acceptodds
Under review as a conference paper at ICLR 2027

BindStyle: SINGLE-IMAGE DYNAMIC GAUSSIAN STYLIZATION VIA MESH-BOUND STYLE PROPAGATION

Abstract

Stylizing a dynamic 3D scene from a single image requires colors, textures, and geometric changes to remain attached to the same scene elements as they move. Independent frame-wise edits do not enforce this association, while appearance-only edits leave spatial support unchanged. We propose , a mesh-bound Gaussian framework that learns time-shared appearance and local geometry from a style reference. Given a pretrained Gaussian scene and a corresponding mesh sequence, the binding transports the learned attributes through the supplied motion while mesh vertices and attachment identities remain fixed. Color changes and geometric changes can explain the same rendered discrepancy, so controls which parameters each supervisory signal updates. Local VGG feature matching guides appearance through spherical-harmonic coefficients, whereas reference-conditioned video diffusion guides local positions, scales, and rotations using mesh depth and temporal context. The reference schedule first adapts appearance, then enables geometry while continuing appearance updates. This gives geometric adaptation an appearance-informed starting point and lets appearance adjust as Gaussian support changes. The resulting model renders the supplied motion without per-frame style optimization or diffusion inference.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Dynamic 3D Stylization, 3D Gaussian Splatting, Single-Image Style Transfer, Mesh-Bound Style Propagation, Video Diffusion Guidance

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WLUROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BindStyle, a framework for dynamic 3D Gaussian stylization from a single style reference image. The proposed method builds on mesh-bound Gaussian Splatting, where Gaussian primitives are attached to mesh surfaces and their appearance and geometry attributes can be propagated through the given mesh motion. To achieve stylization, the authors introduce a parameter-routed supervision strategy, which uses VGG feature matching to optimize Gaussian appearance and video diffusion guidance to optimize local Gaussian geometry, including position, scale and rotation. The authors further adopt a staged optimization strategy to first optimize appearance and then jointly optimize appearance and geometry. Experiments demonstrate improvements in artistic quality and temporal consistency compared with existing stylization methods.

Critical strengths
  1. The paper addresses an interesting problem of dynamic 3D stylization, especially on maintaining consistent appearance and geometry during motion.

  2. The proposed parameter-routed supervision is an interesting design, which separates the optimization of appearance and geometry using different supervision signals.

  3. The experimental results are promising, including comparisons with existing dynamic stylization methods and ablation studies on different optimization strategies.

Critical weaknesses
  1. The paper claims to achieve both appearance and geometry stylization. However, the geometric changes are limited to Gaussian positions, scales and rotations, while the underlying mesh geometry remains unchanged. Besides, most quantitative evaluations focus on rendering quality and temporal consistency. I think it would be important to provide more direct evaluation or visualization of the geometric changes introduced by the proposed method.

  2. The proposed method relies on a given mesh sequence with explicit correspondence, while dynamic baselines such as 4DStyleGaussian and Style4D use learned Gaussian deformation fields without requiring such mesh information. Since the provided mesh correspondence can serve as a strong motion prior, I wonder how much of the improvement in temporal consistency comes from the mesh-bound representation rather than the proposed stylization method. Although Table 4 provides some useful ablations, I think further discussion on this difference would be helpful.

  3. The evaluation is conducted on a relatively small number of scenes and style references. Considering that different artistic styles may require very different geometric and appearance changes, I think a more comprehensive evaluation on diverse styles and complex dynamic scenes would further validate the generalization of the proposed method.

Questions for authors
  1. Is it possible to provide more direct analysis on the geometric stylization? For example, visualizing the Gaussian geometry before and after stylization, or measuring the magnitude of geometric changes.

  2. Could the authors further analyze how much of the temporal consistency improvement comes from mesh binding itself versus the proposed parameter-routed optimization? A comparison under the same mesh correspondence would be helpful.

  3. How does the proposed method perform when the mesh sequence contains inaccurate correspondence or large non-rigid deformations? Since the method strongly relies on mesh binding, I wonder how robust it is to imperfect mesh motion.

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

BindStyle addresses dynamic 3D stylization from a single style reference, given a pretrained Gaussian scene and a corresponding mesh sequence. It learns shared appearance and mesh-local Gaussian parameters that are transported through the supplied motion using persistent mesh attachments.

The method assigns different supervision to different parameter groups. VGG nearest-neighbor feature matching updates spherical-harmonic appearance coefficients, while a reference-conditioned video diffusion teacher updates local Gaussian positions, scales, and rotations.

Optimization first adapts appearance and subsequently enables geometry updates while continuing appearance optimization. Experiments compare stylization quality and spatial–temporal consistency against static and dynamic baselines, with additional controls examining parameter capacity and gradient routing.

Critical strengths
  • The method identifies a concrete optimization ambiguity: rendered discrepancies can be explained by either appearance changes or changes in Gaussian support. Explicitly assigning supervisory signals to parameter groups provides a clear and testable response to this ambiguity.

  • The evaluation includes native dynamic stylization baselines, common-mask consistency measurements, and an unrouted control using the same teachers. These comparisons are more informative than comparisons limited to independent frame-wise stylization or removal of an entire supervision branch.

