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

Yuxuan Chen; Feng Xie; Jiagao Hu; Fuhao Li; Zepeng Wang; Fei Wang; Daiguo Zhou

Abstract

Video stylization typically transforms entire scenes, while local artistic editing requires precise region-style association, spatial control, and temporal consistency. We introduce LociArt, a unified paradigm for free-style local video stylization that explicitly specifies where to stylize and what style to apply while preserving non-target content.

At its core, GRNStyle builds on Generative Refinement Networks and introduces a polar style-aware mask that embeds style identities as equal-norm directions within selected regions. Complementary representation-geometry conditioning combines encoded mask features with raw-mask spatial support. To provide paired supervision, we develop an automated pipeline combining global stylization, video object removal, matting, and compositing, and construct the LociArt-100K Dataset. We further establish the LociArt Benchmark, comprising 80 source videos and 84 target masks covering 168 foreground/background cases, with a Style Prototype Similarity (SPS) evaluation scheme.

LociArt-100K Dataset

LociArt-100K contains 100K paired 720P videos for free-style local video stylization. The dataset is built by separating global style generation from spatial composition. Source and object-removed videos are globally stylized, filtered, matted, and recombined into complementary foreground and background targets.

Dataset construction

Data construction pipeline

The pipeline combines global stylization, object removal, alpha matting, and mask-guided compositing to produce paired local-stylization supervision.

100K720P video pairs

Source, mask, instruction, and locally stylized target.

18artistic styles

From Ghibli and claymation to pixel art and watercolor.

52.4%VLM pass rate

Global stylization candidates entering human verification.

65%human pass rate

Local composites accepted after boundary and temporal review.

Dataset composition

Dataset statistics

LociArt-100K covers 18 artistic categories and a broad distribution of foreground mask area ratios.

Style-category distribution in LociArt-100K — **Style distribution.** The final dataset covers 18 style categories.

Foreground object mask area ratios in LociArt-100K — **Mask area ratios.** Foreground object masks are not the complementary background editing masks.

Interactive sample browser

Pipeline stages

Inspect source, mask, removal, stylization, and final foreground/background composites.

Construction analysis

Data-construction ablation

Direct compositing can leave ghosts or clip stylized contours; object removal and alpha matting resolve different failure modes.

GRNStyle

GRNStyle binds style identities to target regions and combines encoded mask features with explicit geometric support. Source evidence guides content preservation throughout refinement.

01

Polar style-aware mask

Each style is assigned a fixed equal-norm direction on a circle. Its coordinates are embedded inside the selected region, making the style-region interaction explicit before video encoding.

M_s(q) = M(q) · ( (1+a_s,1)/2, (1+a_s,2)/2, 1 )

02

Representation Geometry

VAE-encoded mask features supply spatiotemporal context, while geometric support from binary occupancy localizes mask evidence and suppresses spatial spillover.

ΔZ^M = Φ(M_s) - Φ(0)

03

Compositional inference

Disjoint regions can carry their own learned style codes through the same conditioning interface, without adding channels or conditioning tokens.

M_s^multi = Σ_j M_s^j

Configuration	SPS ↑	CLIPT ↑	PSNR ↑
Base	0.541	0.180	20.09
Base + Polar	0.553	0.181	21.34
Base + RG	0.544	0.179	25.07
GRNStyle	0.554	0.181	25.78

Component ablation. Representation-geometry conditioning (RG) and polar style-aware masks.

**Mask-conditioning ablation.** Representation-geometry conditioning improves spatial localization over raw encoding, while the polar code improves style identity.

LociArt Benchmark

The benchmark separates regional style quality, non-target preservation, and temporal consistency, rather than collapsing everything into a single global score.

80source videos

Each with 81 aligned frames.

84object-mask units

Some videos contain two selected subjects.

168stylization cases

Foreground and background editing for every unit.

93.61%SPS validation

Macro-averaged video-level accuracy under leave-one-source-out validation.

t-SNE visualization of SPS prototypes — **Style prototypes.** 80 CSD prototypes across 18 styles.

Confusion matrix for nearest-prototype validation — **Prototype validation.** Row-normalized confusion under leave-one-source-out validation.

SPSstyle quality

Similarity to a class-specific prototype bank in CSD feature space.

CLIPTtext alignment

Regional alignment between the edited region and the canonical style phrase.

PSNRpreservation

Fidelity to the source outside the intended edit region.

TCtemporal consistency

Adjacent-frame CLIP feature similarity.

StyleVLM score

Style correctness, coverage, and cross-frame stability.

PreserveVLM score

Source fidelity and temporal stability outside the edit.

Local Stylization Results

GRNStyle achieves the best SPS for foreground and background stylization, the best background CLIPT, and a strong trade-off between regional style and non-target preservation.

