I predict that you think artwork created with StyleGAN by definition cannot have artistic merit on its own.
Which is amusing because when people look at StyleGAN artwork and they don’t realize it, like my anime faces, they often quite like it. Perhaps they just haven’t seen anime faces drawn by a true Scotsman yet.
When TWDNE went up, I asked ‘how long will I have to read and mash refresh before I see a cute face with a plot I would probably be willing to watch while bored at 2am’ The answer was ‘less than 10minutes’, and this is either commentary on the effectiveness of the tool, or on my (lack of?) taste.
I have a few pieces of artwork I’ve made using StyleGAN that I absolutely love, and absolutely could not have made without the tool.
When I noticed a reply from ‘gwern’, I admit was mildly concerned that there would be a link to a working webpage and a paypal link, I’m pretty enthusiastic about the idea but have not done anything at all to pursue it.
Do you think training a language model, whether it is GPT-2 or a near term successor entirely on math papers could have value?
It seems that my ignorance is on display here, the fact that these papers are new to me shows just how out of touch with the field I am. I am unsurprised that ‘yes it works, mostly, but other approaches are better’ is the answer, and should not be surprised that someone went and did it.
It looks like the successful Facebook AI approach is several steps farther down the road than my proposal, so my offer is unlikely to provide any value outside of the intellectual exercise for me, so I’m probably not actually going to go through with it—by the time the price drops that far, I will want to play with the newer tools.
Waifulabs is adorable and awesome. I’ve mostly been using style transfers on still life photos and paintings, I have human waifu selfie to anime art on my to do list but it has been sitting there for a while.
Are you planning integration with DeepAnime and maybe WaveNet so your perfect waifus can talk? Though you would know if that’s a desirable feature for your userbase better than I would...
On the topic, it looks like someone could, today, convert a selfie of a partner into an anime face, train wavenet on a collection of voicemails, and train a generator using an archive of text message conversations, so that they could have inane conversations with a robot, with an anime face reading the messages to them with believable mouth movements.
I guess the next step after that would be to analyze the text for inferred emotional content (simple approaches with NLP might get really close to the target here, pretty sure they’re already built), and warp the voice/eyes for emotional expression (I think WaveNet can do this for voice, if I remember correctly?
Maybe a deepfake type approach that transforms the anime girls using a palatte of a set of representative emotion faces? I’d be unsurprised if this has already been done, though maybe it’s niche enough that it has not been.
This brings to mind an awful idea: In the future I could potentially make a model of myself and provide it as ‘consolation’ to someone I am breaking up with. Or worse, announce that the model has already been running for two weeks.
I suspect that today older style, still image heavy anime could probably be crafted entirely using generators (limited editing of the writing, no animators or voice actors), is there a large archive of anime scripts somewhere that a generator could train on, or is that data all scattered across privately held archives?
Which is amusing because when people look at StyleGAN artwork and they don’t realize it, like my anime faces, they often quite like it. Perhaps they just haven’t seen anime faces drawn by a true Scotsman yet.
When TWDNE went up, I asked ‘how long will I have to read and mash refresh before I see a cute face with a plot I would probably be willing to watch while bored at 2am’ The answer was ‘less than 10minutes’, and this is either commentary on the effectiveness of the tool, or on my (lack of?) taste.
I have a few pieces of artwork I’ve made using StyleGAN that I absolutely love, and absolutely could not have made without the tool.
When I noticed a reply from ‘gwern’, I admit was mildly concerned that there would be a link to a working webpage and a paypal link, I’m pretty enthusiastic about the idea but have not done anything at all to pursue it.
Do you think training a language model, whether it is GPT-2 or a near term successor entirely on math papers could have value?
Oh, well, if you want to pay for StyleGAN artwork, that can be arranged.
No, but mostly because there are so many more direct approaches to using NNs in math, like (to cite just the NN math papers I happened to read yesterday) planning in latent space or seq2seq rewriting. (Just because you can solve math problems in natural language input/output format with Transformers doesn’t mean you should try to solve it that way.)
Thank you for this!
It seems that my ignorance is on display here, the fact that these papers are new to me shows just how out of touch with the field I am. I am unsurprised that ‘yes it works, mostly, but other approaches are better’ is the answer, and should not be surprised that someone went and did it.
It looks like the successful Facebook AI approach is several steps farther down the road than my proposal, so my offer is unlikely to provide any value outside of the intellectual exercise for me, so I’m probably not actually going to go through with it—by the time the price drops that far, I will want to play with the newer tools.
Waifulabs is adorable and awesome. I’ve mostly been using style transfers on still life photos and paintings, I have human waifu selfie to anime art on my to do list but it has been sitting there for a while.
Are you planning integration with DeepAnime and maybe WaveNet so your perfect waifus can talk? Though you would know if that’s a desirable feature for your userbase better than I would...
On the topic, it looks like someone could, today, convert a selfie of a partner into an anime face, train wavenet on a collection of voicemails, and train a generator using an archive of text message conversations, so that they could have inane conversations with a robot, with an anime face reading the messages to them with believable mouth movements.
I guess the next step after that would be to analyze the text for inferred emotional content (simple approaches with NLP might get really close to the target here, pretty sure they’re already built), and warp the voice/eyes for emotional expression (I think WaveNet can do this for voice, if I remember correctly?
Maybe a deepfake type approach that transforms the anime girls using a palatte of a set of representative emotion faces? I’d be unsurprised if this has already been done, though maybe it’s niche enough that it has not been.
This brings to mind an awful idea: In the future I could potentially make a model of myself and provide it as ‘consolation’ to someone I am breaking up with. Or worse, announce that the model has already been running for two weeks.
I suspect that today older style, still image heavy anime could probably be crafted entirely using generators (limited editing of the writing, no animators or voice actors), is there a large archive of anime scripts somewhere that a generator could train on, or is that data all scattered across privately held archives?
What do you think?