snake wearing a crown slithering through a grassy plain
An impressionist oil painting of a snake wearing a crown
Borg queen
Borg queen, digital art
Systems as sophisticated as that (and future improvements to come along) are going to be insanely disruptive. Whole lot of artists are going to be looking for new jobs.
I’m wondering how clever it can be in understanding prompts. For example, say, if you asked for “a man in the style of van gogh holding a painting of a woman in the style of picasso”, would it render the man like a Van Gogh painting and the woman like Picasso? Craiyon gets partway there, but doesn’t quite get it.
Not very. It is already established that it has trouble with parsing multiple entities, e.g. if you ask for a blue apple in a bowl of oranges, some of the time you will get an orange in a bowl of blue apples.
Supposedly Google’s version has mitigated this problem to a large extent, allowing you to specify complex and detailed prompts, as well as introducing further refinements to the algorithm itself.
Anyway, here is what it did. I did not try refining any of the results.
PS the original beta gave you 10 variations; then it was reduced to 6; now it’s 4…
Yeah, that, while being more coherent, is actually further off course than Craiyon. That first one is pretty hilarious though. Thanks for giving it a try.
OK, I found this link explaining some Dall-E prompts that are known to work:
Maybe it is also valid for all the other CLIP-based models.
I tried a number of them in Craiyon and they worked (within the quality limitations of that system).
One of the mods that works is “contact sheet”. At the 256x256 resolution of Craiyon the resulting images are way too small, but with the higher resolution you are getting on Dall-E 2, there might be enough pixels per image to be meaningful, if you want to see multiple takes on a prompt at once. (Which you then can supersample on another site, if you wish.)
min-dalle — seems to essentially be a fork of Dall-E Mini?
Borg queen, digital art
a man in the style of van gogh holding a painting of a woman in the style of picasso
One of the (Nightcafe?) Stable Diffusion mystery models is supposed to drop on github soon, remains to be seen if they will release pre-trained checkpoints
I haven’t had time lately to play with my pretty pictures on nightcafe or look into these other websites.
Is there somewhere else I should be logging on every day to get credits for the future though?
The output seems to be about the same. But it is nice to be able to generate 25 versions at once instead of 9. Of course what would really be nice would be the ability to render in higher resolutions than 256x256.
Trying “child in halloween costume” crossed my mind for some reason. I did Studio Ghibli, Tim Burton, Ralph McQuarre, Wayne Barlowe, Michael Whelan, Margaret Keane, Mark Ryden, Picasso, van Gogh, Junji Ito, Anne Leibovitz. Also did “children” a couple of times, but the resolution is too low to even bother. I suspect Anne Leibovitz would be a good mod for Dall-E 2.
Nothing that I know about.
If you look at the code, it does not look like you can change the render resolution without re-training the model. I do not doubt there will be some open-source trained Dall-E-2’s and Imagens within the year, though (hopefully at useful resolutions).
I tried putting together a complicated scene, I mean with a reasonably long string of modifiers, in Dall-E, and while it basically worked in that all the stuff I asked for was present, either I still do not know how to use it right or that already pushes it to the limit, because I kept running into issues like the hands would not look quite right, or some objects were rendered without sufficient detail in what was supposed to be a photorealistic render, etc. The result was burning through all the free credits, so I can’t test any mods ATM. My earlier experiments appear to show that, generally speaking, style mods work if the artist/style was sufficiently represented in the training data.
Another Google one, with no diffusion:
https://parti.research.google/
Maybe a good candidate for those complex photorealistic (or stylized) scenes with multiple elements described by a complex prompt. Doesn’t look like it will run on your current consumer graphics card.
Love Spongebob Godzillapants.
Min-dalle has a specific interpretation of the word “cute”. Go over there and try it (that word alone).
Also try “ugly cute”.
(“Ugly man”/“ugly woman” are quite interesting, too.)
Weird AI art is not to be confused with Weird Al art.
If that’s too confusing:
Weird AI art is not to be confused with Weird Al art.
I also tried ‘cute woman’, and one of them had a cute Glasgow smile.
I’ll thank you to refer to me as Grand Master from now on.
1000 publications and 50 credits. And of course the title.
Ugly scary cute photo
Beautiful hideous
(Each 4 blocks of 25 images combined.)
This may sound like an egotistical question but is there a way I can get my published pics to show up in my feed?
