Hacker News Reader

Comment by aerodog | Hacker News Reader

Has the same problems as midjourney et al: you can feed it pictures of a friend or yourself or a celebrity, and the result is always off - not recognizable as them

View on HN

That's pretty easily fixable if you train a LoRA or similar so that the model has a specific likeness in mind. (You can look at - and despair at - Civitai if you want proof.)

It's harder to do at inference time without training, but I wouldn't assume it'll be impossible forever, especially with the existence of ControlNet.

This is a GAN, so you can just project the image of yourself into the latent space (which will give you a near-pixel-perfect reconstruction), fix the identity-relevant variables in the _z_, and edit it as necessary. (No workarounds like finetuning necessary. Just one of the many forgotten advantages of GANs.)

You can project an image into the latent space with diffusion model too, DDIM inversion.

>fix the identity-relevant variables in the _z_

Is that how the latent space works though; Like if it's a 300-dim vector, is the face at locations 0-10?

I'm a bit surprised this is the only comment mentioning ControlNet.

Some will surely consider that a feature rather than a bug.

But then how am I supposed to be the kind of creepy person that jerks it to poorly modified "nudes" made from Facebook profile pics?

"Back in my day stalkers had to use their imagination, by gum."

Reply on news.ycombinator.com