acceptodds
Withdrawn from ICLR 2027

LociArt: A Unified Paradigm for Local Video Stylization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

12 views

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 (GRN) 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, integrating spatiotemporal context with geometric control. To provide paired supervision, we develop an automated pipeline combining global stylization, video object removal, matting, and compositing, and construct the LociArt-100K Dataset with 100K local stylization pairs. We further establish the LociArt Benchmark, comprising 80 source videos and 84 target masks covering 168 foreground/background stylization cases, with a Style Prototype Similarity (SPS) evaluation scheme for regional style quality and complementary measures of non-target preservation and temporal consistency. Experiments show that GRNStyle outperforms existing methods in regional style quality and spatial control while maintaining strong temporal consistency. Qualitative examples further illustrate joint stylization of multiple regions with the same or different learned styles, despite training only on single-region, single-style pairs. Together, these components provide a unified foundation for learning and evaluating controllable local video stylization. Anonymous demo page: https://lociart-page.pages.dev/.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
local video stylization, video style transfer, region-conditioned generation, video diffusion models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_ZJfdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This work focuses on localized video stylization and introduces a new dataset, LociArt-100K, which contains 100,000 pairs of localized stylization data. The authors also present the LociArt Benchmark—comprising 80 source videos and 84 target masks—and propose a new evaluation metric called "Style Prototype Similarity" (SPS). Based on these, the authors introduce a unified framework, LociArt. Building upon the existing Generative Refinement Network, the framework incorporates a "polar style-aware mask" and a "representation-geometry condition," enabling the integration of spatiotemporal context with geometric control.

Critical strengths
  1. The scope of the work is substantial; it introduces a dataset, a benchmark, and an evaluation metric, while also improving upon existing methods.
  2. The dataset, as showcased on the website, appears to be of high quality.
Critical weaknesses
  1. Although the method significantly outperforms existing approaches in terms of quantitative metrics, the visual quality of the results remains somewhat coarse.
  2. The video examples provided on the anonymous website—both in the dataset and those generated by the method—exhibit very limited motion; in particular, the overall movement within the frames is minimal. It is unclear whether the method faces limitations when applied to scenarios with more significant motion. The authors are encouraged to provide additional examples that demonstrate the boundaries of the method's capabilities.
  3. The paper does not showcase longer sequences; it appears the method is currently limited to generating clips of approximately 3 seconds in duration.
  4. In the ablation study (Table 2), the "RG" (representation-geometry condition) and "polar encoding" components yield only marginal improvements in the SPS and CLIPT metrics, with only the RG component showing a notable impact on PSNR. It is worth questioning whether the polar encoding is truly necessary.
Questions for authors

See weaknesses

1 comment
AuthorsOpenReview ↗

Response to Reviewer ZJfd

We thank the reviewer for recognizing the scope of our contribution and the quality of the dataset. We address the four concerns below, with additional motion examples and mask-robustness results.

1. Visual quality and comparison with existing methods

Our visual evaluation considers both regional style expression and preservation of non-target content. The full SOTA video comparisons show stronger local stylization while preserving non-target content, rather than merely achieving higher numerical scores. For example, GRNStyle applies pixel-art stylization to the dog while preserving the armchair and room. These videos make the qualitative difference directly inspectable during playback. Some results remain imperfect, particularly around fast-moving boundaries, as already discussed in the limitations section.

2. Motion range and additional continuous videos

The original version of our anonymous page already includes appreciable motion, such as the first swinging example under Claymation and the first horse-riding example under Oil Painting. Importantly, the SOTA comparison videos retain all frames but are played at 16 fps instead of 24 fps, making playback two-thirds of the original speed. This presentation choice makes motion appear slower; it does not indicate that the underlying sequences contain only small movements. We have clarified this playback setting on the page.

We have also added 10 additional examples with substantial motion, covering five foreground and five background edits, with synchronized source, mask, and output videos. Most of the displayed examples retain coherent local stylization and preserve non-target content, although rapid motion can still affect quality and boundary stability. They provide direct video evidence beyond the limited-motion examples, while also making the remaining limitations visible.

3. Video duration and scope

Our focus is stylization quality and local region–style control, not extending the temporal horizon of the underlying generator. Our inference frame count and duration follow the backbone, whose temporal capacity determines the supported video length. The proposed region–style conditioning does not impose a separate temporal window and is, in principle, compatible with the backbone’s longer-video generation strategies. Extending the backbone’s temporal horizon is therefore orthogonal to our contribution and outside the scope of this work.

