Mask says where, not what style: GRNStyle writes styles into the mask as polar directions and reaches 0.705 SPS on background edits
Synopsis
Local video stylization used to work in two steps: a mask marked where to edit, and a separate text prompt named the style, leaving the model to link the two. GRNStyle, built on Generative Refinement Networks, instead writes each style into the mask as a fixed polar direction of equal length, so region and style arrive as one conditioning field. Trained on 100K pairs built from global stylization, object removal, matting and compositing, it reaches a Style Prototype Similarity of 0.705 for background editing and 0.407 for foreground editing on the 168-case LociArt Benchmark, the highest in the comparison.
Interpretation
A clip can be restyled only inside the mask while the rest of the frame keeps its original pixels, with one model delivering both. Earlier local pipelines encoded the edit extent as a binary mask and the appearance as separate text, so the mask said where to edit but nothing in it said which style belonged to the selected region. On the 168 cases of the LociArt Benchmark, GRNStyle records the best regional Style Prototype Similarity in the table for foreground (0.407) and background (0.705) editing, while its background PSNR of 26.15 dB ranks second.
Style stops arriving as separate text and becomes geometry carried inside the mask. The polar style-aware mask places each style at a fixed angle on the unit circle at equal radius, so styles differ by direction and never by an arbitrary scalar order, and representation–geometry conditioning anchors VAE-encoded mask features to the encoding of an empty mask while adding pooled raw-mask occupancy as explicit spatial support. Training uses 100K pairs over 18 styles, produced by global stylization, object removal, matting and compositing in sequence and then human-verified.
Perspective
The demonstrated scope is local stylization evaluation on the LociArt Benchmark: 80 source videos and 84 object–mask units run under foreground and background editing protocols, 168 cases in total, with videos standardized to 81 frames at 720p, alongside a user study of 20 participants on a 160-case subset.
Preservation figures need to be read alongside style strength: the highest PSNR in the table comes from VACE-1.3B, whose stylization is minimal, so that number does not indicate a better style–preservation trade-off. Multi-region, multi-style compositions appear only as qualitative examples after single-region, single-style training, and codebook extensibility was tested only by extending a 10-class model to 15 classes with no replay of the original data.