Is it broken right now? Makes my fans spin but it never finishes.
DALL·E: Introducing Outpainting
51–60 of 302 posts
Re: DALL·E: Introducing Outpainting
#52I can't help but feel like they're adding this at this particular point since Stable Diffusion has announced they're releasing their 'inpainting' model next week.
I really doubt it's related at all, though everyone would think it looks that way. SD has only been out a week and this feature would have taken much more than that to build, test, enroll demo users, make a webpage for, etc.
I can't prove it of course but it wouldn't surprise me if they had this pretty much done already long ago (dall-e has been out for several months at this point). The actual implementation doesn't look like it'd take more than a few days to code honestly (and they've got quite competent coders over there). Only speculation of course.
Re: DALL·E: Introducing Outpainting
#53Re: DALL·E: Introducing Outpainting
#54Feels like a race to the bottom. More features, lower cost, every week. No idea where it’ll level out, but I like it. Just bought some more Dalle credits today because it’s so much fun. This is a revolution in ‘art technology’ it’s like Steve Job’s bicycle for the mind. Best I could do a month ago was a stick figure in MS Paint, but now..
Stable Diffusion is arguably better, has more features, and is free. OpenAI can't compete with free. Even if you don't want to take the 30 seconds to set it up in a free Google Colab environment, the paid DreamStudio version is still half the price of Dalle.
Do you know the best Google Colab tutorial / repo?
Re: DALL·E: Introducing Outpainting
#55Feels like a race to the bottom. More features, lower cost, every week. No idea where it’ll level out, but I like it. Just bought some more Dalle credits today because it’s so much fun. This is a revolution in ‘art technology’ it’s like Steve Job’s bicycle for the mind. Best I could do a month ago was a stick figure in MS Paint, but now..
Stable Diffusion is arguably better, has more features, and is free. OpenAI can't compete with free. Even if you don't want to take the 30 seconds to set it up in a free Google Colab environment, the paid DreamStudio version is still half the price of Dalle.
dreamstudio is also waaaay faster than openai. generally a second or two for 512x512 at 50 steps.
Re: DALL·E: Introducing Outpainting
#56This is cool and useful! Putting "Girl with a Pearl Earring by Johannes Vermeer" in the kitchen in 2022 does not look good!
Because depicting a woman in a kitchen is perpetuating the pernicious male patriarchy? Sorry we're not doing that. You might find some reception for this sort of thing on Twitter though.
Re: DALL·E: Introducing Outpainting
#57Re: DALL·E: Introducing Outpainting
#58A few weeks ago I was skeptical that this technology would get past the emotional response we get from procedurally generated game environments, but I've been convinced otherwise. The emotional response I get from some of the best of these images are novel and thought provoking. Makes me wonder what percent of what makes us human is now algorithmically solved....
Maybe. I've been often reminded lately of Herbert Goldstone's "Virtuoso" (1958):
http://elateachers.weebly.com/uploads/2/7/0/1/27012625/virtu...
Re: DALL·E: Introducing Outpainting
#59I think Dall-E would benefit from a "sketch-based" prompt in addition the text based. This was mindblowing - https://andys.page/posts/how-to-draw/
Re: DALL·E: Introducing Outpainting
#60Why would Google hold back on releasing Imagen if there are competitors that are publicly available already? Imagen isn't special anymore.
A few possible theories, some might be mutually exclusive:
Organizational scar tissue making them more risk averse about the PR risks of letting the genpop use AI generation tools, and create something offensive. With the safe assumption that Google will get blamed, not the user.
Fear of government regulation on AI if they don't self-regulate.
No need to actually release it, since this isn't the core business but just research. (While openai needs to actually create the business.) Corollary: more to lose -- a scandal around offensive content will not hurt openai's non-existent other businessess. It might make some advertisers pull their ads from Google.
The opportunity cost of building a self serve platform is too high. (Can't pull in people writing those kind of apps from projects with more commercial importance. Can't make the ML researchers do that.)
They misjudged how much demand there would be, and thought that building a platform would not be useful for a few years. And if it now turns out to actually be a great business it'll now take them a year to productionize and build a platform.
Their compute requirements are so high that selling access is not viable, the costs are prohibitive for real users.
It's not that different from e.g. self driving cars. Pretty obviously they had better tech from early on, but were not willing to take the risks that Tesla was.