Skip to content

Tune-Your-Style: Intensity-Tunable 3D Style Transfer with Gaussian Splatting

Paper Information

  • Conference: ICCV 2025
  • arXiv: 2602.00618
  • Authors: Yian Zhao, Rushi Ye, Ruochong Zheng, Zesen Cheng, Chaoran Feng, Jiashu Yang, Pengchong Qiao, Chang Liu, Jie Chen
  • Institutions: Peking University Shenzhen Graduate School, Tsinghua University, Dalian University of Technology
  • Code: Project Page
  • Area: 3D Vision / 3D Style Transfer
  • Keywords: 3D Gaussian Splatting, style transfer, intensity-tunable, diffusion model prior, cross-view consistency

TL;DR

This paper proposes Tune-Your-Style, the first intensity-tunable 3D style transfer paradigm, which explicitly models style intensity via Gaussian neurons and parameterizes a learnable style tuner. Combined with a two-stage optimization strategy, the method enables users to freely adjust the degree of style injection without retraining.

Background & Motivation

Problem Definition

3D style transfer aims to transfer the artistic effect of a reference style image onto a 3D scene while preserving content consistency and multi-view consistency. The core challenge lies in balancing content preservation and style injection.

Limitations of Prior Work

Existing mainstream methods (StyleGaussian, G-Style, InstantStyleGaussian, etc.) all adopt a fixed-output paradigm: - Each training run produces a stylized result at a single fixed content–style balance point. - Different users have varying needs for this balance, yet no flexible adjustment is possible. - If a user finds the style too strong or too weak, the only recourse is to retrain the model or manually tune hyperparameters.

Core Idea

An intensity-tunable paradigm is introduced: after a single training run, users can freely adjust style injection intensity via a style tuner (range 0%–100%) without retraining. The key challenge is how to explicitly model the concept of "style intensity."

Method

Overall Architecture

Tune-Your-Style comprises two core components: 1. Intensity-tunable Style Injection (ISI): explicit modeling of style intensity with a learnable style tuner. 2. Tunable Stylization Guidance (TSG): diffusion-model-generated multi-view consistent stylization guidance with a two-stage optimization strategy.

1. Intensity-Tunable Style Injection (ISI)

Gaussian neuron modeling of style intensity: A learnable neuron is assigned to each Gaussian primitive to predict offsets for all its attributes (position, scale, rotation, opacity, color):

\[\mathcal{G}(S_k, \Theta) = \{\Delta_k \boldsymbol{\mu}_i, \Delta_k \boldsymbol{S}_i, \Delta_k \boldsymbol{R}_i, \Delta_k \sigma_i, \Delta_k \boldsymbol{c}_i\}_{i=1}^{N}\]

The stylized scene is the original scene plus these offsets: \(\hat{\Theta}_k = \Theta + \mathcal{G}(S_k, \Theta)\).

Style tuner parameterization: A step function \(\mathcal{H}(\beta)\) is introduced to quantize the continuous style control signal into \(Z=10\) discrete levels, and a bijective mapping to learnable embeddings is established: \(\mathcal{V}_\beta = f(\mathcal{H}(\beta))\):

\[\hat{\Theta}_k^{\beta} = \Theta + \mathcal{V}_\beta \odot \mathcal{G}(S_k, \Theta)\]

When \(\beta = 0\%\), \(\mathcal{V}_\beta\) suppresses the offsets to achieve zero style; when \(\beta = 100\%\), the full offset is retained.

3D Gaussian filter: An importance score \(\psi_i\) is computed for each Gaussian across all training views, and the least important 50% are filtered out to avoid artifacts introduced by redundant Gaussians during stylized rendering:

\[\psi_i = \sum^{C} \sum^{H \times W} \kappa(\Theta_i) \cdot \sigma_i \cdot \prod_{j=1}^{i-1}(1-\sigma_j)\]

2. Tunable Stylization Guidance (TSG)

Diffusion model stylization guidance: IP-Adapter-SDXL is used as the 2D diffusion model to perform style transfer on rendered views, generating stylized views \(\mathcal{I}_v^k\).

Cross-view style alignment: To address the lack of 3D consistency across multi-view outputs from the diffusion model: - A random anchor view is selected, and its features are injected into the self-attention layers of other views. - Cross-view feature matching is performed for content calibration (back-projecting anchor view content features into 3D and re-projecting them onto target views). - Mutual self-attention is applied:

\[\mathcal{A}_{v_c}^{t} = \text{softmax}\left(\frac{Q_{v_c}^{t} \cdot ([K_{v_a \rightarrow v_c}^{t}, K_{v_c}^{t}])^{T}}{\sqrt{d}}\right) \cdot [V_{v_a \rightarrow v_c}^{t}, V_{v_f}^{t}]\]

Two-stage optimization:

