Stochastic Interpolants for Revealing Stylistic Flows Across the History of Art

Generative models have made rapid progress in content creation, particularly in synthesizing artworks and capturing stylistic variation. However, most methods operate at the level of individual images, limiting their ability to reveal broader stylistic trends and temporal transitions. We address this by introducing a framework that models stylistic evolution as an optimal transport problem in a learned style space, using stochastic interpolants and dual diffusion implicit bridges to align artistic distributions across time without requiring paired data. A central contribution is a diverse dataset of over 650,000 artworks spanning 500 years, curated with metadata across multiple genres. Together, our method and dataset enable tracing long-range stylistic transitions and plausible futures of individual artworks, supporting fine-grained temporal analysis. This offers a new tool for modeling historical patterns in visual culture and opens up promising directions in visual understanding. Code and dataset: https://github.com/CompVis/Art-fm.

• Equal Contribution lution of artistic movements and styles over time.

In this paper, we ask: How would a piece of art evolve? What might it look like if created centuries earlier or later? Naturally, no ground-truth pairs exist for such a task, and the artistic style of a creator is inherently dynamic and historically constrained. To address this challenge, we aim to analyze the flow of artistic styles over time. Given a single artwork, we try to determine its closest "projection" onto the manifold of artistic expressions corresponding to a given historical period. While optimal transport provides a natural formulation for such a task, considering only a single isolated artwork is insufficient. Instead, we reformulate the problem as a distribution matching task, where we study the evolution of the underlying distribution of art over time.

Recent advances in diffusion bridges such as DDIB [83], DSB [13,85], and DDBMs [100] have shown that domain translation can be performed implicitly, without the need for paired supervision. We build on this idea in the context of art history by incorporating stochastic interpolants [1], enabling us to model the temporal evolution of artistic styles without requiring explicitly aligned training data. Unlike traditional style transfer methods that transform images based on predefined and discrete styles, our approach projects styles onto their corresponding distributions across time. Operating in a compact latent space allows us to abstract away pixel-level distraction and focus on meaningful stylistic shifts.

To support this temporal modeling, we require a dataset that spans centuries and reflects a wide variety of artistic styles. While datasets like WikiArt [71] offer breadth in genre and metadata, they lack the scale and historical depth needed to model long-term stylistic progression. Others, such as BAM , are either restricted in access or skewed toward modern art. To address this, we curate a new unified dataset of over 650,000 artworks annotated with creation year, location, artist, medium, and other metadata (Fig. 1). This dataset enables the study of artistic evolution at scale and over time. In summary, our key contributions are as follows:

• We propose the first generative framework that explicitly models the temporal progression of artistic styles across This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.