Multi-StyleGS: Stylized Gaussian Splatting with Multiple Styles

Authors: Yangkai Lin, Jiabao Lei, Kui Jia

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental As demonstrated by our comprehensive experiments, our approach outperforms existing ones in producing plausible stylization results and offering flexible editing. Extensive experiments conducted on various datasets (Knapitsch et al. 2017; Mildenhall et al. 2019) substantiate the efficacy of our method in generating high-quality, locally matched stylized images in real-time.
Researcher Affiliation Academia 1South China University of Technology 2School of Data Science, The Chinese University of Hong Kong, Shenzhen EMAIL, EMAIL, EMAIL
Pseudocode No The paper describes the methodology in prose, detailing the pipeline, Gaussian Splatting with Semantic Features, Preliminary of Style Loss, and Semantic Multi-style Loss, but does not include any formal pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an unambiguous statement about releasing code, nor does it include any links to a code repository.
Open Datasets Yes We conducted extensive experiments on a diverse set of real-world scenes, including outdoor environments from the Tanks and Temples (shortened as tnt in our paper) dataset (Knapitsch et al. 2017) and forward-facing scenes from the llff dataset (Mildenhall et al. 2019).
Dataset Splits No The paper mentions using 'tnt datasets' and 'llff dataset' but does not specify exact training, validation, or testing splits, percentages, or sample counts.
Hardware Specification Yes Our novel semantic style loss can achieve memory-efficient training, which enables efficient training on a single RTX 3090.
Software Dependencies No The paper mentions models like VGG19, DINOv2, and SAM, but does not provide specific version numbers for programming languages, libraries, or other software dependencies used for implementation.
Experiment Setup Yes Our GS model is trained with Lrecon+λseg Lseg+λKNNLKNN+λNELNE+λmask Lmask, (13) where Lrecon is the Mean Squared Error (MSE) reconstruction loss as outlined in (Kerbl et al. 2023). We typically assign values of λseg = 0.02, λKNN = 0.005 , λNE = 0.005.