  • The representation offers a useful inference workflow for assets with reliable mesh correspondence. Once optimized, the edited Gaussians can follow the supplied animation without per-frame stylization, diffusion inference, or additional motion prediction.

Critical weaknesses
  • The contribution is primarily the supervisory assignment and optimization schedule, while motion correspondence and propagation are inherited from the existing mesh-binding framework. Stronger evidence is needed to establish that this assignment provides a generalizable insight beyond an effective configuration of established components.

  • The claimed geometric stylization is not directly evaluated. RGB-based quality and consistency metrics cannot distinguish meaningful stylistic shape changes from adjustments in scale, offset, or visibility that compensate for appearance mismatch, and routing diffusion gradients to geometry does not establish that those gradients encode desirable geometric changes.

  • The comparisons do not fully separate the proposed optimization mechanism from the benefit of additional inputs. BindStyle receives mesh correspondence and depth unavailable to some baselines, while the nine scene–style quality comparisons and three-case routing study provide limited evidence of robustness across different motions, geometries, and styles.

  • The supplied PDF ends after the references, although the main text repeatedly relies on Appendices A–C, additional tables, and Figure 6 for implementation and evaluation details. These materials were therefore unavailable for verification, limiting assessment of the optimization settings, evaluation protocol, statistical analyses, and direct geometry evidence.

Questions for authors

See weaknesses above

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

BindStyle addresses the stylization of dynamic 3D Gaussian Splatting (3DGS) guided by a single reference image, given a known underlying mesh motion. It adopts a MaGS-style mesh-binding approach, where the appearance and local-space support of the Gaussians are propagated alongside the mesh as time-shared parameters, eliminating the need to learn new motion. A style-matching loss is used to update the Spherical Harmonic (SH) appearance, while a combination of style reference, mesh depth, and guidance from a video diffusion model (operating over short temporal windows) is used to update the local position, scale, and rotation of the Gaussians. The optimization process first refines the appearance before enabling geometric updates; during inference, no per-frame style optimization or diffusion-based processing is required. The experimental evaluation includes comparisons against dynamic baselines, assessments of view and temporal consistency, human subjective evaluations, and an ablation study on gradient routing.

Critical strengths

(1) The problem definition and the division of labor within the optimization process are clearly articulated. A grid-based structure handles motion correspondence, while time-shared Gaussian attributes facilitate style propagation. VGG-based matching updates only the Spherical Harmonics (SH) coefficients, whereas signals derived from a reference-conditioned video diffusion model update only the local spatial support. Although this constrained optimization approach integrates off-the-shelf components, it offers a practical design that effectively addresses the tendency for appearance and spatial support to compensate for one another during dynamic stylization.

(2) The key ablation studies are robust and comparable. As shown in Table 4, the comparisons—with and without "Gradient Routing"—utilize identical teacher models, initialization settings, and computational budgets. Across the three matching scenarios, the CM (Common Mask) Loss decreased from 0.801 to 0.715, Temporal LPIPS dropped from 0.183 to 0.146, and RMSE fell from 0.023 to 0.015. Furthermore, the supplementary material provides details on the "common mask" protocol and a paired evaluation involving 24 human participants; this evidence offers a more compelling demonstration that the routing mechanism genuinely improves rendering quality than a mere selection of qualitative visual examples.

Critical weaknesses

(1) The role of the geometric branch remains limited to indirect verification at the rendering level. Video diffusion signals adjust position, scale, and rotation via RGB rendering; in theory, they could also influence appearance loss by altering pixel coverage and visibility, without necessarily inducing geometric deformations that align with the target style. There is currently no direct reporting on Gaussian displacements, scale distributions, or consistent albedo/depth maps at specific time steps; furthermore, the "constant color" mentioned in the appendix still retains spatially varying source colors. Given the absence of explicit surface distance constraints, the plausibility of these geometric changes warrants further investigation.

(2) The overall comparison is constrained by the inclusion of additional geometric inputs and the scale of the experiments. While BindStyle incorporates mesh motion, patch correspondences, and depth, the dynamic baselines primarily rely on their own native representations; consequently, any advantages in temporal consistency cannot be attributed solely to gradient routing or the video teacher model. Although shared masks serve as a useful control, metrics like warped LPIPS/RMSE may still be influenced by texture complexity. The quality assessment is limited to only nine scene-style combinations, and the ablation study on routing covers just three cases—each with three random seeds—which is insufficient to determine whether these performance gains remain stable across diverse motion patterns and styles.

Questions for authors

Q1. Is it possible to provide a side-by-side comparison—under the same camera view—of the albedo, depth, contours, and Gaussian displacement/scale variations? I would like to see evidence that the geometry branch genuinely introduces shape modifications linked to the artistic style, rather than primarily altering the coverage.

Q2. Could you provide a baseline for appearance editing that utilizes the same mesh sequences, depth, and optimization budget, and share the routing comparison results on a per-case, per-seed basis? Additionally, providing statistics on the output textures or edge density would be helpful to rule out the possibility that a reduction in warping error is simply a consequence of texture simplification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.