Hacking Generative Models with Differentiable Network Bending
This is an incremental method for artists and researchers to creatively alter generative model outputs.
The paper tackles the problem of manipulating generative models to produce outputs deviating from the original training distribution by injecting a trainable module between intermediate layers, resulting in uncanny images useful for artistic purposes.
In this work, we propose a method to 'hack' generative models, pushing their outputs away from the original training distribution towards a new objective. We inject a small-scale trainable module between the intermediate layers of the model and train it for a low number of iterations, keeping the rest of the network frozen. The resulting output images display an uncanny quality, given by the tension between the original and new objectives that can be exploited for artistic purposes.