4. Complementary roles of RG and Polar

RG primarily improves spatial preservation, which is a central objective of local editing: adding RG increases PSNR from 20.09 to 25.07 dB without materially changing stylization scores.

Polar serves two complementary purposes: improving stylization on top of RG and supporting replay-free style expansion. First, adding Polar to RG increases SPS from 0.544 to 0.554, while also improving PSNR from 25.07 to 25.78 dB. Second, it supports learning additional styles without replaying the original training data, while better retaining previously learned styles, as discussed in Sec. 5.5. In expansion from 10 to 15 styles, training uses only the five new classes: the old-class SPS drop is 0.022 with Polar versus 0.102 with binary masks, and the CLIP-T drops are 0.008 versus 0.019. These results demonstrate benefits in both stylization quality and replay-free extensibility.

Additional robustness evidence. As complementary evidence of practical local control, the added appendix Section G “Robustness to Imperfect Masks” evaluates five perturbation types at two severity levels, with 400 video–style evaluations per variant. Under light boundary noise, dilation, erosion, and temporal jitter, SPS and CLIP-T changes remain within 0.006 and PSNR changes within 0.49 dB. Light frame dropout causes a larger background SPS decrease of 0.022, while stronger perturbations expose further degradation. These results demonstrate that GRNStyle, with its complementary RG and Polar conditioning, maintains robust local stylization under moderate mask inaccuracies.

Reviewer_Abd6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work introduce a dataset for local video stylization, built by object removal, global stylization, matting and compositing with VLM + human filtering, a model and a region-aware benchmark.

Critical strengths
  1. This work introduce a benchmark and a training dataset.

  2. This paper is well written and easy to understand.

3.This work is clearly organized.

Critical weaknesses
  1. Video stylization is not a very difficult task. Existing solutions including a lot trained/ train-free works for this topic. This method's novelty is limited, it collects a dataset and use the pairs to train a model used for local video stylization. The results seems not natural and it is not clear the novelty and usefullness of local video stylization.The novelty and effectiness are not clear to me.

  2. Most of the gain looks like data, not architecture.

  3. "Free-style" is a stretch: styles are a closed 18-class codebook and a new style needs paired data + fine-tuning, while Seedance 2.5 and CoinVE take open text. Frame it as a trade-off.

4.Multi-region composition is in the abstract and contributions but only shown qualitatively.

  1. Image-based and training-free baselines aren't strong competitors; the real video comparison is CoinVE-Edit and a closed commercial system. Dataset release is "planned, subject to licensing", which matters for a dataset paper.
Questions for authors

Please address the concerns raised above

3 comments
AuthorsOpenReview ↗

Response to Reviewer Abd6

We thank the reviewer for recognizing the quality of the dataset and benchmark. We clarify below the motivation for local video stylization, the respective contributions of the data and model, the meaning of “free-style,” and the scope of our multi-region and baseline claims.

1. Why local video stylization matters

Local video stylization lets users stylize a subject while preserving its surroundings. This selective control is useful beyond global stylization. Recent methods such as UniVST (TPAMI '25), RegionRoute (CVPR '26), and StyleGallery (CVPR '26) also study local stylization (Sec. 2). The task requires both region control and style control, with coherent boundaries over time. Simple mask compositing can clip changed contours or leave residual content, as shown in Sec. 5.6.

Our contributions cover both data and model design:

  • Data (Sec. 3): We introduce a paired-data construction pipeline and LociArt-100K.
  • Polar Style-aware masks (Sec. 4.2): Equal-norm polar codes explicitly bind styles to regions. They distinguish style identities without imposing a scalar ordering.
  • Representation–geometry conditioning (Sec. 4.3): We combine VAE-encoded mask features with raw-mask spatial support. This brings together spatiotemporal context and spatial control.

These designs go beyond fine-tuning on paired data. The component ablation (Sec. 5.4) shows improved non-target preservation. Replay-free expansion (Sec. 5.5) shows less forgetting with polar masks than with binary masks. These principles may also inform other mask-guided local-editing methods. We will release the dataset and model weights upon acceptance, subject to applicable licenses.

Some cases remain challenging. The side-by-side videos on our anonymous project page (https://lociart-page.pages.dev/#sota) show visible limitations across the compared state-of-the-art methods. Our method provides the strongest overall local-stylization results in these comparisons, consistent with Sec. 5.2. These remaining challenges highlight the difficulty of the task.

2. Data contribution versus model contribution

Both the paired data and our model designs contribute to the improvements. The model contribution is supported by two experiments:

  • Component ablation (Sec. 5.4). Adding representation–geometry conditioning raises non-target PSNR from 20.09 to 25.07 dB. Adding polar conditioning raises SPS from 0.541 to 0.553. Together, they achieve 25.78 dB PSNR and 0.554 SPS. Thus, the model gains extend beyond style scores to non-target preservation.
  • Replay-free style expansion (Sec. 5.5). We expand from 10 to 15 styles using only the five new styles. Old-style SPS drops by only 0.022 with polar codes, versus 0.102 with binary masks. New-style SPS remains comparable (0.590 vs. 0.588). Polar conditioning therefore reduces forgetting without compromising new-style learning.

The VACE comparison provides complementary evidence (Sec. 5.2, Table 1). Fine-tuning VACE on our data raises foreground/background SPS from 0.156/0.191 to 0.386/0.676. This shows the value of the data. GRNStyle reaches 0.407/0.705 on the same training dataset. Since the backbones differ, the component ablation above is the more direct evidence for our model designs.

These results show why the contribution is not limited to data construction. The proposed conditioning improves spatial preservation and supports style expansion with less forgetting.

3. Meaning of “free-style”

In our work, “free-style” refers to flexible region selection, composition of learned styles, and an expandable style set. To avoid ambiguity, we have clarified this meaning in the abstract and Sec. 1.

Compared with the open-text prompting supported by Seedance 2.5 and CoinVE, our approach makes a different trade-off. Our current model operates within an 18-class style codebook (Sec. 5.1). Adding a new category requires paired data and fine-tuning. We therefore prioritize explicit region–style control over open-ended style prompting. Our model binds styles directly to selected regions (Sec. 4.2). On the evaluated styles, it achieves stronger regional style scores than these methods (Table 1). The component ablation also shows improved non-target preservation (Sec. 5.4). The trade-off is stronger control and demonstrated quality on the evaluated styles, at the cost of a restricted inference-time style set and additional training for new categories.

Beyond the currently learned styles, our framework also supports expanding the style set through training. Sec. 5.5 demonstrates expansion from 10 to 15 styles without replaying old data. Old-style SPS drops by only 0.022 with polar masks, versus 0.102 with binary masks. New-style performance remains comparable. This provides a practical path to adding styles while better retaining existing capabilities. It is this trainable extensibility, together with flexible regional control, that we intend by “free-style.”

AuthorsOpenReview ↗

Response to Reviewer Abd6 - continue

4. Multi-region composition

Multi-region composition is an additional generalization capability of our model. Training uses single-region, single-style pairs. At inference, the same interface assigns learned styles to multiple disjoint regions (Sec. 4.4, “Compositional inference”). All regions are processed in just one generation run, rather than through repeated single-region editing.

We add a quantitative comparison with sequential editing in Appendix H. LociArt-Bench contains four cases with multiple mask regions, each with two masks. We evaluate all four using our main GRNStyle model without additional training. All ordered pairs of distinct styles from ten styles give 4 × 10 × 9 = 360 cases for each protocol. We additionally evaluate same-style assignments to both regions:

ProtocolSPSCLIP-TPSNR
Same style, sequential0.4640.17022.45
Same style, compositional0.4670.16824.01
Different styles, sequential0.4520.16722.40
Different styles, compositional0.4290.16027.01

Compositional inference preserves non-target content better while maintaining comparable same-style quality; different-style composition involves a modest stylization-preservation trade-off. Compositional inference also processes both regions in one generation run rather than two, avoiding a second full-video generation and reducing total inference computation under the same per-run settings. The anonymous project page also provides qualitative examples.

5. Baselines and release plan

To our knowledge, ours is the first method trained specifically for local video stylization with paired supervision. This leaves no directly comparable prior method trained for the same task. We therefore select baselines for their local-stylization capability, covering three groups (Sec. 5.2, Table 1):

  • Local image stylization: LEAST, RegionRoute, and StyleGallery. These methods test regional stylization quality and spatial control.
  • Local video stylization: UniVST, the only prior dedicated video stylization method with local control that we are aware of. It is training-free.
  • General-purpose video editing with local-stylization capability: CoinVE-Edit, VACE, and Seedance 2.5. We also include VACE-FT, fine-tuned on our paired data, as a task-adapted baseline.

