Styl3R

Instant 3D Stylized Reconstruction for Arbitrary Scenes and Styles

Peng Wang1,2*, Xiang Liu2*, Peidong Liu2
1Zhejiang University     2Westlake University
*Indicates Equal Contribution
arXiv Code

Styl3R predicts stylized 3D Gaussians in less than a second using a feed-forward network given unposed sparse-view images and an arbitrary style image.

Summary

  • We introduce a feed-forward network for 3D stylization that operates on sparse, unposed content images and an arbitrary style image, does not require test-time optimization, and generalizes well to out-of-domain inputs.
  • We design a dual-branch network architecture that decouples appearance and structure modeling, effectively enhancing the joint learning of novel view synthesis and 3D stylization.
  • Our method achieves state-of-the-art zero-shot 3D stylization performance, surpassing existing zero-shot methods and approximate the efficacy of style-specific optimization techniques.

Architecture

Overview of Styl3R. Our model comprises a structure and an appearance branch, each predicting different Gaussian attributes. The structure branch encodes sparse, unposed images with a shared content encoder, then feeds the resulting tokens into per-view structure decoders with cross-view information sharing. Structure-related attributes are regressed from decoder outputs. The appearance branch encodes a style image into tokens, which attend to content tokens from all views in a stylization decoder. The resulting blended tokens predict Gaussian colors. Alternatively, a content image can provide original colors, enabling stylization or reconstruction.

Comparisons with Baselines

We present qualitative comparisons with the following state-of-the-art models:

Stylization Results on RealEstate10K dataset.

Stylization Results on Tanks and Temples dataset.

More Stylization Results

More scene style combinations on out-of-domain data.

Train from Tanks and Temples dataset.

Room from NeRF LLFF dataset.

Style Interpolation

Stylization results interpolated inbetween 3 styles.

BibTeX

@misc{wang2025styl3rinstant3dstylized,
      title={Styl3R: Instant 3D Stylized Reconstruction for Arbitrary Scenes and Styles}, 
      author={Peng Wang and Xiang Liu and Peidong Liu},
      year={2025},
      eprint={2505.21060},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.21060}, 
  }