ANSWERS TO THE LAST AI CHALLENGE ARE 9,11,20,21,24 BTW (pret..
Added 2023-06-06 19:34:53 +0000 UTCANSWERS TO THE LAST AI CHALLENGE ARE 9,11,20,21,24 BTW (pretty sure I mentioned that already but just in case)
Here is the previously mentioned Jisel Lynn AI hybrid set. It's a scene that we shot at the end of a long day of shooting where we just said fuck it to wardrobe and just shot some straight up nudes on the couch. Her face and body was a bit oily still, too, which carried over into the AI's model. It was a simple set but we only got like 5-7 good shots out of it. The quality varies greatly as I'm not trying to pull any trickery here, I just want to show more of the potential of these hybrid shoots.
For a primer on this, check the AI art tag for the previous post where I discuss the process for training these sets. tldr - instead of a variety of images of the model from different locations, looks, times, lighting, etc, I instead bombard the AI with a ton of pictures from ONE set which allows me to make AI images that can blend in with the *real* images.
All images are txt2img with no other controlnets or factors. I imported them into LR and tried to do some color matching but ultimately failed lol. I think I'd have to train the AI on fully rendered and processed images in order to capture the actual colors that I originally intended for the set. I'll continue working on this.
1-4: I heard that putting clothes back on models was actually difficult if the model wasn't trained on images of her wearing clothes. Well, Jisel is 100% nude and I still managed to put some clothes on her lmao. The background goes nuts in these images, btw, as the AI doesn't know where to put picture frames or curtains or anything for that matter, so you'll see all kinds of chaos happening in the background of these.
5-13: Real photos! For science purposes, take note of her “vaginal area” in image 10.
13-18: These would be your "filler" photos that, in the future, we can use to flesh out a set. Not all are perfect, but some are DAMN close to it. Note 14 and 15, and how the vag area doesn't match the *source material*, there are ways to fix this by painting over that area and importing (and Im not kidding here) a model that was trained by someone else’s that matches the aesthetic and using that when painting over the incorrect one (edit: I just came back to say that I have NOT done this for countless reasons, i’m just over-explaining the tech again.)
19-25: Latex bodysuits, along with a AI model someone trained of "latex" textures really fucking works here. I started off just prompting in a latex bodysuit and evolved it into more of a sexy superhero vibe. I was shocked at the flexibility of the model, allowing me to remove her from the scene (with a LOT of prompt engineering) and place her in other situations. Sometimes artifacts of the original training will still carry over. For example, the AI constantly pushed this blue color onto her attire, likely because the only other color that has any prominence in the original set was this specific shade of blue, crazy right?
As always, I welcome your thoughts!
