With this workflow, you can partially automate the testing of block weights in Flux LoRAs.

This is particularly interesting when trying to understand what the individual blocks are responsible for and what they control.

Once you know the single transformer blocks, it’s possible to train those exact blocks in Flux LoRAs.

The training then results in much more flexible LoRAs, as bleeding effects are reduced. Additionally, the LoRAs take up significantly less storage space, and training the single blocks is much faster compared to full LoRA training.

maDcaDDie has significantly improved the workflow and added a heatmap that makes it easier to see the changes.
https://civarchive.com/models/886270