Style4D-Bench: 4D Stylization Benchmark

• Style4D-Bench is a benchmark that evaluates 4D stylization by preserving scene geometry, ensuring temporal stability, and achieving multi-view consistency in dynamic 3D scenes.
• It integrates a benchmark protocol, curated high-resolution dynamic scenes, and a baseline method built on 4D Gaussian Splatting to address challenges beyond 2D video and static 3D stylization.
• The suite employs comprehensive metrics and user studies to assess imaging quality, aesthetic appeal, and consistency, providing actionable insights for stylization research.

Style4D-Bench is a benchmark suite for 4D stylization, defined as stylizing dynamic 3D scenes over time—$4\mathrm{D} = 3\mathrm{D}$ space $+ 1\mathrm{D}$ time—so that rendered outputs from arbitrary viewpoints and times preserve spatial fidelity, temporal coherence, and multi-view consistency. It is presented as the first benchmark suite specifically designed for 4D stylization and combines a benchmark protocol, a curated set of high-resolution dynamic scenes, and a baseline method, Style4D, built upon 4D Gaussian Splatting (4DGS) (Chen et al., 26 Aug 2025).

1. Scope and problem formulation

In this setting, spatial fidelity denotes preservation of geometry and structural details of the original scene; temporal coherence denotes the avoidance of flickering or jitter across time; and multi-view consistency denotes stylistic consistency across viewpoints. The benchmark therefore targets a joint problem: stylized rendering of dynamic 3D scenes under both view variation and temporal evolution (Chen et al., 26 Aug 2025).

The paper distinguishes this problem from two adjacent regimes. In 2D video stylization, computation occurs in image space; temporal constraints may be imposed, but the methods lack explicit 3D geometric reasoning and therefore cannot guarantee multi-view consistency or geometry-faithful stylization for novel views. In 3D/NeRF stylization, multi-view consistency can be enforced for static scenes, but temporal dynamics are typically ignored, so temporal coherence under motion, deformation, and occlusion is not guaranteed. 4D stylization is therefore defined by the simultaneous requirement to satisfy 3D multi-view constraints and 1D temporal stability on dynamic scenes.

Within this formulation, 4DGS functions as the geometry backbone. The benchmark describes 4DGS as extending 3D Gaussian Splatting to dynamic scenes by combining canonical geometry and appearance with a deformation field over time. In Style4D, this provides both accurate dynamic-scene reconstruction and a differentiable renderer whose appearance can be modulated for stylization while preserving geometry and enabling consistent multi-view rendering. This suggests that the benchmark treats geometry preservation not as an auxiliary desideratum but as a central precondition for evaluating stylization in dynamic 3D scenes.

2. Benchmark composition and data regime

Style4D-Bench is instantiated on the Neu3D dataset, curated here as dynamic 4D scenes for stylization evaluation. The benchmark contains 6 dynamic scenes: cook_spinach, flame_salmon_1, sear_steak, flame_steak, coffee_martini, and cut_roasted_beef. Each scene contains 15–20 static cameras distributed in space, 300 frames per scene, and a frame resolution of $1352 \times 1014$. The scenes exhibit diverse motions, including cooking actions and pouring, together with complex backgrounds, nontrivial viewpoint variation, and occlusion (Chen et al., 26 Aug 2025).

Evaluation uses both fixed test viewpoints and helical/spiral novel-view trajectories rendered across time. The fixed viewpoints probe held-out-view generalization, while the helical or spiral trajectories probe multi-view consistency and temporal stability under continuous camera motion. The use of long sequences at native resolution is a defining property of the suite.

Style specification is image-based rather than prompt-based. Style images are sampled from WikiArt to define target styles, and content images used to train the 2D stylization module are drawn from COCO2014. The paper evaluates similarity between stylized outputs and the target style using CKDN and LPIPS-to-style. It does not enumerate fixed style categories, and no text prompts are used.

The project page is listed as https://becky-catherine.github.io/Style4D/. The dynamic scene data derive from Neu3D, and usage follows Neu3D’s license and terms. The paper states that the benchmark protocol, baseline code pointers, and evaluation details are provided via the project page.