This selection covers both specialized local-stylization methods and broader video editors. It also includes a baseline trained on the same dataset, rather than relying only on image or training-free methods.

Image methods run independently on each frame. Their comparison tests regional style quality, not just temporal consistency. Even excluding TC, GRNStyle leads these image baselines on both regional style metrics (Table 1):

  • Foreground/background SPS: 0.407/0.705, versus the best image-baseline values of 0.317/0.491.
  • Foreground/background CLIP-T: 0.160/0.199, versus 0.158/0.183.

To complement the local comparisons in Table 1, we have added new global-stylization results in Appendix F of the revised manuscript. We compare with StyleMaster, TeleStyle, Ditto, EchoStyle, and PickStyle. We reuse the locally trained checkpoint and simply set the mask to all ones, without any additional training or fine-tuning. Among methods evaluated on all ten styles, GRNStyle achieves the highest CSD. These results show that our locally trained model also performs competitively in global stylization.

Regarding data availability, we will release LociArt-100K and the model weights upon acceptance. In the meantime, reviewers can inspect dataset samples on our anonymous project page, alongside the construction and verification procedure described in Sec. 3. These examples provide a preview of the dataset before its full release.

Thank you for the experiments and clarifications. While several concerns have been addressed, however, I remain unconvinced about the methodological novelty and practical effectiveness,especially regarding the naturalness of the generated results.

Reviewer_jzdKOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces LociArt, a unified framework for free-style local video stylization that enables region-specific artistic transformations while preserving non-target content. The proposed GRNStyle incorporates polar style-aware masks and representation–geometry conditioning to improve style–region association and spatial control. The authors also construct the LociArt-100K dataset and establish a benchmark for evaluating regional stylization, content preservation, and temporal consistency. Experimental results demonstrate promising performance in localized stylization and multi-region style composition.

Critical strengths
  1. The paper addresses the challenging problem of local video stylization, where precise region control, style assignment, and content preservation must be achieved simultaneously. The polar style-aware mask provides an explicit representation of style–region associations, while the complementary representation–geometry conditioning balances spatiotemporal context with spatial localization. The framework also supports multi-region, multi-style composition without requiring explicit compositional training pairs.

  2. The authors construct LociArt-100K through an automated data-generation pipeline with quality filtering and human verification. The accompanying benchmark evaluates foreground and background stylization across diverse styles and scenarios, using complementary metrics for regional style fidelity, non-target preservation, and temporal consistency. The extensive experiments, component ablations, and robustness analysis provide useful evidence of the effectiveness and practical potential of the proposed framework.

Critical weaknesses

1、Figure 4 mixes image-based stylization methods with general-purpose video generation and editing systems, rather than consistently comparing against methods specifically designed for video stylization. The authors should clarify how image-based baselines are adapted to video inputs and ensure that the comparison accounts for temporal consistency. Additional comparisons with dedicated video stylization methods, together with multi-frame visualizations, would provide a more convincing evaluation of the proposed approach.

2、Although the paper focuses on video stylization and reports a temporal consistency metric, the qualitative comparisons mainly present static frames, making it difficult to assess temporal stability. The sparse frame visualizations in the appendix are also insufficient to reveal potential flickering, style drift, or boundary inconsistencies during continuous playback. The authors should provide more comprehensive temporal visualizations, such as consecutive-frame comparisons or video sequences, to better demonstrate the temporal consistency of the proposed method.

3、StyleMaster and TeleStyle are representative methods for video stylization and content-preserving video style transfer, respectively. The authors are encouraged to include comparisons with these methods to provide a more comprehensive evaluation of the proposed approach, particularly in terms of stylization quality and temporal consistency.

4、The authors could consider including TeleStyle, particularly its Qwen-Image-Edit-based image stylization branch, as an additional baseline for evaluating single-frame stylization quality.

[1] StyleMaster: Stylize Your Video with Artistic Generation and Translation. CVPR 2025. [2] TeleStyle: Content-Preserving Style Transfer in Images and Videos. 2026. arXiv:2601.20175.

Questions for authors

see Weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer jzdK

We thank the reviewer for the constructive suggestions on video-aware evaluation and temporal visualization. We have revised the manuscript to clarify the local stylization scope and protocol of our comparisons, and added an additional global video-stylization comparison in the appendix.

1. Baseline scope and temporal evaluation

We have clarified Table 1 by adding an Image/Video type column. Our comparison is organized around the target task—local video stylization—rather than restricted to methods designed only for global video stylization. Accordingly, we include local image stylization methods to assess per-frame regional stylization and spatial control, UniVST as a video-based local stylization method, and video generation/editing systems that support localized editing as video-capable baselines. These categories serve different comparison purposes and are now explicitly identified in Table 1.

The image baselines are applied independently to each frame, without temporal post-processing. Their SPS, regional CLIP-T, and PSNR are computed framewise and then aggregated over the video. Thus, image methods are not treated as temporally modeled video methods; they are included to isolate single-frame local stylization and region-control quality. Importantly, even without relying on TC, GRNStyle outperforms all three image baselines on both foreground/background SPS and regional CLIP-T.

For dedicated video stylization, we include UniVST and CoinVE-Edit in the local comparison and add a separate global video-stylization comparison in Appendix F. There, using an all-one mask, we compare GRNStyle with StyleMaster, Ditto, EchoStyle, and PickStyle under a common protocol. This global evaluation complements rather than replaces the local comparison, since global stylization does not directly test mask-guided regional editing.

2. Temporal visualization

We agree that still frames cannot fully reveal flicker, style drift, or boundary instability. Our anonymous project page provides complete multi-frame video comparisons with the compared methods, allowing direct inspection of temporal stability, style drift, and boundary consistency during playback: https://lociart-page.pages.dev/#sota. These dynamic comparisons complement the aggregate TC scores; we therefore do not interpret TC alone as evidence of artifact-free temporal behavior.

3. StyleMaster and global video stylization

StyleMaster is a global video-stylization method, whereas Table 1 evaluates local stylization/editing. We therefore added a separate appendix Section F “Global Stylization Comparison,” rather than treating global stylization as a direct local-editing baseline. With an all-one mask and the same GRNStyle checkpoint (without retraining), we compare against StyleMaster, Ditto, EchoStyle, and PickStyle on the shared 80-video, ten-style setting (EchoStyle supports five styles and is reported separately for reference).

In this global setting, GRNStyle obtains CSD 0.683, CLIP-T 0.190, TC 0.977, and VLM Style 4.67. Compared with StyleMaster, GRNStyle achieves higher CSD (0.683 vs. 0.428), CLIP-T (0.190 vs. 0.189), TC (0.977 vs. 0.962), and VLM Style (4.67 vs. 2.12). We report these results as complementary global-stylization evidence, not as a replacement for the local comparison.

4. TeleStyle

We have clarified TeleStyle's scope in Related Work: TeleStyle includes both image and video stylization, but it is not designed for mask-guided local editing. Its Qwen-Image-Edit branch is a global image stylization method and therefore is not a directly comparable local baseline for Table 1.

AuthorsOpenReview ↗

Added TeleStyle Results

We have added TeleStyle scores for global video stylization in Appendix Sec. F and Table 6. Please check it.

Reviewer_DYZjOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper is about local video stylization: you give it a video, a mask, and a style name, and it repaints only the masked region while leaving everything else alone and keeping things temporally stable. It comes with three pieces. (1) LociArt-100K — a data pipeline that globally stylizes both the original video and an object-removed version (using VACE / Ditto), filters the results with GPT-5.5, mattes the original object and the stylized object separately (VideoMaMa), and then composites them into fg/bg-stylized training targets; 100K pairs, 18 styles, human-checked. (2) GRNStyle — built on top of GRN/GRNEdit (a non-diffusion binary-token generator, initialized from an already-trained GRNEdit), adding a polar style-aware mask (basically a per-style unit vector painted into the mask channels) and RG conditioning (subtract the empty-mask encoding, then gate with a pooled occupancy map). (3) LociArt Benchmark — 168 fg+bg cases scored with a CSD-based Style Prototype Similarity (SPS) metric, plus CLIP-T, PSNR, temporal consistency and VLM scores. GRNStyle comes out on top on SPS and VLM style against seven baselines, including a VACE-1.3B fine-tuned on the same data; they also run ablations, a 20-person user study, a replay-free style-expansion test, and show some multi-region composition examples.

