Women Wrestlers

This was a huge pain in the butsuska to fabricate without using real images. However, I did reference real images in order to fabricate replacement images to get the job done. It may be a bit interesting if you want to know the process.

I grabbed 30 images of professional female wrestlers for visual reference, as if I were going to draw something. From those I generated captions, using Joy-Caption, which, when pushed through Z-Image, gave me a base to work with.

Cherry picking just as many images, 30, from some hours of attempts to match the original I was able to get some decent results for the first stage and, just in case Z-Image knew too much, I masked out the faces (colored them white), it'll have another chance to irritate me in stage_2 (the final one).

I pushed those through OneTrainer, giving the usual system prompt (ignore the white mask) and used this first stage to produce images for the Qwen to work with.

Z-Image knows some of these people so, in order to prevent mistakes, I took any of those out that resembled the identify of those in the original pictures, so that I can honestly say there were no real images used, of real people, living or dead, and then processed to mask out the face of those that remained, moving forward with all good intentions. Then I altered the images with an overlay of white or black, detailing some things that I want to change when pushing them through Qwen-image-edit, and drawing some designs where needed to replicate costume decorations.

Then I did just that, explaining to Qwen what the scribbles were for and how I want them translated into whatever thing I determined were necessary to achieve my goal.

In this way I was able to accentuate muscle, harder facial features, add emblems to outfit (costumes as it were), and manually fabricate, with assistance, the likeness of the characters that the entities were portraying, keeping only the allusion of those characteristics.

From these I cherry picked pairs, 2 each, of each of the 30 chosen different wrestling characters, resulting in 60 images for the final dataset.

I then pushed those through OneTrainer for the final stage.

I hope I was diligent in my scrambling of likeness in order to perfectly align with the guidelines of acceptable, safe content. There is absolutely no way that a likeness could have been passed through since it was never provided by me, and even filtered when Z-Image chose to reproduce a maybe, almost, a little bit, kind of, could be, semblance of the "who is that person" people.

However, if there's any likeness of the original real person, that performed the character that they portray, in any images that are generated by the end user then this is either humanistic interpretation, incidental partial resemblance in that we're limiting the dataset to humans or Z-Image itself just knows the person and neither of these have anything to do with my LoRA, except that it alters the pathways in a way that may make it easier to access that deep data, because of a combination of colors, scene, objects and/or skin color, or any number of normally incidental visual imagery unrelated directly to the person.

whew