Visual parameter space exploration for AI art design

Abstract

Generative AI gained great popularity in recent years and it is important that the users understand how the input parameters of such models relate to the generated output. The term parameter space exploration describes the systematic variation of model input parameters and the generation of the corresponding outputs which can then be used to better understand the relations between the parameters and the outputs. With this work, the DiffusionExplorer is proposed, incorporating a visual and interactive framework for parameter space exploration in the context of latent diffusion models. The Iterative View, the Projection View and the Pipeline View build up the major views of the DiffusionExplorer, while the Iterative View offers the user the iterative refinement of a given image by tuning the model parameters. The Projection View embodies the visualization of image samples by using a 2D projection based on similarity metrics. The last view, namely the Pipeline View provides the user with an image history consisting of nodes and edges representing the generation steps. In order to allow a seamless integration of drawing and other editing techniques, the tool is integrated into the painting program Krita via a plugin. After the implementation phase, the DiffusionExplorer has been put to the test, being evaluated with a user study, proving the effective application of the DiffusionExplorer in the context of AI art design. Therefore several participants successfully completed a given task using the tool in combination with Krita and accomplished the synthesis of images that could not be generated in a single synthesis step with a classical approach.