StyleGAN2 locally. Part 2: Infinite Latent Space Browser.

You can see the machine’s visual memory.

See also Part 1: StyleGAN in ComfyUI.

Infinite Latent Explorer

In the old days, if you were looking for a specific programme or app, you only had two choices:

• Wait until someone might program it (in case someone begins to develop it at all),
• or do it yourself (learn programming languages from scratch).

Both options took time, which neither of us had. Sometimes, wonderful apps were developed and you worked with them until the providers simply closed them down. Vendor lock-in. Like RunwayML did to ML Lab.

Their most fascinating tool was the Latent Space Browser:

You would train a GAN model on your dataset, open it in a browser, and then explore the Endless Latent Space full of machine vision. Then it was gone. Bard Ionson, objektpermamenz, myselle, me and many other artists were desperately looking for a similar solution until I just did it, of course not alone, but with Claude’s help — for years.

Now I did it… of course, not alone, but by vibe coding with Claude. In 2 days.

And it works! So, as a gift, I want to give you a functioning, locally running standalone Latent Browser Suite!

What does it contain?

The Tool Suite contains:

• Latent Browser — shows you a variety of samples from your GAN model
• Walk Explorer — creates endless interpolations, that you can export as videos
• Infinite Latent Map — the heart of the suite, enabling you to roam inside the machine’s memory endlessly.

You can also send images between apps, for example to discover similar latent images around your favourite one.

How to install it?

You can do it via my GitHub repository.

[embed]GitHub - Merzmensch/stylegan2-latent-explorer: 2D Latent Space Browser, Walk Explorer and Infinite… 2D Latent Space Browser, Walk Explorer and Infinite Map for StyleGAN2 models - Merzmensch/stylegan2-latent-explorergithub.com

The following data is contained within:

• index.html: the main page with app selection
• latent_browser.html: providing a bunch of random images
• latent_explorer.html: the Latent Image Walk
• latent_infinite.html: Infinite Latent Browser
• server.py: virtual server instance
• start.bat: starting the system with one click (don’t forget to change YOUR-USERNAME to your username)

You will also need:

• Miniconda (just the small version, not Anaconda). Do not “add to PATH”, and do not register as default Python during installation.
• Clone StyeGAN2-ADA-PyTorch into /stylegan2-explorer/
• To export videos, you will need ffmpeg.
• Put your GAN model in .pkl format into /stylegan2-explorer/models folder

All instructions can be found in ReadMe.

How does it work?

Once you have started the server, you can open it in your browser at http://localhost:5000.

Main page

Here you can select between the apps.

Latent Browser

1. Load your model (in my case, Mem.pkl).
2. Choose the Grid sizeM; the larger the grid, the longer the building time. However, you will have more details.
3. Truncation is a technique that generates higher-quality images with less diversity (lower truncation values) versus lower-quality images with more aesthetical variations.

Left: truncation set on 0.10, right: truncation set on 1.35

You can see a 4x4 grid of images here, and you have two options:

1. You can select an image by pointing to it with the mouse and clicking immediately. The selected image will then be visible in the top right-hand corner.
2. Alternatively, if you wait three seconds and move the mouse over an image, you will move within the latent space and see the dynamic changes in the preview window in the top right corner.
3. Clicking “s” will save the image with the filename [your dataset]_[SEED].png

But you also can send the image to Walk Explorer or Infinite Map.

Walk Explorer

This tool allows you to create a Latent Walk, i.e. interpolation from one latent image to another, while it is running. You can control the walk speed, frame rate (i.e. how many frames per second) and truncation (i.e. how similar or different the images should be).

By clicking on ‘Start Recording’ and then the play button, the images will begin to be collected into an MP4 file.

Infinite Latent Map

This is the main part. It was inspired by the ML Labs Vector Browser.

You will see the selected, ‘pinned’ image in the centre of the map, and around it you will see the neighbouring latent images. The degree to which they are similar or distant from each other can be controlled with Truncation and Sampling Distance.

The map is endless and you can move it in all directions in 2D. Use the mouse wheel to zoom in or out.

This is the real Latent Space you see here.

This is how you work with AI without prompts: by searching actively for aesthetics and visions that might inspire you.

This is GANism, an art style term invented by François Chollet and now a common part of generative art terminology.

Enjoy exploring machine memories!