The LoRa was trained on just two single transformer blocks with a rank of 32, which allows for such a small file size to be achieved without any loss of quality.
Since the LoRa is applied to only two blocks, it is less prone to bleeding effects. Many thanks to 42Lux for their support.
Description
FAQ
Comments (8)
Any good tutorials on achieving such a small LoRA feat?
This was trained on single Transfomrers 7 and 20. Two blocks training.
At Ostris Github under "Training Specific Layers" u can find the configuration. https://github.com/ostris/ai-toolkit?tab=readme-ov-file#training-specific-layers
@denrakeiw Wonderful - thanks for the link
insane discovery for lora training, i'll be sure to try that out, have you found other blocks use ?
The blocks for people were found by LastBen, and after two days of testing, I found the blocks for clothing.
How did you ensure only 2 blocks are trained? and which of the 25 blocks?
It's 19 Dbl blocks and 38 single layer Blocks.
I trained 7 & 20 Single
Here is a good article:
https://civitai.com/articles/8362/lora-block-weight-testing-for-flux