News for hackers. Stuff that matters.
Re:Text to Image Generation by flowery on Monday March 29, 2021 @01:26AM (#26617982)
For some reason Colab let me run this all last night. Wow just wow. Eventually I got the feeling that it matters if I'm watching.

Each of these links has about 9 images and the prompts that made them. Sometimes the image does not look like an animal right off the bat, then it seems like you asked for something the network had to say.

https://postimg.cc/vDvYdBhC https://postimg.cc/XpHnpzT8 https://postimg.cc/HVspWgPn

That Syd Mead style ... mwa

Re:Text to Image Generation by indigodaddy on Monday March 29, 2021 @02:45AM (#26618370)
How long did it take for you.. I just did one text to image and Colab was churning away for a few hours but then choked after maybe 300ish image iterations saying something about some kind of space limit being reached..

Were you able to complete the 1050 iterations, and how long did it take?

Re:Text to Image Generation by indiv0 on Monday March 29, 2021 @03:13AM (#26618524)
Not the OP, but some tips:

- the model usually locks in within 200-300 iterations, so if you don’t like the result by then, retry

- in fact, you can tell if the model is off to a good start within 25-50 iterations and I encourage you to cherry-pick runs early and often; don’t be afraid to restart

- time to render depends on which GPU you get from colab, but I usually run the renders for 10 minutes a pop. About 1-2 minutes if I run them on a 3090 locally

- the prompt plays a big role in the quality of the result; “A painting of a dog playing fetch” will usually turn out better than “dog playing fetch”

- lucidrains/bigsleep produces better results generally than lucidrains/deepdaze (this is my subjective preference)

- the colabs linked to from the big-sleep GitHub repo produce poorer results than running them as a python package locally (this one might honestly be placebo)

Re:Text to Image Generation by Hard_Space on Monday March 29, 2021 @03:21AM (#26618557)
> - the prompt plays a big role in the quality of the result; “A painting of a dog playing fetch” will usually turn out better than “dog playing fetch”

However, it can get taken very literally, in that you might get a picture that features a frame around the painting.

Re:Text to Image Generation by indiv0 on Monday March 29, 2021 @03:27AM (#26618584)
Indeed!

Another thing that comes to mind as a corollary is that the AI seems to like being constrained in its outputs. So adding something like “in the style of Monet” to the end of the prompt will return much more coherent results.

Re:Text to Image Generation by Hard_Space on Monday March 29, 2021 @03:43AM (#26618648)
True. For the article, I experimented with a number of 'in the style of' prompts. Where there's a distinctive visual iconography with strong key features, BigSleep [1]does an amazing job of abstracting and reproducing that style. Besides artists, it also does very well with iconic movies like Blade Runner and The Matrix.

[1]The Deep-Daze author's most popular T2I mashup: https://github.com/lucidrains/big-sleep

https://colab.research.google.com/drive/1NCceX2mbiKOSlAd_o7I...

Re:Text to Image Generation by indigodaddy on Monday March 29, 2021 @03:33AM (#26618602)
Very helpful feedback, thanks very much!
Re:Text to Image Generation by indigodaddy on Monday March 29, 2021 @01:31AM (#26618007)
Wow! Fascinating images. The “Chinese ink painting made of poetry” one is particularly interesting and beautiful.