3. Evaluation protocol and metric decomposition

The benchmark decomposes evaluation into “4D stylization quality” and “4D stylization consistency”, comprising six dimensions and 12 metrics. Scores are computed per frame and averaged over sequences; for multi-view assessment, they are averaged over frames along spiral trajectories. The paper explicitly states that it does not use FID or CLIPScore, and that PSNR is used only in an auxiliary reconstruction ablation, not as a primary stylization metric (Chen et al., 26 Aug 2025).

Dimension	Metrics	Direction
Imaging Quality	UIQM, CLIP-IQA+, MUSIQ	Higher is better
Aesthetic Quality	Q-Align, MUSIQ-PAQ2PIQ	Higher is better
Spatiotemporal Consistency	DISTS, Warp loss	DISTS lower, Warp lower
Subject Consistency	Inter-frame DINO feature similarity	Higher is better
Style Consistency	CKDN, LPIPS-to-style	CKDN higher, LPIPS lower
Content Consistency	SSIM, LPIPS-to-content	SSIM higher, LPIPS lower

Imaging quality is measured with UIQM, CLIP-IQA+, and MUSIQ. UIQM is described as a no-reference score combining colorfulness, sharpness, and contrast. CLIP-IQA+ is a CLIP-based no-reference image quality assessment measure. MUSIQ is identified as a multi-scale transformer-based NIQE. Aesthetic quality is measured with Q-Align, an LMM-based discrete-level aesthetic assessment, and MUSIQ-PAQ2PIQ, trained on PaQ-2-PiQ for aesthetics.

Spatiotemporal consistency is assessed with DISTS and an optical-flow-based Temporal Warping Error on helical-trajectory videos. Using consecutive frames $I_t$ and $I_{t+1}$, and RAFT optical flow $F_{t \to t+1}$, the benchmark defines:

$\hat{I}_{t+1}(x) = I_{t+1}(x + F_{t \to t+1}(x))$

$$L_{\mathrm{warp}}^t = \frac{1}{H W}\sum_x \left| \hat{I}_{t+1}(x) - I_t(x) \right|$$

$$L_{\mathrm{warp}} = \frac{1}{T-1}\sum_{t=1}^{T-1} L_{\mathrm{warp}}^t$$

DISTS is used as a reference-based perceptual similarity between consecutive frames, with lower values preferred when comparing frame-to-warped-frame. Subject consistency is measured by inter-frame DINO feature similarity at a fixed viewpoint.

For style consistency, the suite uses CKDN and LPIPS-to-style. For content consistency, it uses SSIM and LPIPS-to-content, both computed against the original content frame from the same camera and time.

Human perceptual evaluation supplements the metric suite. The study uses 34 participants in two parts. In the single-frame study, participants evaluate 5 pairs of images, with four extracted frames per video, rating stylization quality and image quality. In the long-video study, participants evaluate two 10-second pairs containing 300 frames each, rating stylization quality, spatiotemporal consistency, and video quality. Results are reported as voting percentages.

4. Style4D baseline architecture

To establish a strong baseline, the paper introduces Style4D, a 4D stylization framework built upon 4DGS. The framework contains three components: a basic 4DGS scene representation to capture reliable geometry, a Style Gaussian Representation using lightweight per-Gaussian MLPs for temporally and spatially aware appearance control, and a Holistic Geometry-Preserved Style Transfer module (HGST) designed to enhance spatio-temporal consistency via contrastive coherence learning and structural content preservation (Chen et al., 26 Aug 2025).

Base 4DGS representation

The canonical or static Gaussians per scene are defined as

$G_i = \{ \mu_i, r_i, s_i, o_i, c_i^{sh} \},$

where $+ 1\mathrm{D}$0 is the center, $+ 1\mathrm{D}$1 the rotation, $+ 1\mathrm{D}$2 the scale, $+ 1\mathrm{D}$3 the opacity, and $+ 1\mathrm{D}$4 the spherical-harmonic coefficients for view-dependent color. A deformation field network $+ 1\mathrm{D}$5 warps the Gaussians across time.

Rendering uses front-to-back alpha compositing of rasterized Gaussian splats:

$+ 1\mathrm{D}$6

where $+ 1\mathrm{D}$7 is the depth-sorted neighborhood of Gaussians contributing to pixel $+ 1\mathrm{D}$8.

