top of page
shuang's itp jorney
welcome to my chelsea hotel
chelsea is a name given by naive shuang
hope you enjoy your stay
class two assignment
feb 15nd
The style photo is a sketch I did on my undergrad thesis. I took the content photo of a friend of mine using 35mm film. The result does not necessarily look like a sketch, but the glitchiness is also quite intriguing. I also tried increasing the max size and the result is less glitchy. Below are the init images and the two results.
Text to image has some great potential for my creation. I tried a bunch of text with different weights. This progression on the attached is one of the most representative example.
For my experiment, the images style and clarity kind of goes back and forth. Some intermediate creations are more appealing to me than the final result. The images plateaued after 200 epoch. The images on left is the earlier generation. As you can also see, the following ones are very similar with variation.
The prompts are "Cat is mad at mom | Cyanotype Image". To be completely honest, I am inspired by the Sofia Crespo's cyanotype works.
class one
class one
feb 7nd
For starters, I tried the circular loop with the FreaGAN-6000-steps pickle file using the provided script by @aydaogman and the tutorial. Other than messing with the parameters, the main change I added is to change how the original script save the file and wrote a piece of hard code that will display the video on google colab directly. To achieve the added feature, I forked the GitHub repository onto my page and changed the code directly there.
Though different parameters will give different results, the main visuals of the final video output are still heavily based on its input pickle file. To do more different things with this script in the future, I plan on collecting my own data and compressing my own pickle file with them.
For some weird reason, I couldn't get the image ratio correct for this given color-filling script. I found the original notebook here at qblocks.
The result is very realistic though. The image on the right is the original photo that I took with a black and white film camera. I have vivid memories of what happened on the day but have never seen this image in colors before. I believe I will be able to tweak it a bit more if I have more time to work on this. For this little experiment, I crossed into out-of-memory error so many times and had to reboot the notebook every time in order for this to work.
The above are the two successful trials I did this week. I tried CLIP and StyleGAN. Different CLIP keep telling me I either do not have pydiffvg or missing some other packages. StyleGAN on the other hand is giving me memory errors.
All in all, the codes are much more hidden than I imagined. I was hoping I could look deeper into the construction and the training, but, at this point, I would have to be patient.
bottom of page