Dall-E
I generated all of the following faces with Dall-E using a variety of prompts. Most contained the word ‘portrait’
The overall facial structure is great. His expression almost says “Come on, You really think I am AI?” I have to peer at the photo a little before I notice that is is AI generated. There is some very nice detail (where its not washed out on the skin, as if with a beautifying filter) The lips and facial hair have a nice definition to them.
I wasn’t going for glamour, though it is definitley capable of those kind of shots as well. (impressive swoopy hair curls here!) While teeth were sometimes a problem for Dall-E, it did surprisingly well with hands, which is probably the biggest stumbling block — for generative AI and art students alike.
Where AI also really falls short, but Dall-E does exceptionally well, is with emotion. I was able to generate a surprising amount of character and emotion in these AI portraits.
I wanted to see if Dall-E could take direction. Portraiture is more than just head-shots. How about more candid shots with different poses? I was pleased to see I could fairly easily get different poses like this three quarter profile.
What about expressions? I was usually able to direct the expressions quite a bit, with a nice amount of subtelty as well. I usually got what I hoped for like this “slight smile”.
While many Dall-E images suffer from flat or cheap flash photo lighting, I was able to get some beautiful studio lighting with these portrait generations.
Stare at any of these for a while and you will spot the flaws. The image above has some eye trouble on the right eye, or is that just a stray contact lens? That one may be up to your brain to choose. The right upper lip is probably more of an AI giveaway however.
Despite these fairly minor AI flaws (notice the uneven bottom lip in the image below) and sometimes thanks to them, DALL-E is able to create some of the most believable AI generated people I have seen.
Failures. While a surprising percentage of the portriat images were usable, probably most of them (something I cant say for most AI generators) there were still plenty of flawed images.
Dall-E seems to have spent all its energy on the remarkably good hand in this next one. The eyes are not just mismatched, the right eye is missing the iris and is nothing but pupil. Most of the image is off in this one which appears to have been ‘unfolded’.
Sometimes there were both eye and hand problems. A fused-tooth mouth with rubber-band lips doesn’t help.
Overall, Dall-E produces amazingly realistic faces which are very controllable, in plenty of ways. To demonstrate that, I’ll close this Dall-E section with a (only slightly) flawed Fisheye lens portrait.
Up Next: Generating Human Faces with Stable Diffusion Dream Studio
- Pages:
- 1
- 2