Xiaomi LociArt Opens 100K-Pair Local Video Stylization Dataset

Xiaomi's MiLM Plus team, working with researchers from Xi'an University of Electronic Science and Technology, posted LociArt, a framework for restyling a chosen region of a video while leaving the rest of the frame alone. The paper bundles three artifacts: a 100,000-pair training dataset across 18 styles, a 2B-parameter model called GRNStyle, and a 168-case benchmark built from 80 source videos and 84 object-mask units.

The system, the authors write, specifies "where to stylize and what style to apply" while preserving non-target content. On their own Style Prototype Similarity metric, GRNStyle reports 0.407 SPS-fg against 0.317 for RegionRoute and 0.386 for a VACE-1.3B baseline fine-tuned on the same data. Temporal consistency lands at 0.983, within rounding of the strongest comparison methods. A 20-person user study rated 160 cases at 4.34 out of 5 for style fidelity and 4.47 for background preservation; the paper reports GRNStyle "wins preservation over Seedance 2.5 for every style" and style for seven of ten.

The method claim worth reading closely is the polar style-aware mask: each style is encoded as an equal-norm direction in a 100-slot circular codebook, which the authors argue lets new styles be added without replaying old training data. On a 10-to-15 class expansion, the polar scheme loses 0.022 SPS on old classes versus 0.102 for a binary-mask baseline. GRNStyle was trained on 8 NVIDIA H200 GPUs. The paper arrives in a dense week for region-aware video work; our AI Video tracker has logged 59 stories in the last 90 days.