Bibi Jones - PDXL/Pony - 55 MB

Bibi Jones lora, trained with basic PDXL. I'm generally happy with the results for realistic Pony checkpoints as well as combining it with various cartoon style loras on other Pony checkpoints.

I've attempted other (bad) 1.5 loras and embeddings and results were subpar, so decided to be serious with this one as a learning exercise. My goal was to make a versatile lora that was flexible and didn't take up a huge amount of space. My biggest aesthetic issue with several character loras is they seem to "pinch" the face and my theory was that this is because the training data includes selfies taken with the small front camera; Loradude kindly made a Katelyn Lordahl SDXL lora with all data from a single camera on a single source and it made me think that I maybe wasn't wrong, but not necessarily was right. So I gathered 90-100 non-selfie photos, cropped them at a variety of the native SDXL resolutions, captioned them, and got to training, aiming to keep them around 50 MB.

There are ultimately three versions of the lora that I'll release this week, each with different training parameters. The captioning improved with each iteration but it didn't really seem to have an effect; as time went on I dropped a couple photos from dataset if an element showed up enough to bother me. v1 (this one) and v3 are the most similar, v2 had a number of training differences. Fun to see how each produces different images after my incredibly basic prompting. Generation parameters were just whatever the checkpoint recommended and I don't think any of the example images used adetailer/inpainting. If generating multiple people or using a complicated prompt I've found inpainting to be useful.

Bibi has a tattoo on her lower back that was present in the training images and I tried my best to caption for it in v3 but had zero luck in generations.

With the large HQ dataset my intent is to make a v4 that's relatively huge in size (300MB?) to see if there are improvements/changes. My guess is "no" but that's why I'm testing it haha. Who knows. And then a v5 with a much smaller dataset to see if 100 pics is just overkill. Also intend to train the dataset on SDXL after that. I'm sure someone has tested these parameters out before but I didn't find it.

I'm obviously a beginner so if you have any insights or thoughts I'm here to learn.

As always, this lora depicts a real person so be responsible and don't generate inappropriately. Due to the nature of the dataset the model will generate accurate if so prompted.