Reproduce “Color Transfer with Modulated Flows” on OpenResearch

About this paper

Maria Larchenko, Alexander Lobashev, Dmitry Guskov, Vladimir Vladimirovich Palyulin

The ModFlows method introduces a novel approach to color transfer by integrating optimal transport theory with neural networks, utilizing rectified flows and a common intermediate latent space. This method achieves competitive performance, demonstrated by the lowest aggregated DISTS scores for content preservation (e.g., 0.129 grayscale) and a mean Wasserstein distance of 0.123 for style matching, while generalizing to new image pairs without requiring fine-tuning.

Key insights

• Decomposing the color transfer into two rectified flows through a common intermediate uniform latent space simplifies the learning task and ensures bijective, invertible transformations.
• Training an encoder to predict rectified flow parameters acts as an implicit regularizer, leading to smoother color transformations and a reduced likelihood of visual artifacts.
• The encoder's output vector effectively functions as an embedding of an image's color palette, enabling applications like style similarity search.

Summary by alphaXiv.