This is my first attempt at a photorealistic LORA of Choa, my favorite idol.

If you're just here for settings: DPMPP 2M SDE Karras is what I use. It works well. Other DPMPP variations probably work well - the typical case with realistic Pony models. I'd recommend CFG of around 4-6.

Version 1.1 is out!

This version addresses many of my concerns about v1. It's far smaller in size, retains the likeness of Choa, and deals with higher CFGs better while removing the blurriness of v1.

While I am sure I could improve this further, I may take a break from this model and move on to the other members of AOA. I'd like to get a LoRA of each of them in this quality at some point!

For v1.1, I've selected a different LoRA from earlier in the training run of my 16th attempt, after trying (and failing) to produce better results with four additional training runs. Testing this, I found the previews during training to be far more unreliable than I thought - while the previews would look bad, not having much likeness, the generated results could look quite good. As a result, I was no longer looking for models from preview images that looked good, but rather looking for images that were of a good fidelity and still had enough training for there to be a resemblance. This seemed to work.

The current iteration is the best of 16 different attempts at training my own LORA in Kohya_SS, and while I believe I have learned a lot and made significant progress, I think I could still do better. As such, any feedback on how to improve would be greatly appreciated.

It seems to work best on DPMPP SDE samplers with karras scheduling. It also seems to work a bit better at lower CFG (3-6 range.)

Training Information

In training this model, I went through 16 different attempts, switching between SDXL (JuggernautXL) and Pony (PonyRealism and ultimately settling on CyberRealistic.) The current version was trained on CyberRealistic. I trained on 94 images total, captioned with Automatic Captioning Workflow using the Florence2 LLM. The captions were neutral language, not they typical danbooru tag-style that we associate with Pony. (This may be an avenue of future improvement as I learn more about LoRA training.)

My Current Concerns with the Model

While I am generally pleased with how it turned out, I am not satisfied with the fidelity of the generated images. It seems like no matter what I do - upscaling, face detailing, etc, the faces always come out with a slight blur to them. The original intention of this LoRA was to use it as part of the OliviaRossi's Photorealism workflow (Article on Civit, NSFW) in order to generate high-fidelity, realistic images. Unfortunately, it doesn't work particularly well with that workflow, at least that I've been able to accomplish. As a result, the images can look good but they have this consistent lack of sharpness that I can't get rid of. Suggestions would be appreciated.