I've noticed something while setting up my datasets that has caused me to start over with some models. When I started I would just describe everything, be as detailed as possible, hoping that the more information I provide, the more concepts the AI can learn, and this is technically true, but if you don't plan to use as many words when prompting this can become an issue. Like, I would describe when an outfit was tight and was showing off the shape of their boobs or butt, but then I learned that the model did a poor job learning the shape of their body without including words like that in my prompt. Same thing for if I mentioned that an outfit showed cleavage, where some models were only learning the correct shape of their boobs when associated with the word cleavage, which I obviously wouldn't use when describing someone who is naked.
Another one I've noticed that it's a good idea not to describe is hair style, length, and color, UNLESS it's specifically in a style or color that you don't want the AI to ever automatically generate without you specifying it, like for example with Karen Gillan looking bald, or Katy Perry with short blonde hair, or Taylor Swift when she went platinum blonde. Personally, I didn't like those styles as much as their other looks, so I'd prefer for those styles to only be generated if I specifically mention them. So I've been going back through my datasets, trimming them down, and retraining models, and I think the results are better than they were before.
Rachel McAdams is a good example where I did both of these things. Especially with her hair, since she's often switched between blonde and brown hair, and some red, and so I've let the model learn all of those styles in context with her name, so that the model will decide which hairstyle to use when, based on the context of the rest of the picture, rather than needing to specify it myself, which I think is pretty cool.
By the way, for those wondering, my next model will be Emma Watson. I've gathered almost 200 images so far but still have a lot to go.
EDIT: Oh and another thing I learned to stop mentioning was if they were smiling, because then the generated images never had a smile unless I specified it. Now the models learn when it makes sense for them to smile or make other facial expressions, rather than needing to specify it
Another one I've noticed that it's a good idea not to describe is hair style, length, and color, UNLESS it's specifically in a style or color that you don't want the AI to ever automatically generate without you specifying it, like for example with Karen Gillan looking bald, or Katy Perry with short blonde hair, or Taylor Swift when she went platinum blonde. Personally, I didn't like those styles as much as their other looks, so I'd prefer for those styles to only be generated if I specifically mention them. So I've been going back through my datasets, trimming them down, and retraining models, and I think the results are better than they were before.
Rachel McAdams is a good example where I did both of these things. Especially with her hair, since she's often switched between blonde and brown hair, and some red, and so I've let the model learn all of those styles in context with her name, so that the model will decide which hairstyle to use when, based on the context of the rest of the picture, rather than needing to specify it myself, which I think is pretty cool.
By the way, for those wondering, my next model will be Emma Watson. I've gathered almost 200 images so far but still have a lot to go.
EDIT: Oh and another thing I learned to stop mentioning was if they were smiling, because then the generated images never had a smile unless I specified it. Now the models learn when it makes sense for them to smile or make other facial expressions, rather than needing to specify it
Last edited:
