README.md

Table of Contents

• Overview
• Full Instructions
  • Environment Setup
  • Dataset Preparation
    • CO3D Dataset
    • DL3DV Dataset
  • Training Guidelines
    • Train Stylos on CO3D
    • Train Stylos on DL3DV
  • Inference and Evaluation
    • Test Data and Model Checkpoints
    • Inference
    • Evaluation
    • Comparison with SOTA Methods
• Quick Inference

Overview

TL;DR. Stylos a single-forward 3D Gaussian framework for 3D style transfer that operates on unposed content, from a single image to a multi-view collection, conditioned on a separate reference style image.

Full Instructions

1. Environment Setup

requirements.txt contains the minimum required packages to train and inference Stylos, which can be installed by running

pip install -r requirements.txt

For cuda-compiled version of RoPE2D, run following command

cd src/model/encoder/backbone/croco/curope
pip install .

2. Dataset Preparation

To modify the dataset location, please edit the following config files, config/dataset/co3d.yaml, config/dataset/dl3dv.yaml, config/dataset/dl3dv960.yaml.

2.1. CO3D Dataset

To download the CO3D dataset, run the following command,

bash scripts/sh_files/datasets/download_co3d.sh

To preprocess the CO3D dataset, run the following command without --dry_run,

python3 -m scripts.python_files.co3d_dataset_preprocess --co3d_root datasets/CO3D --dry_run

Note: Since Stylos is based on Anysplat and VGGT, it does not require image poses during both training and inference.

2.2. DL3DV Dataset

Please download the DL3DV dataset from their official website.

3. Training Guidelines

Please download the checkpoints from huggingface link to Stylos to checkpoints in the current workingspace. We use the pre-trained VGG weights for computing losses.

3.1. Train Stylos on CO3D

We first train Stylos to learn geometry-related knowledge, on 8 NVIDIA H200 GPUs.

python src/main.py +experiment=co3d_geo_global_base trainer.num_nodes=1

After that, we load the pre-trained Stylos weights obtained from the previous step and further train the model for style learning, on 4 NVIDIA GH200 GPUs.

python src/main.py +experiment=co3d_style_3d_loss_4gpus trainer.num_nodes=1

3.2. Train Stylos on DL3DV

We first train Stylos to learn geometry-related knowledge.

python src/main.py +experiment=dl3dv_geo trainer.num_nodes=1

After that, we load the pre-trained Stylos weights obtained from the previous step and further train the model for style learning,

python src/main.py +experiment=dl3dv_style trainer.num_nodes=1

Note: We have trained multiple versions of Stylos on DL3DV, and released two of them. Certain training settings, e.g., loss configurations, number of iterations, number of view, and etc., may vary. Please adust the training hyperparamters according to your needs. In the released training codes, only Wikiart is supported as the style reference while DELAUNAY could be added easily.

4. Inference and Evaluation

4.1. Test Data and Model Checkpoints

The 50 test styles are located in examples/styles. Please copy that folder to the datasets directory.

Download all needed test data and checkpoints in huggingface link to Stylos.

Please use the output/exp_dl3dv_old. checkpoint to reproduce the quantative results of Stylos in the paper.

4.2. Inference

To test Stylos on the CO3D dataset using a frame stride of 3, run:

bash scripts/sh_files/co3d_3d_loss/inference_frame_stride_3.sh

To test Stylos on the TNT dataset, run the following command,

scripts/sh_files/dl3dv2tnt/inference.sh 

4.3. Evaluation

After inference, compute consistency metrics and Artscore on CO3D scenes by running,

scripts/sh_files/co3d_3d_loss/eval_frame_stride_3.sh 

To compute consistency metrics and Artscore of Stylos on each TNT scene, run,

scripts/sh_files/dl3dv2tnt/eval.sh 

The implementation of consistency metrics is modified from StyleGaussian.

4.4. Comparison with SOTA Methods

We reproduce or directly evaluate several 3D stylization models, which include StyleGaussian, G-Style, SGSST, StyleGaussian, and Styl3R. We sincerely appreciate their open-source contributions to the 3D stylization community. The visual results are available at the following link, huggingface link to SOTA comparisons.

Quick Inference

For demo only, please switch to quick_inference branch. To reproduce results in our paper, please use the main branch.

Timeline & TODO

The complete codebase will be fully released soon. We appreciate your patience and interest. Thanks for your attention and support!

• Sep 2025 — Paper available on arXiv
• Jan 2026 — Paper accepted by ICLR 2026
• Jan 2026 — Hugging Face demo released
• Jan 2026 — Inference pipeline released
  (please refer to the quick_inference branch)
• Feb 2026 — Full training code released
• Feb 2026 — Evaluation codes released
• Feb 2026 — Comparison results released
• Feb 2026 — Paper final version available

---

⭐ If you find this project useful, please give us a star to help more people discover it.
👀 You can also watch the repository (top-right corner) to stay updated on new features, papers, and releases.

Your feedback and contributions are always welcome! If you have any question, feel free to leave an issue or email Hanzhou Liu @ hanzhou1996@tamu.edu

Citation

@article{liu2025stylos,
  title={Stylos: Multi-View 3D Stylization with Single-Forward Gaussian Splatting},
  author={Liu, Hanzhou and Huang, Jia and Lu, Mi and Saripalli, Srikanth and Jiang, Peng},
  journal={arXiv preprint arXiv:2509.26455},
  year={2025}
}