Using Aesthetic Images Embeddings to improve Dreambooth or TI results

[bmaltais]

While doing some experimentation I discovered you can greatly improve overall result produced by Dreambooth or TI by adding an Aesthetic embedding overlay.

Here is an example of a Dreambooth image I created from a series of pictures of me. Not a great result in that case:

But if I add an Aesthetic Embedding from the same photos on top, bam, much better:

And here is an example of the config for the Aesthetic Embedding I used:

So this is one way to improve those bad results from Dreambooth or TI.

[bmaltais]

To create the Aesthetic Images Embedding just point at the folder where the images are and name the embedding. Takes literally a few seconds to create:

[Hugo-Matias]

Using the same dataset as the one for the dreambooth model I'm getting vastly different results, the resemblance is lost with the embedding. Does the batch size influence the output or is just to speed up the creation of the embedding?

[bbecausereasonss]

Was going to ask what the batch-size did?

[Hugo-Matias]

That's probably a way to speed the training of the embedding by storing more images in memory for each step, much like textual inversion, etc. I tried to use batch size of 1 now but the end result is the same, the training is very quick regardless so it doesn't really matter I guess.

I must be missing something. Lowering the aesthetic weight gets some of the facial resemblance back but it still goes the opposite way of what is the intended purpose.

[bbecausereasonss]

Hmm strange, let's see what @bmaltais has to say? :)

[bmaltais]

Another potential use. I applied the embedding over a low res picture of me and it was able to "upres" it pretty well:

Original:

Same resolution but "upresed" sort of:

[NapalmCandy]

Hey! it stole you the alarm motion device!

[grexzen]

I have been working so much on aesthetics scorers lately on gens, that I forget how non-gens look irl.

Looking at your real image, the scorer would have not chosen it at all due to 'unreal' features, while it would have chosen the bottom image lol.

Real space is now uglier than latent space.

[Jonseed]

It does three rounds of "aesthetic optimization" before each generation. Is that normal? Might be quicker on the training, but not on the inferencing.

[bbecausereasonss]

Is it better to train aesthetics on less images that look more similar? Not having the best results.

[RomixERR]

How did you make the top menu? At least tell me where to look.

[Jonseed]

Those are "quick settings," which can be set up in the settings tab. Look for quicksettings list.