MIT License
0 stars 0 forks source link

Proper model analysis #3

Open Karonar1 opened 3 months ago

Karonar1 commented 3 months ago

In theory we can analyze the text encoder part of a LoRA to determine what tokens it changes the response to. This will allow us to find the most relevant tags even in the absence of metadata. This is really beyond the original intent of the application, but might be interesting to investigate.

Karonar1 commented 3 months ago

Note that One Trainer does not populate the tag_frequencies metadata, so this will be required for more general support. However, it's not going to be possible to do this for UNet-only LoRAs: there must be a change to a text encoder for us to determine what is being changed.