LociArt: A Unified Paradigm for Free-Style Local Video Stylization

LociArt. A Unified Paradigm for Free-Style Local Video Stylization. By Yuxuan Chen and six others, from Xiaomi and Xi'an University of Electronic Science and Technology. Published on the twenty-seventh of October.

Video stylization typically transforms the entire scene. Local artistic editing asks for something harder. It needs a precise association between each region and the style meant for it, spatial control over where that style goes, and temporal consistency so the result holds steady from frame to frame. Restyling everything at once avoids the first two questions, and local editing cannot avoid them.

LociArt is a unified paradigm for free-style local video stylization. It specifies explicitly where to stylize and what style to apply, while preserving the content outside the target. At its core sits GRNStyle, built on Generative Refinement Networks, or GRN. Its key addition is a polar style-aware mask, which embeds style identities as equal-norm directions within the selected regions. A region, in other words, is marked not only as a place but with a direction that stands for its style.

Alongside the mask comes a complementary representation and geometry conditioning. It combines encoded mask features with raw-mask spatial support, so that spatiotemporal context and geometric control are brought together in one signal. That answers the where and the what in the same breath.

Training such a model needs paired examples, and the authors build them automatically. Their pipeline combines global stylization, video object removal, matting, and compositing. The result is the LociArt-100K Dataset, holding one hundred thousand local stylization pairs.

For evaluation, the authors establish the LociArt Benchmark. It holds eighty source videos and eighty-four target masks, which together cover one hundred and sixty-eight foreground and background stylization cases. Regional style quality is scored with a scheme called Style Prototype Similarity, or SPS. It is supported by complementary measures of how well non-target content is preserved, and of temporal consistency.

In experiments, GRNStyle outperforms existing methods in regional style quality and in spatial control, while maintaining strong temporal consistency. Qualitative examples go further. They show several regions stylized jointly, with the same learned style or with different ones. That happens even though training used only pairs with a single region and a single style, so the multi-region behaviour was never directly taught.

What arrives here is a package with three parts: a model, the data to supervise it, and a benchmark to judge it by. The authors present them together as a unified foundation for learning and evaluating controllable local video stylization, the kind that says exactly which part of the frame changes, and how.