V3 is a large improvement over my other attempts but this is still a WIP. the lora works well between weights like 0.60-0.90, setting it to full has given me mixed results. this seems to be a difficult concept for FLUX to learn so I edited the images in my dataset using gimp to make the veins more prominent, after 15 edits I really nailed down the look but was not jumping to go back and re-edit everything. I ran this as is at a pretty low learning rate ( 0.000035 ) with a dataset of about 30 images for 9000 steps. at some point I'll re-edit what I have and add some more images to the dataset. once I'm happy with my dataset I'll probably end up training this on the civit trainer since training at 1024x1024 is just not possible on my hardware and I think this lora could especially benefit from being trained at a higher res. currently it seems to struggle with full body shots, but close-ups and medium shots seem to be fine.

Files with SHA256: 1a070cf96b5fae08f548e45f128bdd50c03aa0f8437ef890021dce0dbace6fc9