GaussianBlender: Instant Stylization of 3D Gaussians with Disentangled Latent Spaces 🎨

Melis Öcal1, Xiaoyan Xing1, Yue Li1, Vien Ngo2, Sezer Karaoğlu1, Theo Gevers1
1University of Amsterdam 2Bosch Center for Artificial Intelligence
arXiv Code

TLDR

  • We introduce, GaussianBlender, a diffusion-based, feed-forward 3D Gaussian splat editor that delivers high-quality stylization instantly at inference, fully eliminates per-asset test-time optimization, generalizes to out-of-domain inputs, supports practical interactive use and large-scale production.
  • We propose an effective strategy for learning 3D diffusion priors in structured, disentangled latent spaces - decoupling appearance from geometry with controlled information sharing - to enable controllable edits and simplify Gaussian parameter learning.

Overview

Overview of our method. (1) Latent space learning: Given input Gaussians, our method first groups them based on spatial proximity and and encodes into group-structured disentangled latent spaces, with controlled cross-branch feature sharing. (2) Latent diffusion pre-training: A denoiser then learns to denoise the noisy appearance latent conditioned on a text embedding. (3) Latent editing: Once 3D priors are captured, denoiser is further trained to learn an editing function that maps the input appearance latent to a modified latent, guided by the geometry latent. At inference, GaussianBlender generates modified high-quality, 3D-consistent assets from text prompts in a single feed-forward pass instantly, fully eliminating test-time optimization. Trainable models at each stage are denoted.

Results

Input Asset

"Make it Starry Night Van Gogh painting style."

Input Asset

"Make it in sunset palette"

Input Asset

"Make it Barbie style."

Input Asset

"Make it in cyberpunk style."

Input Asset

"Make it pop art neon duotone."

Input Asset

"Make its colours look like rainbow."

Input

"Make it marble."

Original

"Make it golden."

Interactive Scene Editing

With its ~0.26 second inference time, GaussianBlender supports interactive scene editing.

Qualitative Comparison

BibTeX

@misc{ocal2025gaussianblenderinstantstylization3d,
      title={GaussianBlender: Instant Stylization of 3D Gaussians with Disentangled Latent Spaces}, 
      author={Melis Ocal and Xiaoyan Xing and Yue Li and Ngo Anh Vien and Sezer Karaoglu and Theo Gevers},
      year={2025},
      eprint={2512.03683},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.03683}, 
}