4 ms·

These generated "anime" pictures look overfitted to me. It's not convincing me that it's in any way original. I bet large parts of these pictures actually do ex

by tinco 6y ago

These generated "anime" pictures look overfitted to me. It's not convincing me that it's in any way original. I bet large parts of these pictures actually do exist, and it's just mashing its memories of these pictures together, instead of actually having internalized how to draw manga pictures. Maybe the source dataset was too small, and it was not trained strictly enough, or it should have actually trained on anime stills instead of these Moe manga style drawings. The state of the art, for example that site that generates images from sentences is much more impressive to me.

Blikkentrekker 6y ago
At the very least it's mashing them together in such a way that about 50% of the time they form anatomically correct humans. So at the very least it is rotating and shrinking limbs in a three dimensional way and internalized that. The other 50% of the time it goes wrong, and the result is either a human that is not anatomically correct, or something that doesn't even seem to be a two-dimensional projection of three-dimensional Euclidean space.
isoprophlex 6y ago
Take into account that the train set is very heterogeneous compared to eg. very well lighted, centered (celebrity) faces or whatever went into thisfacedoesnotexist. From the detailed writeup: > Broadly, his StyleGAN2-ext increases the model size and disables regularizations, which are useful for restricted domains like faces but fail badly on more complex and multimodal domains.
sillysaurusx 6y ago
The dataset was 2.4 million images. I assure you that a ~1GB stylegan model isn’t overfitting on that. The fact that it looks overfitted is a sign that the model is legit. The interpolation videos are much more convincing. https://images-ext-2.discordapp.net/external/4Om03hvtvjDhT4culcJGlci43yYmFhCidoK_wmqiuZk/https/i.4cdn.org/a/1611088549640.gif https://images-ext-2.discordapp.net/external/4Om03hvtvjDhT4c...
- pontifier 6y ago
  That is truly amazing. It makes AI feel like a wild thing. It's got something... A creative stamina and output that is astounding.
- numpad0 6y ago
  Since discovering an anime-manga franchise of biped daikons existed a decade prior to DALL-E I’m not so much convinced...
freeone3000 6y ago
I think this is more an indication of homogeneity in current anime art than an indictment of the GAN.
petters 6y ago
There is a slider that allows you to view the same image in many versions. Does that convince you otherwise?
gwern 6y ago
> The state of the art, for example that site that generates images from sentences is much more impressive to me. Come on, DALL-E came out like last week - give us a break! That was a major research project contributed to by at least 11 fulltime OpenAI researchers using 400 million images; and this was like 3 hobbyists in a Discord occasionally.
- tinco 6y ago
  I mean obviously it's very interesting to look at and no small feat. I suppose I wrote my comment in the wrong context, my apologies. If 3 hobbyists moonlighting in a Discord could casually match Dall-E it would definitely spell utter doom for us all :P Anyway, I did the most hated thing on HN, casually dismissing someone's hard work or careful thoughts. I'll go chew on some sand now.

by tinco 6y ago

Blikkentrekker 6y ago

At the very least it's mashing them together in such a way that about 50% of the time they form anatomically correct humans. So at the very least it is rotating and shrinking limbs in a three dimensional way and internalized that. The other 50% of the time it goes wrong, and the result is either a human that is not anatomically correct, or something that doesn't even seem to be a two-dimensional projection of three-dimensional Euclidean space.

isoprophlex 6y ago

Take into account that the train set is very heterogeneous compared to eg. very well lighted, centered (celebrity) faces or whatever went into thisfacedoesnotexist. From the detailed writeup: > Broadly, his StyleGAN2-ext increases the model size and disables regularizations, which are useful for restricted domains like faces but fail badly on more complex and multimodal domains.

sillysaurusx 6y ago

The dataset was 2.4 million images. I assure you that a ~1GB stylegan model isn’t overfitting on that. The fact that it looks overfitted is a sign that the model is legit. The interpolation videos are much more convincing. https://images-ext-2.discordapp.net/external/4Om03hvtvjDhT4culcJGlci43yYmFhCidoK_wmqiuZk/https/i.4cdn.org/a/1611088549640.gif https://images-ext-2.discordapp.net/external/4Om03hvtvjDhT4c...

pontifier 6y ago
That is truly amazing. It makes AI feel like a wild thing. It's got something... A creative stamina and output that is astounding.
numpad0 6y ago
Since discovering an anime-manga franchise of biped daikons existed a decade prior to DALL-E I’m not so much convinced...

freeone3000 6y ago

I think this is more an indication of homogeneity in current anime art than an indictment of the GAN.

petters 6y ago

There is a slider that allows you to view the same image in many versions. Does that convince you otherwise?

gwern 6y ago

> The state of the art, for example that site that generates images from sentences is much more impressive to me. Come on, DALL-E came out like last week - give us a break! That was a major research project contributed to by at least 11 fulltime OpenAI researchers using 400 million images; and this was like 3 hobbyists in a Discord occasionally.

tinco 6y ago
I mean obviously it's very interesting to look at and no small feat. I suppose I wrote my comment in the wrong context, my apologies. If 3 hobbyists moonlighting in a Discord could casually match Dall-E it would definitely spell utter doom for us all :P Anyway, I did the most hated thing on HN, casually dismissing someone's hard work or careful thoughts. I'll go chew on some sand now.