Hacking Generative Models with Differentiable Network Bending

TL;DR

This paper introduces a method to manipulate generative models by inserting a trainable module that shifts outputs away from the original distribution, enabling artistic and novel image generation.

Contribution

It presents a novel approach to 'hack' generative models through differentiable network bending by inserting and training a small module while keeping the rest of the network fixed.

Findings

Generated images exhibit high quality and uncanny effects.

The method effectively shifts outputs towards new objectives.

Few training iterations are sufficient for significant output changes.

Abstract

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.