SD1.5 version of the LoRA. This took me an awfully long time to put together after trying various schedulers, optimizers, etc. The end result is fairly decent.

For reviewers, a couple of things to keep in mind: there weren't a heck of a lot of images without watermarks that were available for this celeb, so I had to run with what I could source. Second, this is AI generated, so it's only going to be so close to what the actual person looks like. Depending therefore on the checkpoint you use, YMMV.

Recommended settings:

1. Steps: 30-60

2. Sampler: DPM++ 2S a Karras, DDIM, euler a

3. CFG Scale: 7-7.5

4. Use ADetailer

Comments and suggestions are welcome, but please include your generations, otherwise it's hard to tell what didn't work for you.