Course experiment
I made Twisted Thought GAN in Python and PyTorch for the Neural Networks 2023 exam. It explores what a generative adversarial network produces when trained on a synthetic dataset of Lissajous- and harmonograph-inspired figures.
The project uses the visual regularity of these curves as a compact test case for a GAN. What interested me was the training process and the images it produced.
Dataset and generation
The synthetic dataset contains plotted figures inspired by periodic curve families. A GAN is trained to generate related image-like outputs from random inputs. The repository documents the code and notebooks used for this work, making the project suitable as a record of a course exercise in generative modelling.
Unexpected results
An important part of the project is that the generated images are not exact Lissajous figures. The repository explicitly describes some outputs as resembling “mind flayer” or upside-down forms. That mismatch is not hidden: it is the creative value of the experiment.
The results show how a model trained on structured synthetic images can produce shapes that retain some visual cues while departing from the intended geometry. For this course project, those surprising failures became the most interesting part of the result.
The repository is kept as a transparent record of that experiment, including the gap between the intended curve family and the forms the model actually generated.