Kind of curious, how do most people run their notebooks? I've been doing it in the browser for a while, but then I realized I could also do it in VSCode. After installing the necessary extensions, I realized that VSCode was ridiculously slow (and was eating up a ton of memory.)
From https://twitter.com/advadnoun/status/1348375026697834496:
>colab.research.google.com/drive/1FoHdqoqKntliaQKnMoNs3yn5EALqWtvP?usp=sharing
>
>I'm excited to finally share the Colab notebook for generating images from text using the SIREN and CLIP architecture and models.
>
>Have fun, and please share what you create!
Change the text in the above notebook in the Params section from "a beautiful Waluigi" to your desired text.
Reddit post #1 about SIREN. Post #2.
Update: The same parameter values (including the desired text) can (and seemingly usually do) result in different output images in different runs. This is demonstrated in the first two examples later in this post.
Update: Steps to follow if you want to generate a different image with the same Colab instance:
- Click menu item Runtime->Interrupt execution.
- Save any images that you want to keep.
- Change parameter values if you want to.
- Click menu item Runtime->Restart and run all.
Update: The developer has changed the default number of SIREN layers from 8 to 16.
Update: This project can now be used from the command line using this code.
Example: This is the 6th image output using notebook defaults after around 5 to 10 minutes of total compute for the text "a football that is green and yellow". The 2nd image (not shown) was already somewhat close to the 6th image, while the first image (not shown) looked nothing like the 6th image. The notebook probably could have been run much longer to try to generate better images; the maximum lifetime of a Colab notebook is 12 hours for the free version (source). I did not cherry-pick this example; it was the only text that I tried.
https://preview.redd.it/r31mtn7phlb61.png?width=512&format=png&auto=webp&s=66560ea899e17b46cd6c83538c576de521b3e7ca
Update: I did a different run using the same parameters as above.
... keep reading on reddit ➡