Critical strengths
  1. The pipeline and dataset are the real contribution here. Paired supervision for local video stylization simply didn't exist before. The remove–stylize–matte–composite design fixes the ghosting/clipping you get from stylization-induced contour changes (Fig. 7), and it's delivered at scale (100K pairs, 18 styles, human-verified). Just fine-tuning VACE-1.3B on this data lifts SPS from 0.156/0.191 to 0.386/0.676 — the data is valuable on its own, and that's the main reason I lean toward accepting.

  2. The evaluation is more thorough than what you usually see. Region-aware metrics, VLM scoring, a user study, component and pipeline ablations, plus a replay-free expansion study — that's a fuller evidence chain than most data+model papers in this space.

Critical weaknesses
  1. SPS is circular with respect to the training data. App. C.2 says the prototypes come from globally stylized videos that passed the pipeline's quality gates — i.e., VACE/Ditto outputs — and the training targets are composited from those same outputs. So SPS really measures "how close are you to VACE/Ditto's rendering of the style", which favors anything trained on LociArt-100K and penalizes baselines (Seedance 2.5, StyleGallery, UniVST, …) that render the style differently. Because of that, the Table 1 lead can't be read as better stylization. On non-circular metrics the picture is mixed (CLIP-T-fg 0.160 vs. Seedance 0.173; Preserve 4.64 vs. VACE 5.00 / StyleGallery 4.77), and the user study leaves out the strongest baseline (VACE-FT).

  2. The modeling contribution is marginal. In Table 2, Polar adds +0.012 SPS and RG basically only moves PSNR; together that's +0.013 SPS. App. E.2 shows validation SPS going from 0.512 to 0.554 over the entire run — and that's starting from a GRNEdit checkpoint the authors themselves describe as having "strong stylization capabilities". Against a VACE-1.3B fine-tuned on the same data, GRNStyle-2B gains roughly 0.02–0.03 SPS, ~0.5 dB PSNR and ~0.1 VLM, with 50% more parameters. And the headline zero-shot multi-region / multi-style composition only shows up in curated figures — no numbers, no failure analysis, no sequential per-region baseline.

Questions for authors
  1. (Clarification, critical) Could you rebuild the SPS prototypes from references that are independent of the training generators (real artworks, or a generator you didn't train on) and show how Tables 1–2 change? It would also help to report the SPS of the raw VACE/Ditto global stylizations as a ceiling.
  2. (Clarification) Are the Table 2 gaps (e.g., SPS 0.541 vs. 0.554) actually significant over 168 cases? Please add std/CIs. Also, was each user-study case rated by just a single participant?
  3. (Clarification) For multi-region composition, could you report per-region SPS/CLIP-T, cross-region style leakage, and PSNR for N regions (same and different styles), against a sequential per-region baseline?
  4. (Clarification) Are Polar and RG backbone-agnostic — could you just add them to VACE-FT? And if Polar mainly helps with forgetting (Table 3), how much does that matter in practice?
  5. (Clarification) At inference, does the user have to pick one of the 18 trained styles, or can the text specify an unseen one? If it's the former, please revise "free-style".
  6. (Suggestion) Please add VACE-FT to the user study and report inter-rater agreement.
2 comments
AuthorsOpenReview ↗

Response to Reviewer DYZj

We thank the reviewer for recognizing the dataset's value and for the specific suggestions. We will release LociArt-100K and the model weights upon acceptance. We address the six questions below.

Q1. SPS circularity and external references

We agree that the original SPS prototypes may favor the training generators. We rebuilt them using style samples from OpenVE-3M[1], RefVideo6M[2], and VISTA-1000[3] datasets: 100 randomly selected videos per style, covering nine styles (Children's illustration style is unavailable). We then rescored all Table 1 methods on these nine styles. Our GRNStyle remains highest on both foreground and background SPS with these external references:

MethodSPS-fg ↑SPS-bg ↑
LEAST0.2210.228
StyleGallery0.2280.337
RegionRoute0.3200.501
VACE-1.3B0.1570.196
UniVST0.1870.327
CoinVE-Edit0.2390.237
SeedDance0.3490.438
VACE-1.3B (our VACE-FT)0.3470.598
GRNStyle-2B (ours)0.3760.657

We also sampled 500 videos per style from the original Ditto-1M dataset. Their mean SPS is 0.773 with the original ten-style prototypes and 0.728 with the external nine-style prototypes. These scores can be viewed as an empirical upper-bound reference for SPS under the respective prototype settings.

