My first attempt to train a lora locally.
Flux loras seem to work pretty well despite this lora actually failing during training.
I've tried FluxGym and Onetrainer, fluxgym stops training after only 8 of 23 epochs and I have no idea why. It did this twice.
Onetrainer stopped because my PC rebooted or something in the middle of the night. That being said I don't know if it worked because it saved some kind of lora halfway through and it didn't seem to work at all.
If anyone has any tips i'm willing to hear em, or knows why the training stopped on FluxGym.
18 images, 23 epochs, 10 repeats
I believe it only finished 7 or 8 epochs so this likely would have worked far far better, but idk how to make it fully tain and it takes hours just to fail so I only have so much patience just to find out that my lora won't train.
See newest version for notes
Anyway, this works okay but not perfect. Longer prompts seem to give it an extra head for whatever stupid reason. Weights higher than 1 seem to help with it.
I used many different faces and captioned them accordingly so it seems capable of doing more than Yes Man (which was not captioned for btw)
Anyway hope you get some cool stuff out of it. If I fix it i'll update this.
Description
So the first one was technically a failed lora. For whatever reason Fluxgym won't train anymore even after reinstalling.
This is trained with Comfy with 100 epochs, 18 images and 1800 steps.
I'm VERY MUCH still figuring all this out but the prospect of training a lora in just a few hours is pretty enticing so I may keep working on this.
I can't say this version is actually better but the images seem cleaner. It still has similar issues putting heads on the robot and it screws up with feet/tires.