AI Magnification

My basic workflow is take a picture of myself, sometimes change my outfit, put me in a fantasy gym setting, and turn me into a video. There is a lot of opportunity in that workflow for the AI to take notice of subtle signals and change my body. I’ve narrowed down some specifics that are kind of interesting.

Slight variations in the first picture get magnified

If you’ve ever taken pictures of yourself, you know how it is. The wrong angle, the wrong lighting, the wrong pose – you can look terrible. Conversely, if you get it right, you can look amazing! Maybe a little pushup bra or a little strategic cleavage, tummy slimming outfits, makeup, flexing as hard as you can… normal amateur model stuff! The gym adds an extra layer, which is that if I’m in the middle of a lift, I can look super muscular, but in real life, I feel like I’m just a normal mostly-fit person.
The AI just takes the picture you give it and is like “that’s what she looks like!” Previously I posted about how AI would make me too skinny or make my boobs smaller. It’s a lot better about that now, and as a result these early subtleties in how I look get amplified because the AI doesn’t know what I really look like, it just knows that first picture. I do train it on my face and body, but this first-picture signal is very strong. If I look busty in the picture, I will be very busty by the end of the process. If I’m flexing, I’ll look pretty strong in the end, but if I’m just doing a thirst trap pose, I might look softer or thinner.

Changing outfits = improved body?

I have to be careful when I ask it to change my outfit, because it can easily decide that if I want to be wearing a bikini, I must have an incredible butt and some luscious boobs. Direction at this point can be counterproductive – for example, if I say hmm, my boobs should be a tiny bit bigger, sometimes it’s like OKAY YOU ARE A PORN STAR. These pictures are hilarious (see the featured image).

Video loses the plot

This is not a problem all the time, but the limits of attention and transformers and the shallow understanding of visual data in these models is pretty clear from my work. This is where I think the most magnification happens – if I am flexing in the original picture, the video shows me completely ripped for whatever five second clip it generates. On the other hand, if it loses “sight” of part of me, especially my face, it kind of reverts to whatever it thinks is the next conglomeration of tokens – usually a generic face, or a generic butt, etc. When I stand up in a bikini or thong, for example, it doesn’t always get my butt right, but it is usually a nice butt!

All this is fine

I can work around this most of the time, but it does mean that some of my videos diverge kind of dramatically in terms of what I look like or how fit/muscular I am. I think the models are getting better, though, and even the improvement in just the last year is kind of incredible. Look for more to come!