Title: Hist2Style: Bilateral Grid-Based Stylization Guided by Histograms

Abstract: The objective of photorealistic style transfer is to align the color and tonal characteristics of an input image with those of a designated style reference, all while retaining the original scene’s content and fine details. While contemporary large-scale image models can enable such appearance modifications, their prohibitive computational requirements, susceptibility to generating hallucinations, and restricted user control render them impractical for high-resolution, real-time applications. To address these limitations, we present Hist2Style, a bilateral-grid approach designed for rapid, edge-aware stylization. This method ensures visual fidelity by restricting operations to locally affine transformations within bilateral space. By leveraging a substantial supervised dataset generated via language and vision-language models, our approach distills a large image editing model into a lightweight network tailored for spatially varying color adjustments. The system utilizes a histogram-based embedding of the style target, offering an intuitive interface that allows users to fine-tune the output style by altering the target color distribution. Consequently, Hist2Style inherently preserves content structure, eliminates hallucinations, and facilitates real-time, high-resolution photorealistic stylization with interactive, user-driven control over color and tone.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC