I have made a few attempts at it. That one is tricky. It does a decent job of adjusting facial features for taller and shorter people, but without someone that is average height standing next to them for comparison, not a lot changes, and cannot generate an average height person at the same time. Ill keep tweaking on it though. Ill try a different approach, maybe trying to make it generate 2 people where only 1 is affected or something like that.

I am working on that one. I have been working on it a very long time actually. It cannot get it to function properly. I have improved my training method for these quite a bit since the last time I tried, so I will try again. That is the one I want the most. Unfortunately, they trained stable diffusion on "a woman" or "a man" and not a "medium close shot of a woman" and because of that, it seems to be very far off from understanding these concepts where I haven't been able to get it to work even with fine tuning on a large hand labeled dataset.

@SLAPaper scratch that, I started it up again with my newer code and looks like it is actually working. It is early, but is constantly moving the shot. Just zoom for now (extreme close up - full shot). I should be able to get this out today hopefully.

https://civitai.com/models/114460 zoom one done. I still want to try to make it better, but it is functional for now. I was having some trouble with viewing angle today, i'll see if I can figure that out tomorrow.

@frfromg I am not sure if there is much I can do about wrecking the face at a distance. Some of that is the VAE. I am actively working on training a better version of that as well. Most of my ML experience for the past 10 years is with enlarging images, which is where the issues with the VAE lie.. Though SDXL changed the encoder, so will have to do that one too.

Thank you for this list. I train from a merge I did a while back of all the top photorealistic models of the time. But prompts like "taken from below" and "taken from above" don't seem to be concepts it understands, so there are no concepts to I cana lock on to to expand. I'll test some of these out to see if I can find some concepts to lock onto. I appreciate it.

Interesting. I'll see if I can find a way to fix that. At higher weights, any slight bias in the Lora stand out. I probably need to tweak the other hair colors to even it out.

I have attempted to isolate just body hair a few times but it always ends up making the entire subject hairy / less hairy. This has been requested more than anything else (not released). I do have a new method I can try to throw at the problem though, I’ll give it a shot again.

It should work on men as well, though, at higher values, it tends to turn them female.

The model is trained with polarity during taring. In a single batch there are items where the model is in a positive weight, and there are items in a negative weight. It allows the model to target the difference, and only the difference, between the two concepts.