
This LoRA is meant to replicate the style and look of the harness and jockstrap in Coyote Jock's Varsity Collection. This is my first clothing/style LoRA so take that into account if you want to use this LoRA. The training set contained 14 images.
What happened to V3 do you wonder? I worked on a new version based on the feedback from V2 and realized that adding a more diverse dataset from social media might help with flexibility. V4 takes into account a larger dataset and V3 was trashed.
Use wearing a jock_harness to trigger the LoRA. It can also replicate the jockstrap and (to a point) socks.
wearing a jock_harness will generate the harness and usually the jockstrap
wearing a jock_harness and jockstrap will give you both the harness and jockstrap
wearing a jock_harness and jockstrap and socks should give you the full outfit
I rebuilt the LoRA a few times and think it is in decent enough shape now.
Use jock_harness to prompt for the harness and jockstrap.
The LoRA likes them beefy, muscular, and hairy.
When prompting for color if you want to replicate the classic or vintage jockstrap look use white and <color>. This should hopefully keep most of the harness and jockstrap white and add colored bands.
When prompting for a green jock_harness the green color will meld into the person's chest hair...not sure what that's about.
Sometimes a collar shows up. Add collar to your negatives to avoid this.
If you're mixing this with other LoRAs use a lower weight.
Changes
Used BLIP2 captioning - not sure if this helped much
Pruned a lot more tags in the captions like gym, jock, harness
Again, this is my first clothing/style LoRA so there are a few caveats:
I find using wearing a <style> cyth works most of the time. For example wearing a (blue) cyth or wearing an (icelandic) cyth.
It's very much tied to muscular and hairy men. It also struggles with darker-skinned colored men, unfortunately. I think these limitations are due to the limited dataset.
It will often draw the shoulder straps, but not the chest strap. If that happens you can use (harness) to influence the model to draw one.
You can use harness and jock to fine-tune the strength of the model. The jock keyword seems to be tied to the face of the male model in the dataset which can limit faces and body styles.



