Racing Women

I made the LoRA with hopes that it will help produce images of woman with alluring styles of racing outfits, not particularly targeting any real style, with some sensual allusion.

My source images were generated with XL, Z-Image and even Qwen Image Edit 2511, but I'm not a fan of the genre in particular, so I don't know how well the faces match the entities, not that I'd care personally, though I do like the potential that the associated outfits can have. My intent is, of course, to avoid the likeness of any real person when creating this LoRA.

Even though real images weren't used I obliterated all of the faces, covering them with white, and created a LoRA from that without system instructions, which produced mostly images of women without faces when inferencing.

I prompted for various details and colors for the outfit and imagined as many poses I could think of, all upright either sitting, leaning or just standing.

From that set I removed the background from most of them, added a white background of various dimensions for each and examined what I wanted changed in order to guide the training into particular styles of outfit, poses and environment.

I then prompted a new set of images, for the final dataset, using qwen-image-edit-2511, making sure it repaired the face and added the details I wanted.

However, half way through my set of several hundred images, we got hit with Flux Klein, which I tested thoroughly using both the base and 4 step versions and settled on a non-quantized version of a 4 stepper, pushing it to 6 steps. I finished up the dataset using this model/work-flow since it produce acceptable source material and did so very quickly.

I gave it a quick once over glance, inspecting the results for glitches and inconsistencies and pushed some images back through to retouch them or add things I didn't think of before. Then I chose 50 of those that fit my desired target.

That was the dataset I used for training and I didn't fiddle with many hyper-parameters, just added 512 and 768 dimensions for some versatility, allowing for better quality with landscape images, turned the dial up to 20k steps and it took about 3 hours.

There was no nudity or sexual situations in the dataset at any stage of the pipeline.

I won't put this much detail into most of my LoRA description but I figure if I do a few like this then others can test out some of my methods and see if it works for them, I'd like to know what your experience has been, with or without utilizing any of the above steps.

I forgot to mention that I did automatic captioning at each stage. I like automatic it because it gives a lot of weighted concept that I believe helps later when combining portions of different learned aspects. Of course it's easier and quicker this way and the mistakes that the resulting captions often contain don't present to me in ways that I can identify.