bruh - what are you doing with my jock?!

--

Flux v1 version:

I had a previous dataset that I improved greatly for another LoRA and saw that there was enough there to revisit this concept but with much higher quality images. So I tossed those into the captioner, ran it in my go-to Flux trainer and after a few minutes in the oven what came out was not bad. Certainly a v1.0 version, but so long as you are thorough in your prompting, as one should be with Flux, you'll ???????? fairly decent results after a couple of generations (if not on first shot). I created it because even non-merged BFL's Flux 1 D model can get most of the way there, it usually looks like the subject is blowing their nose on a towel requiring a second pass inpaint correction process. Not end of the world and this concept usually doesn't come up but seeing that I could easily revisit this concept with some images made across a few other projects I figured what the heck and give it a go.

You can get aggressive with the strength on this one if it's not getting what you want after a few tries.. I usually start it at 1.0 and increase up to 2 depending on the model checkpoint and scenario. It also made sure the dataset was varied and cropped enough that it will stack well with character LoRAs - my own testing, I really didn't see much, if any, character identity drift when using it but results may vary.

--

SDXL versions:

Will generate full image a man sniffing a jockstrap - fairly temperamental but will generate decent results most of the time. Color saturations aren't what I'd like so you can balance out with tagging, TIs, or vaes.

Can trigger with words like jocksniffer, sniffing a jockstrap, smelling a jockstrap, smelling a jock, etc.

Although it doesn't work well w/ other LoRAs where you start with character LoRA + this LoRA = final output in one shot, it can be used for inpainting if you increase the denoising strength high enough and paint over shoulders/head and give it plenty of wiggle room - if a specific subject is important then you can go back and inpaint to get to intended subject or or do cleanup.

Latest version I moved the rank down from 128 but still keeping the general fidelity of original making it a 892+MB to about a 32 MB LoRA. No other difference so if you liked the 1st one, you can use the 2nd version without really a noticeable difference in outputs.

SDXL Note: I merged this concept with my more recent all-in-one Jockstrap LoRA