I tried making a prompt-only waist concept slider initially and it provided unusable results. I would suspect the main culprit was dectupling the learning rate as per Ostris' tutorial. Try leaving it at 1e-4.

@jenniferhoustonpancakes 1e-4 is SLOOOOOWWW... It's painful. I'll try try one around .0005. I think the subtle part is that you probably need to only go for a few steps (50? 100?) and adjust what you're seeing going wrong (color balance, plastic vs hyperreal, whatever) via the prompts and let it go for another segment. I suspect that it's about 10x more art than science.

How detailed did you get on the prompt? "person who is fat" or did you include "person who is fat, plump, slob, heavy, obese"... I think it doesn't matter if you do multiple iterations during training, but there probably is a "magic" prompt that would work from the get go. You would just never know what it is.

@fohowop124538fdfdfdfds Five terms each on the positive and negative prompts. Positive was "woman who has a tiny waist, narrow waist, small waist, thin waist, hourglass shape" or something similar and Negative was "woman who has a large waist, thick waist" and three other things. Anchor class was "man". Target class was "woman." There was also the same dataset with four pairs with head in each compass direction facing every possible way that I used for my waist set. Same settings used from Ostris's concept slider video.

What I would do now is pretty much just use the default settings and offload the text encoder if it's low on VRAM. Flip both vertical and horizontal. Also cache latents. But since you can rent an A40 for 40 cents an hour (You will want that for Qwen) that really isn't a problem. Also disable sampling since I can do that at home. 3000 steps, maybe more.

It actually did show some signs of being useful but started to get ruined at 800-1000 steps iirc. Actually it showed better results on the large side than small.