Category: aiarts

For this week’s assignment I played around with BigGAN. In class we experimented with how truncation would affect single images, but I wondered how it would affect the video animation from one object into another.

I wanted to morph an object into another one that is already similarly shaped, so at first I chose guacamole and ice cream on truncation 0.1. This turned out to be…really disgusting looking.

Video: https://drive.google.com/file/d/1mAewM63SA8vT1rez3u7co2fDE3eP7d0C/view?usp=sharing

For some reason the guacamole didn’t really seem to be changing at all at the beginning, and when it did begin to morph into ice cream it just looked like moldy food. The ending ice cream picture also didn’t really look like ice cream.

So…I decided to change the objects to a cucumber and a burrito. This worked a lot better. I created four videos, one with truncation 0.2, one with 0.3, one with 0.4 and one with 0.5. I then put these into a collage format so you could see the differences between all of them:

Though it’s subtle, you can definitely tell that there is a difference between the four videos. Theoretically, the top left corner is 0.2, the top right corner is 0.3, the bottom left is 0.4 and the bottom right is 0.5, however I am not super well-versed in video editing and when I put this together in iMovie it was hard to tell which one was which.

Problems:

This week we have learned BigGan and deepDream and I have done some exploration on BigGan in the class. Thus, for this week’s assignment I choose to explore more about deepDream. I met some errors when I train the model. One is that I ran the python script without installing keras and tensorflow so that I couldn’t run the script. The other one is that I didn’t change the a+=b to a=b+. After I changed that, the model could run successfully.

Process:

This is my original image:

The default version looks like this:

Then I changed the step from 0.01 to 0.1 and the result looked vague.

The next step I changed the number of scales from 3 to 6 and the result looked like this: it looked more like the original one except for the ring in the middle.

Then I changed the size ratio from 1.4 to 1.8 and the result looked like this:

Then I changed the max-loss and found this change doesn’t have great effect on the result:

For this week’s assignment, I played with DeepDreaming a little to generate images. I ran the model with the image below and changed the settings to see how each setting influences the output image. The difficulty I met when running the model is that when I first run the deep_dream.py file the error message “RuntimeError: Variable += value not supported.” showed up. I later changed the code in the file into “loss = loss+coeff * K.sum(K.square(x[:, :, 2: -2, 2: -2])) / scaling” and it worked. But I still don’t really know why it couldn’t run when the code is ” lose += ….” . But the program did begin to generate output after I changed that. The results are as follows.

The input image:

The output image when no setting is changed:

The output when I changed all “feature settings” into 2:  

(The changes seem to be smaller than the original image.)

The output when I changed step into 0.1 from 0.01 and left anything else unchanged:

Overall this was a super fun experience. The image it generated has a really powerful and interesting feeling. Even though I haven’t seen  how I can apply this technology in my final project, still I find this the deep dreaming is a very powerful technology.