StructSplat: Generalizable 3D Gaussian Splatting from Uncalibrated Sparse Views

Jia-Chen Zhao^1,2  Beiqi Chen¹  Xinyang Chen¹  Guangcong Wang^2,3  Liqing Nie¹

¹Harbin Institute of Technology (Shenzhen)   ²Great Bay University   ³Guangzhou CloudButterfly Technology Co., Ltd.

 

TL;DR: We present StructSplat, a feed-forward and generalizable NVS framework that predicts 3D gaussians from uncalibrated images without requiring camera parameters.

Abstract

We present StructSplat, a feed-forward and generalizable 3D Gaussian reconstruction framework that operates directly on uncalibrated images without requiring camera parameters. Existing methods either rely on per-scene optimization or assume known camera poses, and often entangle geometry and appearance within a unified backbone, limiting reconstruction fidelity and generalization. Our key idea is to adopt a structured representation that organizes geometry, semantic, and texture cues with explicit roles in the reconstruction process. Specifically, we introduce a pixel-aligned feature injection mechanism to enable accurate texture modeling from 2D observations, incorporate semantic-aware priors to improve global consistency, and design a camera alignment strategy to prevent information leakage and improve generalization. Experiments show that our method significantly outperforms prior approaches on challenging benchmarks.

Demo Video

Installation

Clone Our Codebase

git clone --recursive https://github.com/iLearn-Lab/ECCV26-StructSplat.git
cd StructSplat

Set Up the Environment

conda create -n structsplat python=3.10.19
conda activate structsplat
pip install torch==2.4.0 torchvision==0.19.0 -i https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

Dataset

• DL3DV

  Download training set and evaluation set. Put them into folder data as:

  data
  ├── dl3dv
  │   ├── 1K
  │   ├── 2K
  │   ├── 3K
  │   ├── ...
  │   └── DL3DV-bm
  └── ...
  

Training

• Download pretrained VGGT and Dino V3 checkpoints. Put them into folder ckpts as:

  ckpts
  ├── dinov3_convnext_large
  │   └── ...
  ├── vggt
  │   └── ...
  └── ...
  

• Run the following command to train the model:

  python train.py -c config/dl3dv.yaml
  

• Importan arguments:

  • --configor-c: Configuration file path.
  • --gaussian_training_stage.data.annotations: List of training data annotation file pathes, default: [dataset_annotations/dl3dv_train_clean.json].

Evaluation

• [Optional] Download our pretrained checkpoint. Put it into folder ckpts as:

  ckpts
  ├── dinov3_convnext_large
  │   └── ...
  ├── vggt
  │   └── ...
  └── structsplat
      └── pytorch_model.bin
  

• [Optional] Convert a trained checkpoint from Deepspeed format into binary format:

  python -m deepspeed.utils.zero_to_fp32 "$deepspeed_checkpoint_dir" "ckpts/structsplat" --max_shard_size 10GB
  

• Run the following command to evaluate the model:

  python evaluation.py -c config/dl3dv.yaml
  

• Importan arguments:
  • --configor-c: Configuration file path.
  • --gaussian_evaluation_stage.data.annotations: List of evaluation data annotation file pathes, default: [dataset_annotations/dl3dv_eva_src-2_tar-2.json].
  • --gaussian_evaluation_stage.ckpt: Checkpoint file path, default: ckpts/structsplat/pytorch_model.bin.

Citation

@inproceedings{zhao2026structsplat,
 title={StructSplat: Generalizable 3D Gaussian Splatting from Uncalibrated Sparse Views},
 author={Zhao, Jia-Chen and Chen, Beiqi and Chen, Xinyang and Wang, Guangcong and Nie, Liqing},
 booktitle={European Conference on Computer Vision},
 year={2026}
}