Style Gaussian Representation

Each Gaussian receives a lightweight, per-Gaussian MLP with a style code $+ 1\mathrm{D}$9 that modulates appearance over space-time while preserving geometry. The inputs include time $1352 \times 1014$0 and the ray–ellipsoid intersection depth $1352 \times 1014$1, computed from pixel $1352 \times 1014$2 and camera pose $1352 \times 1014$3, following the 3D/4DGS ray–Gaussian intersection. Rendering with style modulation is defined as:

$1352 \times 1014$4

Here $1352 \times 1014$5 is the base view-dependent color, $1352 \times 1014$6 is the MLP-predicted color increment, and $1352 \times 1014$7 together with $1352 \times 1014$8 denotes opacities modulated per Gaussian by the MLPs.

The paper states that this formulation preserves 3D geometry because modulation is applied at intersections and uses depth and time, enables temporally aware stylization, and improves multi-view consistency. In the reported experiments, each per-Gaussian MLP consists of two layers. The appendix specifies that the MLP takes time and intersection depth, uses 4 hidden units, and outputs 4 channels corresponding to RGB deltas and opacity.

Holistic Geometry-Preserved Style Transfer

HGST is an encoder–transformer–decoder stylization backbone that stylizes 2D frames while preserving structure and coherence. It fuses content and style via multichannel correlation and is regularized by two key consistency terms.

The local contrastive loss aligns local differences using an attention-guided design with CBAM attention. For features at scales $1352 \times 1014$9, after sampling $I_t$0 locations $I_t$1 and their $I_t$2-NN $I_t$3, and defining local differences $I_t$4 and $I_t$5, the loss is:

$I_t$6

The global content loss preserves structure:

$I_t$7

These define the combined consistency loss:

$I_t$8

Additional stylization objectives are also included:

• Style perceptual loss:

$I_t$9

• Identity loss:

$I_{t+1}$0

• Illumination loss:

$I_{t+1}$1

• Inner-channel similarity loss $I_{t+1}$2

The total HGST objective is:

$I_{t+1}$3

with appendix weights

$I_{t+1}$4

The paper’s stated rationale is sequential: geometry is learned once by 4DGS from content images; HGST produces structurally faithful, temporally consistent stylized supervision frames; and the Style Gaussian Representation applies per-Gaussian, depth- and time-aware modulation to yield fine, view-consistent style details without geometry drift.

5. Training protocol, rendering regime, and reproducibility

Training is staged sequentially. First, 4DGS geometry is trained on multi-view content images $I_{t+1}$5 using the 4DGS setup, with hyperparameters following Wu et al., CVPR’24. Second, HGST is trained on COCO2014 content images and WikiArt style images using $I_{t+1}$6 crops. Third, the Style Gaussian module is trained against HGST-generated stylized frames (Chen et al., 26 Aug 2025).

For stylization supervision, the Style4D objective is:

$I_{t+1}$7

where $I_{t+1}$8 is the rendered image from Style4D and $I_{t+1}$9 is the HGST stylized target, and $F_{t \to t+1}$0 is total variation regularization.

Initialization uses the official 4DGS point cloud and downsamples it to approximately 3–4k Gaussians per scene. The per-Gaussian MLPs are again specified as two-layer networks, also reported with 4 hidden units and 4 outputs, trained with a learning-rate schedule from $F_{t \to t+1}$1 to $F_{t \to t+1}$2 and delay_mult $F_{t \to t+1}$3.

The main training logistics are explicit:

• Batch size: $F_{t \to t+1}$4
• Iterations: up to 14,000 per scene
• Hardware: single NVIDIA A40 (48 GB)
• Training time: approximately 2 hours per scene
• Peak memory: approximately 10 GB

Inference is performed at the native Neu3D resolution of $F_{t \to t+1}$5. Although the method builds on real-time 4DGS rendering, the paper does not report frames per second for Style4D. Rendering complexity is described as scaling with the number of Gaussians and the cost of the per-Gaussian MLPs.

The project page is listed as the primary access point for the benchmark protocol, baseline, and links. The data are obtained under Neu3D’s license and terms. The setup is summarized as: Stage 1, train 4DGS on content frames; Stage 2, train HGST on COCO2014 and WikiArt with the specified loss weights; Stage 3, train per-Gaussian MLPs with $F_{t \to t+1}$6 against HGST stylized supervision.

