Hacker News Reader

Comment by guerrilla | Hacker News Reader

Why do you believe that?

Coming up with ok-looking generated faces is not ground-breaking in the ML world and when the state-of-the-art model botches it completely, you start looking for the reason.

And coming up with very good looking modifications/poses of known actual faces is very well achieved.

How compose-able are ML solutions? I'm wondering if coming up with ok-looking generated headshots (face at a specific, known, predetermined angle, in a neutral and standardized context, independent of background features) is not ground-breaking, but embedding that capability in a dynamic image generator with myriad other objects, headgear, lights and shadows, etc. may be.

Facebook's Make-A-Scene, which looks almost as good as DALL-E 2, does something very much like that, using face-specific targeting to boost the quality of that part of images: https://arxiv.org/abs/2203.13131#facebook

It was mentioned in the DALL-E github repo.

Because they said so:

  Preventing Harmful Generations

  We’ve limited the ability for DALL·E 2 to generate violent, 
  hate, or adult images. By removing the most explicit content 
  from the training data, we minimized DALL·E 2’s exposure to 
  these concepts. We also used advanced techniques to prevent 
  photorealistic generations of real individuals’ faces, 
  including those of public figures.

Is there a fork that unhandicaps it?

To do such a fork, wouldn't you have to build a completely new model with the same training data and everything else the same except leaving out the restrictions?

You should be able to finetune on a dataset that has captions of human faces.

No, you can use transfer learning.

Like a lot of other Open AI work, DALL-E 2 isn't open. There are people working at re-implementing it, but training has considerable computational costs.

Reply on news.ycombinator.com