Stage 1 (full-style guidance, 2000 steps): - Only the Gaussian neurons \(\mathcal{G}\) and the full-offset embedding \(\mathcal{V}_{full}\) are optimized. - Loss: \(\mathcal{L}_{full}^{s_1} = \mathcal{L}_1(\mathcal{I}_v^{\tilde{t}_1}, \mathcal{I}_v^k) + \mathcal{L}_{lpips}(\mathcal{I}_v^{\tilde{t}_1}, \mathcal{I}_v^k)\)

Stage 2 (tunable guidance, 2000 steps): - The neurons and the full-offset embedding are frozen; only the remaining level embeddings are optimized. - Intermediate \(\beta\) values are randomly sampled, and zero-style and full-style guidance are mixed with a weighted combination:

\[\mathcal{L}_{tunable} = (1-\beta_{\tilde{t}_2}) \cdot \mathcal{L}_{zero} + \beta_{\tilde{t}_2} \cdot \mathcal{L}_{full}^{s_2}\]

Loss & Training

The overall training objective is a weighted combination of L1 loss and LPIPS perceptual loss, applied respectively under zero-style guidance (compared against the original rendering) and full-style guidance (compared against the stylized views).

Key Experimental Results

Main Results: Comparison with 3DGS Style Transfer Methods

Method Short-range Consistency↓ (LPIPS/RMSE) Long-range Consistency↓ (LPIPS/RMSE) CLIP_S↑ CLIP_Sdir↑ User Study↑
StyleGaussian 0.067 / 0.070 0.126 / 0.108 0.2134 0.2223 2.79±0.16
G-Style 0.044 / 0.059 0.093 / 0.096 0.2406 0.2391 3.10±0.40
InstantStyleGaussian 0.053 / 0.062 0.108 / 0.113 0.2204 0.2160 2.06±0.22
Ours 0.033 / 0.035 0.062 / 0.067 0.2619 0.2881 3.97±0.13
  • Multi-view consistency (both short-range and long-range) is comprehensively superior, far outperforming competing methods.
  • CLIP similarity and directional similarity scores are both highest, indicating superior style fidelity.
  • User study score of 3.97/5.0, significantly ahead of all baselines.

Ablation Study

Two-stage optimization ablation: - Removing two-stage optimization (jointly training zero-style and full-style) leads to a severe degradation in stylization quality (over-stylization) and renders the style tuner ineffective. - Root cause: random mixing of zero-style and full-style guidance produces unstable supervision signals.

Cross-view style alignment ablation:

Configuration Effect
Raw diffusion output Visually plausible but lacks 3D consistency
+ Feature injection Good consistency near anchor views; content distortion in distant views
+ Content calibration Content and layout well preserved in distant views; style textures remain consistent

Key Findings

  1. Intensity tunability is genuinely effective: the style tuner smoothly controls style injection intensity from 0% to 100%.
  2. Multi-style composition: combined with SAM segmentation, different styles can be applied to different scene regions and adjusted independently.
  3. Two-stage training is critical: learning the full-style offset stably in the first stage before learning intermediate-level embeddings in the second stage is essential.
  4. Training efficiency: approximately 20 minutes per scene on a single V100 GPU.

Highlights & Insights

  1. Paradigm innovation: the first intensity-tunable 3D style transfer paradigm, representing a qualitative leap from "fixed output" to "continuously adjustable."
  2. Explicit modeling of style intensity: the three-layer design—Gaussian neuron attribute offset prediction, step-function quantization, and learnable embeddings—achieves elegant intensity control.
  3. Cross-view style alignment: the depth-based back-projection–reprojection content calibration scheme effectively resolves multi-view inconsistency in diffusion-generated results.
  4. Full-attribute offsets: offsets are predicted not only for color but also for position, scale, rotation, and opacity, enabling simultaneous transfer of geometric and appearance styles.

Limitations & Future Work

  • The quantization level \(Z=10\) is manually set and may not generalize optimally to all scenes.
  • Generating stylization guidance via the diffusion model increases preprocessing time.
  • The Gaussian filter removes 50% of primitives, which may cause information loss in certain scenes.
  • Fine-grained control for local hierarchical editing is not yet supported.
  • vs. StyleGaussian/G-Style: the proposed method shifts from VGG feature alignment to diffusion prior guidance, avoiding costly encoder–decoder training.
  • vs. SDS/IDU: the proposed cross-view style alignment is better suited to handling fine style textures than SDS loss and iterative data updates.
  • Insight: the Gaussian neuron + style tuner design is extensible to other 3D attribute editing tasks (e.g., lighting, material).
  • Multi-style composition capability has direct application value in the gaming and film industries.

Rating ⭐⭐⭐⭐

Strong novelty (first intensity-tunable 3D style transfer paradigm), comprehensive experiments, and substantial quantitative and user study advantages over baselines. The cross-view style alignment design is elegant. However, some design choices (quantization level, filter ratio) lack sufficient empirical justification.