In this video, a computer invents faces that look like celebrities. They're 100% artificial, created by extracting information from a huge database of real celebrity images. (Link to video)
The faces slowly morph from one individual to another, shifting from male to female, melting gradually from one ethnicity to another. The only thing that stays constant is the position of the eyes.
Some faces almost look plausible, even though they're synthetic.
But there are in-between stages that look monstrous, bizarre, and sometimes beautiful. Textures become liquid or crusty like lava. The transitional images appear to be creative interpolations, rather than linear in-betweens.
The images are produced by an artificial-intelligence algorithm called a generative adversarial network (GAN).
One algorithm is optimized to generate images while another is fine-tuned to distinguish a plausibly real face from a fake one. The computer is trained using this "creator vs. critic" dynamic, starting with low resolution images, and adding detail and resolution in stages.
This kind of algorithm can also be applied to objects. If you start this video at 4:00, it shows computer-invented forms that morph into each other, but still stay within certain categories. (YouTube)
For example, the synthetic images above are generated within the categories of horse, sofa, bus, church/outdoor, and bicycle.
----
Read the scientific paper
For example, the synthetic images above are generated within the categories of horse, sofa, bus, church/outdoor, and bicycle.
----
Read the scientific paper
Publication: Progressive Growing of GANs for Improved Quality, Stability, and Variation
Authors: Tero Karras, Timo Aila, Samuli Laine, Jaakko Lehtinen (NVIDIA and Aalto University)
Previously on GJ
Text-to-image synthesis
Authors: Tero Karras, Timo Aila, Samuli Laine, Jaakko Lehtinen (NVIDIA and Aalto University)
Previously on GJ
Text-to-image synthesis
Wow. Watched about 11 minutes. Strongly reminiscent of certain experiences from the years 1967 - 1972. Found my breathing fell into synch with the rhythm of the shifts.
ReplyDeleteSteve, yes, we don't need drugs anymore to experience waking hallucinations.
ReplyDeleteLooks like dark souls facegen
ReplyDeleteThis comment has been removed by the author.
ReplyDeleteSomething that amazes me about computers:
ReplyDeleteOnce a computer can do something, it can easily do it a million times. Once a computer is programmed to morph two faces, it has no difficulty morphing hundreds, thousands, perhaps millions of faces. Is there a name for this sort of phenomenon?
In a strange way that may only be partially relevant, it reminds me of a scene in Disney's Sword In The Stone. After countless brawny men have repeatedly failed to pull the sword from the stone, a scrawny boy Arthur miraculously does so without effort. The crowd returns the sword to the stone, reasoning that a boy can't be the next king. Then steps up Kay, a man of great physical strength who failed before Arthur's success. He dismisses the challenge now saying, "Anyone can pull it once it's been pulled."
Do the accomplishments of computers ruin the miracle and spoil the challenge for the rest of us?
This is all very Kool and Trippy (and I do enjoy the occasional peak-experience and related time-flow distortion) but I am skeptical:
ReplyDelete“invents faces that look like celebrities”
and
“They're 100% artificial”.
I reviewed the article, and my interpretation is that they started w a data-base of celebrity photos of in-the-wild low-rez images and applied their magical-manipulations in order do obtain a consistently well oriented sampling of high-rez images to morph.
While the Sourcerer’s Apprentice’s magic-spell is quite impressive, and the millions of walking brooms mesmerizing, Mickey started with a real broom… Just saying. –RQ
From: ‘Progressive Growing of GANs for Improved Quality, Stability, and Variation’ (page 14)
C) CELEB A-HQ DATASET
In this section we describe the process we used to create the high-quality version of the CELEB A dataset, consisting of 30000 images in1024 × 1024 resolution. As a starting point, we took the collection of in-the-wild images included as a part of the original CELEB A dataset. These images are extremely varied in terms of resolution and visual quality, ranging all the way from 43 × 55 to 6732 × 8984. Some of them show crowds of several people whereas others focus on the face of a single person – often only a part of the face. Thus, we found it necessary to apply several image processing steps to ensure consistent quality and to center the images on the facial region.
Jim, I agree. As with the Fosbury Flop in high jumping, it literally raised the bar for everyone, setting a new standard that everyone reached. Also I find it amazing how quickly we accept the accomplishments of machine learning systems, and the wonder of it all rapid wears off. What I worry about is that we will lose our confidence in basic human skills such as wayfinding, remembering names, basic reasoning—even planning our next actions— and that we end up taking on the role of the subdued spouse that you see in some marriages where one of the partners is so adept and outspoken that their mate becomes mute.
ReplyDeleteRoberto, You're right, but I think I mentioned that they extracted information from a database of real celebrity images. However, even though they're drawing on existing imagery, they're creating novel specific faces that don't exist as such, and they're doing it by "growing" them through stages of ever greater complexity. I think it was important to distinguish this process from some kind of photo enhancement algorithm, or the older digital morphing techniques that transitioned in a linear way from one photo to another, which existed as far back as the late 1980s.
Jim, interesting point about the "subdued spouse" effect. Personally, I have a natural inclination to strive for independence. That may sounds conceited, but I'm not tooting my own horn. In fact, it's worth noting that this inclination sometimes makes it difficult for me to delegate tasks. While reliance can be a weakness, it can also be the foundation for building trust, promoting teamwork, developing young talent, or increasing productivity. I've developed a guideline that works for me: Never delegate any task that you couldn't do yourself if need be. Where the computer fits into that is a tougher question...
ReplyDelete