Is there a way to wire this up to ffhq dataset?

[johndpope]

Did anyone try?

I see there’s some LDM models here
https://github.com/CompVis/latent-diffusion

I see your script is downloading celebrity dataset as week as this one.

[rinongal]

The FFHQ models are not text conditioned (or conditioned at all?) so you'll likely have to train a new conditional model for this. I believe there's a CelebA-like set with full textual descriptions of the faces that might be useable for this. But really, if you're working with just FFHQ, I'm not sure there's currently any advantage to this approach over the StyleGAN toolbox.

[johndpope]

https://github.com/orpatashnik/StyleCLIP - @orpatashnik thoughts ?

[orpatashnik]

Hi 😊
I'm not sure I understood the question

[johndpope]

Do you think the latent diffusion model / stable diffusion offer any breakthroughs around the generation of faces using ffhq datasets and latest code? It seems like with the styleclip - it's already proven to work - are you looking into diffusion models? wouldn't this repo provide some insight?
if you scroll down - you can see some images I've made using stable diffusion and priming it with a dataset - and keyword - it reminds me of efforts to create / train global directions
#8