Applying a Color Palette with Local Control using Diffusion Models

We demonstrate two novel editing procedures in the context of fantasy art. Palette transfer applies a specified reference palette to a given image. For fantasy art, the desired change in palette can be very large, leading to huge changes in the ``look'' of the art. We show that a pipeline of vector quantization; matching; and ``dequantization'' (using a diffusion model) produces successful extreme palette transfers. A novel training loss measures the match between color distribution in control and generated images even when a ground truth target is not available. This measurably improves performance. Segment control allows an artist to move one or more image segments, and to optionally specify the desired color of the result. The combination of these two types of edit yields valuable workflows. We demonstrate our methods on the challenging Yu-Gi-Oh card art dataset.
Comment: 14 pages, 14 figures