Styl3R: Instant 3D Stylized Reconstruction for Arbitrary Scenes and Styles

Peng Wang* · Xiang Liu* · Peidong Liu

(* Equal Contribution)

NeurIPS 2025

Paper | Project Page

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.
🤖 Our method could potentially improve the efficiency of domain randomization in sim2real transfer as in Deep Drone Racing.

📢 Updates

• 2025.10.31: The training and inference code is updated. Please inform us if you have encountered any issues.

⚙️ Installation

Our code relies on Python 3.10+, and is developed based on PyTorch 2.2.0 and CUDA 12.1, but it should work with higher Pytorch/CUDA versions as well.

1. Clone Styl3R.

git clone https://github.com/WU-CVGL/Styl3R
cd Styl3R

2. Create the environment, here we show an example using conda.

conda create -y -n styl3r python=3.10
conda activate styl3r
pip install torch==2.2.0 torchvision==0.17.2 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

3. Optional, compile the cuda kernels for RoPE (as in CroCo v2).

# Styl3R relies on RoPE positional embeddings for which you can compile some cuda kernels for faster runtime.
cd src/model/encoder/backbone/croco/curope/
python setup.py build_ext --inplace
cd ../../../../../..

📦 Pre-trained Checkpoints

Our models are hosted on Hugging Face 🤗

Model name	Training resolutions	Training data	Input view num.
re10k_2v.ckpt	256x256	re10k	2
re10k_4v.ckpt	256x256	re10k	4
re10k_dl3dv_4v.ckpt	256x256	re10k, dl3dv	4

📚 Datasets

Please refer to DATASETS.md for detailed dataset preparation.

After setting up the datasets (RE10K, DL3DV, and WikiArt):

1. Modify the dataset paths marked with # TODO in
   generate_scene_style_correspondences.py.
   Then, run the script to generate scene-style correspondence .json files.

2. Update the roots paths in:

   • re10k_style.yaml
   • dl3dv_style.yaml

3. Update the style_root paths in:

   • re10k_3view_style_1x.yaml
   • re10k_3view_style_8x8.yaml
   • re10k_dl3dv_3view_style_1x.yaml
   • re10k_dl3dv_3view_style_8x8.yaml

🚀 Training

First download the MASt3R pretrained model and put it in the ./ckpts directory.

Stage 1: NVS Pretraining

(1) 2-view Model (RE10K)

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5 python -m src.main_style +experiment=re10k_3view_style_8x8 \
    wandb.mode=online \
    wandb.project=noposplat_xiang_token_style_pretrain \
    wandb.name=re10k_multi-view_tok-sty-NVS-pretrain \
    data_loader.train.batch_size=10 \
    dataset.re10k_style.view_sampler.num_context_views=2 \
    trainer.val_check_interval=500 \
    checkpointing.every_n_train_steps=3125 \
    model.encoder.stylized=False \
    model.encoder.gaussian_adapter.sh_degree=0

(2) 4-view Model (RE10K)

Initialize from the pretrained 2-view model by setting checkpointing.load to the corresponding checkpoint path.

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m src.main_style +experiment=re10k_3view_style_8x8 \
    wandb.mode=online \
    wandb.project=noposplat_xiang_token_style_pretrain \
    wandb.name=re10k_4multi-view_tok-sty-NVS-pretrain \
    data_loader.train.batch_size=6 \
    dataset.re10k_style.view_sampler.num_context_views=4 \
    dataset.re10k_style.view_sampler.num_target_views=6 \
    trainer.val_check_interval=500 \
    checkpointing.every_n_train_steps=3125 \
    model.encoder.stylized=False \
    model.encoder.gaussian_adapter.sh_degree=0 \
    checkpointing.load='outputs/exp_re10k_multi-view_tok-sty-NVS-pretrain/2025-04-29_19-47-17/checkpoints/epoch_0-step_15000.ckpt'

(3) 4-view Model (RE10K + DL3DV)

Initialize from the pretrained 4-view RE10K model.

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m src.main_style +experiment=re10k_dl3dv_3view_style_8x8 \
    wandb.mode=online \
    wandb.project=noposplat_xiang_token_style_pretrain \
    wandb.name=re10k_dl3dv_4multi-view_tok-sty-NVS-pretrain \
    data_loader.train.batch_size=3 \
    dataset.re10k_style.view_sampler.num_context_views=4 \
    dataset.re10k_style.view_sampler.num_target_views=6 \
    dataset.dl3dv_style.view_sampler.num_context_views=4 \
    dataset.dl3dv_style.view_sampler.num_target_views=6 \
    trainer.val_check_interval=500 \
    checkpointing.every_n_train_steps=3125 \
    model.encoder.stylized=False \
    model.encoder.gaussian_adapter.sh_degree=0 \
    checkpointing.load='outputs/exp_re10k_4multi-view_tok-sty-NVS-pretrain/2025-05-03_14-44-57/checkpoints/epoch_0-step_18750.ckpt'

Stage 2: Stylization Fine-tuning

(1) 2-view Model (RE10K)

Initialize from the 2-view NVS-pretrained checkpoint.

