This woman is a unicorn. There is only one Juliette Lake Lewis in the world, and there will never be another woman like her.

Quirky, dark, unpredictable, ingenious, rebellious, pioneering... and so fucking hot. Damn.

However. This LoRA has been the most challenging of my Flux LoRAs. It's the first LoRA I have trained on my 3060, which requires lots of significant differences in set up, from optimizer to batches, the settings are quite different, so I'mma blame that.

The end epochs are obviously overfit, just a few epochs before and you already start to lose likeness. It just doesn't follow prompts that well. It has great likeness, and it isn't broken. It doesn't even suck, but it's below par for my uploads. I see it put out "blurry" images and fake jpg artifacts every once in a while. It likes photographers' bright lights in her face. It likes a certain type of hat. Lots of quirks of being overfit. You can tame it some lowering strength, but below .9 it's not worth it. .925 ain't so bad. In the end, even with strength lowered and very explicit prompts, it still isn't "easy" to guide.


Play with it. Use very precise and lengthy prompts. This is the final epoch of two training sessions. It is obviously therefore the strongest and most overfit of all the checkpoints, but I feel it performs better than the lower epochs. I will upload another less-trained version if anyone wants it.

I apologize for lowering my standards for this. It was a request, too, which makes it worse. I just have a lot on my plate. Probably 6 HY LoRAs in the pipeline to be tested and uploaded. I spent what I consider to be "too much" time on this. I'll revisit it in the near future if peeps ask I guess, just right now it's a thorn in my face and I'm pulling it out.