Lately I have been spending a lot of time learning deep learning with PyTorch. I am still very much in the learning phase, but I find the process interesting in the same way I first learned graphics programming. It feels similar to learning shaders or OpenGL, where you start with basic building blocks and slowly begin to understand how data flows through the system.
To keep things grounded and creative, I decided to make a small example video using my existing tools as part of a pipeline. I used PyTorch to generate a latent animation from a set of images, then passed that output through Glitch.GUI to experiment with visual distortion and motion. After that, I mapped the result to a model and applied shaders in ACMX2.
What I like about this approach is that it connects machine learning with tools I already understand. Instead of treating deep learning as something abstract, I can see the results directly and shape them with code and shaders. Each stage of the pipeline does something simple on its own, but together they create something that feels more complex and expressive.
This example video is not meant to be perfect or final. It is more of a proof of concept and a way for me to explore how these systems can work together. I am learning by experimenting, breaking things, and adjusting parameters until something interesting emerges.
I plan to keep pushing this further as I learn more about PyTorch and model behavior. For now, this was a fun and motivating step forward, and it feels good to combine deep learning with my existing creative coding workflow.
Today I spent some time experimenting with a new process that blends neural style transfer with the distortion filters inside ACMX2. The results created a series of portraits that I find interesting from both a technical and creative angle. These images are built from a mix of PyTorch’s neural style reconstruction and the shader based transformations that ACMX2 applies. When they overlap, the two systems produce visuals that neither would generate on their own.
In this set, the artwork forms a hybrid of digital interference, neural color mapping, and CRT style line patterns. The scanline structure appears because the style network creates dense micro textures that ACMX2 then stretches, compresses, and bends like an old analog signal. The circular waves around the face come from the radial warp logic in ACMX2, which treats the center of the image like a gravitational lens. The results look like a portrait being pulled through a field of magnetism, or a signal bending through a distorted display.
The color movement is also interesting. The neural network clusters blues in shadow areas and pushes yellows and reds toward higher intensity regions. When the ACMX2 distortion layers apply after that, the color map bends into patterns that look similar to heat signatures or energy ripples. The two systems merge into a style that looks part psychedelic rendering and part malfunctioning retro hardware.
These experiments are not meant to be perfect or polished. I am just exploring how different tools behave when they interact in unexpected ways. It is fun to see how far a single frame can be pushed when you combine AI based texture reconstruction with shader based warping. This is something I plan to experiment with more as I continue to develop ACMX2 and my artwork.