GANS – phil caridi

March 10, 2022

A little from column A and some from column B

Just a little bit of practice with two different GANS this week. I used a version of Next Frame Prediction (NFP) and StykeGan2-ADA from Derrick Schultz, both notebooks can be found below:

NFP: https://colab.research.google.com/drive/1bOuMWqnXty6F-RX6xO_uRJ9RFY7y-P1q?usp=sharing

StyleGan-ADA: https://colab.research.google.com/drive/19wmrT2NOcCAfJRWIc-DpXIP8NY0MJ9B-?usp=sharing

Both notebooks are pretty straight forward to use. For the NFP I used a trippy animation by Anthony Francisco and can be found here

For StyleGan-ADA I trained a custom dataset with microscopic images of cells, the model needs way more training but its nice to see where it stands currently.

StyleGan-ADA truncation

StyleGan-ADA interpolation noiseloop

StyleGan-ADA interpolation linear

NFP

February 14, 2022

The rabbit hole is quite deep

This week I mostly worked with text-to-images notebooks in Google Colab. I also tried some style transfer notebooks, but to be honest I don’t really find them that interesting. Maybe I’m not using them right but a lot of my outputs feel like they are just layered photoshop files.

Lets get started with text-to-image. First up

Hypertron 2.0

Prompt: A spaceship, by Hakashina Mokinato. A brutal Style.:1 | Trending on ArtStation, unreal engine. 8K 3D. Volumetric lighting.:0.75

I used 2000 steps (iterations); the imagenet_16384, coco, and wikiart_1024 models; and 480×480 to get the final image.

Google Colab Notebook:

https://colab.research.google.com/drive/1N4UNSbtNMd31N_gAT9rAm8ZzPh62Y5ud?usp=sharing

Final image after 2000 steps

Step #95

Step #1000

Step #200

Step #1250

Step #300

Step #1500

Step #500

Step #1750

Below are a selection of steps, as the steps progress it get to a point where the changes in image are less significant. If time was a factor I probably would have around 900 steps.

Illustrip 3D

I made two different trips with this notebook. Iused the same prompt as the Hypertron 2.0 to see the difference in the output.

Prompt: A spaceship, by Hakashina Mokinato. A brutal Style.:1 | Trending on ArtStation, unreal engine. 8K 3D. Volumetric lighting.:0.75 3d hyper real

I used 200 steps to create this video.

Google Colab Notebook

https://colab.research.google.com/github/eps696/aphantasia/blob/master/IllusTrip3D.ipynb

Step #1, this starts with noise

Step #150

Step #50

Step #175

Step #100

Step #200

Prompt: A colorful snake, by HR Giger. Blood surrounds it. Trending on ArtStation, unreal engine. 8K 3D. Volumetric lighting

I wanted to see what more steps (iterations) produced, for this video I used 2000 steps.

Google Colab Notebook

https://colab.research.google.com/github/eps696/aphantasia/blob/master/IllusTrip3D.ipynb

Again we start with noise, Step #1

Step #500

Step #150

Step #750

Step #375

Step #999

Zeotrope 5.5

For Zeotrope I used both text input and images.
Prompt: A grotesque face of evil, blood, deformed teeth.

Below are the image inputs I used.

Disco Diffusion 4.1

I started with a text prompt: Spiders crawl across the landscape in space, by HR Giger, Trending on artstation. 3d.

Google colab notebook

https://colab.research.google.com/drive/1sHfRn5Y0YKYKi1k-ifUSBFRNJ8_1sa39

NueralStyleTransferFor

For this style transfer I started with a content image and a style image. After I made the composite from the two images I also tried giving the content image a weight of 0 to create a new texture image. I did this twice once with 100 steps and 1000 steps.

Google colab notebook

https://colab.research.google.com/github/dvschultz/artificial-images/blob/master/neural_style_tf.ipynb

Content Image

Style Image

Output Image

500 iterations scale 0.5

500 Iterations Scale .5

Lucid 2D Style Transfer

Again with this style transfer I used a content image and a style image.

Google colab notebook

https://colab.research.google.com/github/tensorflow/lucid/blob/master/notebooks/differentiable-parameterizations/style_transfer_2d.ipynb

Content Image

Style Image

Output Image

February 6, 2022

Down the Rabbit Hole

This week I looked at seven different Machine Learning models in Google Colab Notebooks. Four of which I had success with the other three there were a number of errors I couldn’t get past. I’ll show examples of the ones I could get working and the error for the ones I couldn’t.

For a lot of the following work I used seed images generated with a notebook created by Jeramy Torman link:

https://colab.research.google.com/drive/1Zny3nZwzkGqzVd-PBQaEK48w9HC2SaL_?usp=sharing

LOOKING GLASS V1.1

First up we have Looking Glass, notebook link:

https://colab.research.google.com/drive/11vdS9dpcZz2Q2efkOjcwyax4oob6N40G

This uses both an input image and input text to generate multiple images file. There weren’t many parameters to play with but the input image and text made major differences in the output image. You can adjust the number of training epochs to change the similarity between images.

Trial #1

Flavor Text: Blood flows from ancient text
Image Name: 01608
Epochs: 50
Images Generated: 9
Image

Output Images

Trial #2

Flavor Text: The void stares back at you
Image Name: 01608
Epochs: 50
Images Generated: 9
Image

Output Images

Trial #3

Flavor Text: The void stares back at you
Image Name: 01625
Epochs: 25
Images Generated: 25
Image

Output Images

Trial #4

Flavor Text: The void stares back at you.
Image Name: 07114
Epochs: 25
Images Generated: 25
Image

Output Images

Trial #5

Flavor Text: Enemy of mine the omega.
Image Name: Day7of50
Epochs: 25
Images Generated: 25
Image

Output Images

Trial #6

Flavor Text: Collapsed across time and space.
Image Name: Day7of50
Epochs: 25
Images Generated: 9
I used the blend feature for this set, 50 additional images were used to create more variation across the output images

Image

Output Images

FLESH DIGRESSIONS

For the next set of experiments I used the “Flesh Digressions” colab notebook link:

https://colab.research.google.com/github/dvschultz/ml-art-colabs/blob/master/flesh_digressions.ipynb

I produced three different video outputs using three different .pkl files they can be found at the link below.

https://drive.google.com/drive/folders/1TtgUHAHH6mdXHeO12wQWWp7_Ev4VLxLn?usp=sharing

This video uses birdtogod.pkl

This video uses landscapetogod.pkl

This uses the visionaryart.pkl by Jeramy Torman.

NEURAL CELLULAR AUTOMATION

Next up is Neural Cellular Automation colab notebook link

https://colab.research.google.com/github/google-research/self-organising-systems/blob/master/notebooks/texture_nca_tf2.ipynb

I used the following image as an input

Output Video

DISCO DIFFUSION

I also tried a purely text to image model called “disco-diffusion”.
my input was the following:

“I’ve been crawling on my belly” Tool Lyric
“I think drugs have done some good things for us” Bill Hicks Quote
“prayed like a martyr dusk till dawn” Tool Lyric
“somniferous almond eyes” Tool Lyric
“blue color scheme” Descriptor

Output Image

I had the following errors for a few notebooks.

Deep Dream

Lucid Style Transfer

IllusTrip

I didn’t spent too much time trying to figure out the issues since there are so many notebooks to try. There are a few that I really want to get going like the illstrip and lucid style transfer.