LociArt Benchmark

Quantitative comparison

Performance is measured with regional style quality, non-target preservation, temporal consistency, and VLM-based perceptual scores.

Method	Venue	SPS ↑		CLIPT ↑		PSNR ↑		TC ↑	Style ↑	Preserve ↑
Method	Venue	fg	bg	fg	bg	fg	bg	TC ↑	Style ↑	Preserve ↑
LEAST^*†‡	arXiv '24	0.283	0.302	0.158	0.173	-	-	0.953	2.20	4.50
VACE-1.3B	ICCV '25	0.156	0.191	0.114	0.132	32.35	29.17	0.981	0.05	5.00
UniVST^*‡	TPAMI '25	0.191	0.286	0.127	0.165	27.69	23.92	0.969	1.83	4.40
RegionRoute^‡	CVPR '26	0.317	0.491	0.154	0.183	29.38	13.17	0.955	2.41	2.75
StyleGallery^*†‡	CVPR '26	0.230	0.325	0.146	0.183	-	-	0.957	3.39	4.77
CoinVE-Edit	arXiv '26	0.248	0.251	0.153	0.159	20.71	16.76	0.983	3.11	3.93
Seedance 2.5	Commercial	0.347	0.358	0.173	0.175	18.50	18.76	0.984	3.98	2.40
VACE-1.3B (our FT)	-	0.386	0.676	0.155	0.185	25.43	25.49	0.980	4.52	4.56
GRNStyle-2B (ours)	-	0.407	0.705	0.160	0.199	25.92	26.15	0.983	4.62	4.64

Benchmark comparison. fg/bg denote editing the target mask and its complement. ^* Training-free methods. ^† Direct mask-compositing methods omit PSNR because non-edited pixels are copied. ^‡ Image-based methods. FT denotes local-stylization fine-tuning.

0.705SPS backgroundBest overall

0.199CLIPT backgroundBest overall

4.62VLM StyleHighest overall

0.983Temporal consistencyNear-best

State-of-the-art samples

Qualitative comparison

Compare GRNStyle with CoinVE-Edit, Seedance 2.5, and RegionRoute where available. All 60 comparison videos in the project sample set are included.

Human evaluation

User study

Twenty participants rated a 160-case subset (16 cases per style) for style fidelity and background preservation.

Human evaluation of style fidelity and background preservation — **Human evaluation.** Twenty participants on 160 cases across ten styles. GRNStyle achieves the highest mean ratings for style fidelity (4.34) and background preservation (4.47).

Key findings

Consistent gains across styles

GRNStyle outperforms CoinVE-Edit on both criteria for all ten styles, improves preservation over Seedance 2.5 for every style, and improves style fidelity for seven styles.

4.34 mean style fidelity
4.47 mean background preservation
20 participants across 160 cases

Replay-free style expansion

New style without replay

Expanding from 10 to 15 classes without replaying old classes retains the original styles with substantially less forgetting and adds the five new styles through polar conditioning.

Configuration	10 classes (old)			5 classes (new)
Configuration	SPS ↑	CLIPT ↑	PSNR ↑	SPS ↑	CLIPT ↑	PSNR ↑
Binary, 10 classes	0.547	0.177	25.42	-	-	-
Binary, 15 classes	0.445 (-0.102)	0.158 (-0.019)	25.43	0.588	0.197	25.39
Polar, 10 classes	0.553	0.180	25.75	-	-	-
Polar, 15 classes	0.531 (-0.022)	0.172 (-0.008)	25.77	0.590	0.199	25.74

Replay-free style expansion. Expanded models are trained only on the five new classes. Retention and acquisition are evaluated on 1,680 original-style and 840 new-style cases, respectively.

Video Gallery

Explore local and whole-frame stylization with source videos and generated results. Local examples also include mask context; the final gallery compares GRNStyle with direct pipeline outputs.

GRNStyle outputs

Foreground, background, and multi-region

Four curated cases in each category, with source and mask context for direct inspection.

GRNStyle outputs / Whole-frame editing

Global stylization

Selecting every pixel applies one style to the complete scene.

End-to-end diagnostics

GRNStyle vs. Pipeline

These examples compare direct pipeline outputs with GRNStyle when a stylized subject no longer matches the source mask or when an object-removal stage damages scene elements.

Citation

@article{chen2026lociart,
  title   = {LociArt: A Unified Paradigm for Free-Style Local Video Stylization},
  author  = {Chen, Yuxuan and Xie, Feng and Hu, Jiagao and Li, Fuhao and Wang, Zepeng and Wang, Fei and Zhou, Daiguo},
  journal = {arXiv preprint arXiv:2610.36456},
  year    = {2026},
  url     = {https://arxiv.org/abs/2610.36456}
}