Ref:

  1. OpenVE-3M: A Large-Scale High-Quality Dataset for Instruction-Guided Video Editing, arXiv:2512.07826
  2. RefVideo-6M: A Reliable Reference-Based Dataset for Instructional Video Editing, arXiv:2608.26101
  3. VISTA: Triplet-Supervised Video Style Transfer with Diffusion Transformers, arXiv:2605.17312

Q2. Table 2 uncertainty and user-study coverage

We now report means ± nominal 95% CI half-widths over 1,680 cases: 10 styles × 84 units × foreground/background. These statistics use the original ten-style SPS protocol, separate from the external-reference evaluation above.

ConfigurationSPS ↑CLIP-T ↑PSNR ↑
Base0.541 ± 0.0050.180 ± 0.00220.09 ± 0.17
Base + RG0.544 ± 0.0090.179 ± 0.00225.07 ± 0.16
Base + Polar0.553 ± 0.0060.181 ± 0.00221.34 ± 0.15
Full model0.554 ± 0.0090.181 ± 0.00225.78 ± 0.16

The raw user-study wording was unclear, and we have corrected it. Each output in the original three-method study was rated by all 20 participants, not one participant. Each participant evaluated 160 cases × 3 methods = 480 outputs. The cases cover 80 videos, foreground/background editing, and ten styles.

Q3. Multi-region composition versus sequential editing

We added this comparison in the appendix Section H “Multi-Region, Multi-Style Stylization.” It uses all four two-mask cases in the LociArt benchmark. Different-style testing covers 360 style assignments across four two-region cases; we also test the same style in both regions.

ProtocolSPSCLIP-TPSNR
Same style, sequential0.4640.17022.45
Same style, compositional0.4670.16824.01
Different styles, sequential0.4520.16722.40
Different styles, compositional0.4290.16027.01

Thus, joint editing improves non-target preservation, but different-style composition involves a style–preservation trade-off. In addition, it needs only one generation run rather than two in sequential editing.

Q4. Transferability and the practical value of Polar

Polar is a general region–style encoding for tasks that take masks as conditioning inputs. It replaces common binary masks with explicit style identities within each region, so its design is not tied to GRNStyle.

RG also has potential to transfer: empty-mask calibration and geometric support address issues that can arise when masks are encoded and compressed in other backbones.

Theoretically, both designs can therefore be adapted to other mask-conditioned models, including VACE, although their gains on each backbone require empirical evaluation.

In practice, Polar reduces the cost of adding new styles by avoiding large-scale retraining on a mixture of old and new data. In our 10-to-15-style expansion experiment (Sec. 5.5), training uses only the five new styles. Polar reduces the old-style SPS drop from 0.102 to 0.022, while maintaining comparable new-style SPS. This allows the style collection to grow without replaying the original training data, while retaining much more of the previously learned capability.

AuthorsOpenReview ↗

Response to Reviewer DYZj - part 2

Q5. Meaning of “free-style”

In this paper, “free-style” refers to flexible regional style control and style expansion, rather than training-free generation of arbitrary styles. Specifically, users can freely choose which regions to stylize, assign different learned styles to different regions, and extend the available style set through paired data and fine-tuning.

At inference, the current model selects from 18 learned styles; a text prompt alone does not introduce an unseen style. However, the style set is extensible rather than permanently fixed. Our replay-free expansion experiment shows that new styles can be learned using only new-style data, while Polar substantially reduces forgetting of existing styles.

We have added this clarification to the abstract and introduction of the revised manuscript. We retain “free-style” to describe the flexibility in region selection, style composition, and style-set expansion, without implying support for arbitrary unseen styles without training.

Q6. VACE-FT in the user study

We have completed the additional evaluation including VACE-FT. Results are means ± reported 95% CI half-widths:

MethodStyle ↑Preserve ↑
CoinVE-Edit2.525 ± 0.2403.144 ± 0.283
SeedDance3.862 ± 0.2212.575 ± 0.211
VACE-FT4.056 ± 0.1454.375 ± 0.127
GRNStyle (ours)4.338 ± 0.1604.475 ± 0.142

VACE-FT ranks second in both style quality and content preservation, while GRNStyle remains the best-performing model.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.