6. Quantitative performance, user studies, and limitations

Quantitative comparisons are reported on three representative scenes—cook_spinach, flame_salmon_1, and sear_steak—against 4DGS(AdaIN), 4DGS(AdaAttN), and 4DStyleGaussian. Across these scenes, Style4D attains the best or tied-best results across most metrics, especially on spatial fidelity measured by SSIM and LPIPS-to-content, aesthetics, CLIP-IQA+, MUSIQ, and DISTS (Chen et al., 26 Aug 2025).

For cook_spinach, Style4D reports UIQM 1.9290, CLIP-IQA+ 0.4437, MUSIQ 53.4681, Q-Align 3.2072, MUSIQ-PAQ2PIQ 65.8520, DISTS 0.0112, Warp 0.0058, DINO 0.9395, CKDN 0.2290, LPIPS-to-style 0.6866, SSIM 0.7771, and LPIPS-to-content 0.1834. For flame_salmon_1, the corresponding values are 1.7529, 0.3962, 55.2302, 3.6030, 63.5178, 0.0138, 0.0067, 0.9415, 0.2354, 0.7602, 0.6963, and 0.2704. For sear_steak, Style4D reports 1.6818, 0.4176, 53.3443, 2.8820, 68.0488, 0.0108, 0.0066, 0.9564, 0.3239, 0.7014, 0.7503, and 0.2146, respectively. The paper notes a single clear exception: on sear_steak, 4DStyleGaussian has the lowest warp loss (0.0050) while Style4D is second-best, though still best on DISTS.

The user studies reinforce the quantitative findings. In the single-frame study, voting percentages for stylization quality / image quality are 11.76% / 2.94% for 4DGS(AdaIN), 14.70% / 8.82% for 4DGS(AdaAttN), 11.76% / 17.64% for 4DStyleGaussian, and 61.76% / 70.58% for Style4D. In the video study, pairwise against Style4D, the percentages for stylization quality / video quality / spatiotemporal consistency are:

• vs 4DStyleGaussian: 23.52% / 30.12% / 44.11% for 4DStyleGaussian versus 76.47% / 69.87% / 55.88% for Style4D
• vs 4DGS(AdaIN): 20.58% / 11.76% / 14.70% versus 79.41% / 88.23% / 85.29%
• vs 4DGS(AdaAttN): 14.70% / 17.64% / 14.70% versus 85.29% / 82.35% / 85.29%

The ablations further isolate component behavior. For 2D video stylization temporal warp loss, the appendix reports AdaIN 0.045824, AdaAttN 0.031671, CCPL 0.013301, MCCNet 0.021858, and HGST 0.021068. For content preservation on 4D stylization, Style4D exceeds the listed baselines on all three benchmarked scenes in SSIM and LPIPS-to-content. For style loss, measured as VGG Gram-matrix MSE to the style image, Style4D is lower than 4DStyleGaussian on sear_steak (0.005687 vs 0.006816), cook_spinach (0.006123 vs 0.008329), and flame_salmon_1 (0.006364 vs 0.007387). An auxiliary reconstruction ablation shows that the style-aware Gaussian representation does not harm rendering quality, with PSNR improvements over 4DGS on all six original scenes, including flame_steak (32.85 vs 32.02) and cut_roasted_beef (32.81 vs 32.55).

The paper identifies several limitations. Computational cost remains nontrivial because of multi-stage training and per-Gaussian MLP overhead; the reported cost is approximately two hours per scene on an A40 with about 10 GB peak memory. Style granularity is limited to a fixed style per scene; rapid style switching and region-specific stylization are not supported. The paper also notes that, although temporal stability is strong, very fast motions or extremely complex occlusions remain challenging in general. The stated future direction is to improve stylization quality further and enable broader, user-controllable style manipulation for interactive applications.

Taken together, the benchmark’s design emphasizes standardized evaluation of dynamic 3D stylization through simultaneous measurement of geometry-faithful rendering, temporal stability, and view consistency. A plausible implication is that Style4D-Bench’s principal contribution lies not only in reporting a strong baseline, but in formalizing a metric and data regime under which future 4D stylization methods can be compared on the same spatiotemporal and multiview criteria.