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5 python -m src.main_style +experiment=re10k_3view_style_8x8 \
    wandb.mode=online \
    wandb.project=noposplat_xiang_token_style_debug \
    wandb.name=re10k_multi-view_tok-sty-stylization_content-h3-h4 \
    data_loader.train.batch_size=14 \
    dataset.re10k_style.view_sampler.num_context_views=2 \
    trainer.max_steps=35001 \
    trainer.val_check_interval=500 \
    checkpointing.every_n_train_steps=5000 \
    model.encoder.stylized=True \
    model.encoder.gaussian_adapter.sh_degree=0 \
    loss=style \
    train.identity_loss=true \
    checkpointing.load='outputs/exp_re10k_multi-view_tok-sty-NVS-pretrain/2025-04-29_19-47-17/checkpoints/epoch_0-step_15000.ckpt'

(2) 4-view Model (RE10K)

Initialize from the 4-view NVS-pretrained checkpoint.

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m src.main_style +experiment=re10k_3view_style_8x8 \
    wandb.mode=online \
    wandb.project=noposplat_xiang_token_style_debug \
    wandb.name=re10k_4multi-view_tok-sty-stylization \
    data_loader.train.batch_size=8 \
    dataset.re10k_style.view_sampler.num_context_views=4 \
    dataset.re10k_style.view_sampler.num_target_views=6 \
    trainer.max_steps=35001 \
    trainer.val_check_interval=500 \
    checkpointing.every_n_train_steps=5000 \
    model.encoder.stylized=True \
    model.encoder.gaussian_adapter.sh_degree=0 \
    loss=style \
    train.identity_loss=true \
    checkpointing.load='outputs/exp_re10k_4multi-view_tok-sty-NVS-pretrain/2025-05-03_14-44-57/checkpoints/epoch_0-step_18750.ckpt'

(3) 4-view Model (RE10K + DL3DV)

Initialize from the 4-view NVS-pretrained checkpoint on RE10K + DL3DV.

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m src.main_style +experiment=re10k_dl3dv_3view_style_8x8 \
    wandb.mode=online \
    wandb.project=noposplat_xiang_token_style_debug \
    wandb.name=re10k_dl3dv_4multi-view_tok-sty-stylization \
    data_loader.train.batch_size=3 \
    dataset.re10k_style.view_sampler.num_context_views=4 \
    dataset.re10k_style.view_sampler.num_target_views=6 \
    dataset.dl3dv_style.view_sampler.num_context_views=4 \
    dataset.dl3dv_style.view_sampler.num_target_views=6 \
    trainer.max_steps=35001 \
    trainer.val_check_interval=500 \
    checkpointing.every_n_train_steps=5000 \
    model.encoder.stylized=True \
    model.encoder.gaussian_adapter.sh_degree=0 \
    loss=style \
    train.identity_loss=true \
    checkpointing.load='outputs/exp_re10k_dl3dv_4multi-view_tok-sty-NVS-pretrain/2025-05-05_21-11-11/checkpoints/epoch_0-step_15625.ckpt'

🎨 Inference

Note: Set checkpointing.load to a checkpoint obtained after stylization fine-tuning.

Stylize a scene in RE10K

CUDA_VISIBLE_DEVICES=0 python -m infer_model_re10k \
    +experiment=re10k_3view_style_1x.yaml \
    wandb.name=re10k_tok-sty_inference \
    model.encoder.stylized=True \
    model.encoder.gaussian_adapter.sh_degree=0 \
    test.pose_align_steps=50 \
    checkpointing.load='outputs/exp_re10k_dl3dv_4multi-view_tok-sty-stylization_b2x3/2025-05-06_19-31-34/checkpoints/epoch_0-step_35000.ckpt'

infer_re10k.mp4

Stylize a scene in Tanks and Temples (COLMAP format)

CUDA_VISIBLE_DEVICES=0 python -m infer_model_colmap \
    +experiment=re10k_3view_style_1x.yaml \
    wandb.name=video_colmap_tok-sty_inference \
    model.encoder.stylized=True \
    model.encoder.gaussian_adapter.sh_degree=0 \
    test.pose_align_steps=50 \
    checkpointing.load='outputs/exp_re10k_dl3dv_4multi-view_tok-sty-stylization_b2x3/2025-05-06_19-31-34/checkpoints/epoch_0-step_35000.ckpt'

infer_tnt.mp4

🧩 Debugging

A launch.json configuration is provided for debugging all training and inference commands in VSCode.

• Update the python field to your Python interpreter path.
• Update checkpointing.load to your corresponding checkpoint path.

🚧 TODO

• Release inference code and pretrained models
• Release gradio demo
• Release training code

🙏 Acknowledgements

This project is based on NoPoSplat. We sincerely thank the authors for open-sourcing their excellent work.

📖 Citation

If you find our work useful, please consider citing our paper:

@article{wang2025styl3r,
  title={Styl3R: Instant 3D Stylized Reconstruction for Arbitrary Scenes and Styles},
  author={Wang, Peng and Liu, Xiang and Liu, Peidong},
  journal={arXiv preprint arXiv:2505.21060},
  year={2025}
}