3 ms·

I noticed there are very few black people and none with very dark skin. But there are lots of people of many other ethnicities. Was dark skin avoided because it

I noticed there are very few black people and none with very dark skin. But there are lots of people of many other ethnicities. Was dark skin avoided because it's inherently harder for the algorithm?

goodside 7y ago
The training data contains examples of very dark skin: https://github.com/NVlabs/ffhq-dataset/blob/master/ffhq-teaser.png https://github.com/NVlabs/ffhq-dataset/blob/master/ffhq-teas...
- OscarCunningham 7y ago
  Interesting. I just refreshed the site 100 times to get a large sample size and the darkest skinned image it produced was this girl (https://i.imgur.com/ggJIG94.png https://i.imgur.com/ggJIG94.png). Her skin is much lighter than many of the people in that sample of the training data. Assuming that that sample is representative of all the training data, it must be that the algorithm is 'choosing' not to produce images of dark skinned people. Perhaps because it isn't very good at them and 'knows' this? I wonder if there are other features it's deliberately avoiding?
  - iamnothere 7y ago
    > Assuming that that sample is representative of all the training data, it must be that the algorithm is 'choosing' not to produce images of dark skinned people. Perhaps because it isn't very good at them and 'knows' this? I wonder if there are other features it's deliberately avoiding? An algorithm would not have this kind of insight. (This is just a GAN trained on face data.) To say that an algorithm like this is "deliberately" doing anything is a misunderstanding. Most generators like this that I've seen, like GPT-2, require you to provide a "seed" of some sort, such as a sentence fragment. They then build off of this seed. I don't know if this is implemented that way, but if so, perhaps the developer provided a set of seed data that leads to this result. There may also be a sort of "averaging" involved (see another commenter's note about the women having similar features), and depending on both the training data and the seed, this may result in a preference towards certain features. Edit: it's StyleGAN.
  - goodside 7y ago
    You can see several examples of dark-skinned outputs around timestamp 2:41 in this video: https://youtu.be/c-NJtV9Jvp0 https://youtu.be/c-NJtV9Jvp0
gwern 7y ago
The original FFHQ was designed to cover a lot of ethnicities, but maybe it didn't cover very dark skin enough? Alternately, it might be related to the 'truncation trick': images at the 'edge' of the latent space get dropped to assure high quality at the cost of covering the entire distribution. If very black skin got mapped to a compact cluster at the edge, say, the truncation trick could hypothetically lead to few such samples despite representation. To test this you would generate images with very high psi to see if they come out more often.

I noticed there are very few black people and none with very dark skin. But there are lots of people of many other ethnicities. Was dark skin avoided because it's inherently harder for the algorithm?

goodside 7y ago

The training data contains examples of very dark skin: https://github.com/NVlabs/ffhq-dataset/blob/master/ffhq-teaser.png https://github.com/NVlabs/ffhq-dataset/blob/master/ffhq-teas...

OscarCunningham 7y ago
Interesting. I just refreshed the site 100 times to get a large sample size and the darkest skinned image it produced was this girl (https://i.imgur.com/ggJIG94.png https://i.imgur.com/ggJIG94.png). Her skin is much lighter than many of the people in that sample of the training data. Assuming that that sample is representative of all the training data, it must be that the algorithm is 'choosing' not to produce images of dark skinned people. Perhaps because it isn't very good at them and 'knows' this? I wonder if there are other features it's deliberately avoiding?
- iamnothere 7y ago
  > Assuming that that sample is representative of all the training data, it must be that the algorithm is 'choosing' not to produce images of dark skinned people. Perhaps because it isn't very good at them and 'knows' this? I wonder if there are other features it's deliberately avoiding? An algorithm would not have this kind of insight. (This is just a GAN trained on face data.) To say that an algorithm like this is "deliberately" doing anything is a misunderstanding. Most generators like this that I've seen, like GPT-2, require you to provide a "seed" of some sort, such as a sentence fragment. They then build off of this seed. I don't know if this is implemented that way, but if so, perhaps the developer provided a set of seed data that leads to this result. There may also be a sort of "averaging" involved (see another commenter's note about the women having similar features), and depending on both the training data and the seed, this may result in a preference towards certain features. Edit: it's StyleGAN.
- goodside 7y ago
  You can see several examples of dark-skinned outputs around timestamp 2:41 in this video: https://youtu.be/c-NJtV9Jvp0 https://youtu.be/c-NJtV9Jvp0

gwern 7y ago

The original FFHQ was designed to cover a lot of ethnicities, but maybe it didn't cover very dark skin enough? Alternately, it might be related to the 'truncation trick': images at the 'edge' of the latent space get dropped to assure high quality at the cost of covering the entire distribution. If very black skin got mapped to a compact cluster at the edge, say, the truncation trick could hypothetically lead to few such samples despite representation. To test this you would generate images with very high psi to see if they come out more often.