README.md

metadata

license: apache-2.0
pipeline_tag: image-to-3d

AdaptSplat: Adapting Vision Foundation Models for Feed-Forward 3D Gaussian Splatting

Paper | Code

AdaptSplat is a lightweight adapter design for feed-forward 3D Gaussian Splatting (3DGS). It introduces a single Frequency-Preserving Adapter (FPA) of only 1.5M parameters into a generic architecture to achieve superior performance in cross-domain generalization and high-frequency geometric fidelity. By extracting direction-aware high-frequency structural priors from a vision foundation model backbone (DINOv3-distilled ConvNeXt), it effectively compensates for high-frequency attenuation caused by over-smoothing in deep features.

Inference

To run inference, please refer to the environment setup in the official repository. Once the environment and weights are prepared, you can run inference using the following commands:

Single-GPU Inference

CUDA_VISIBLE_DEVICES=0 python inference.py --config configs/inference.yaml

Multi-GPU (DDP) Inference

torchrun --nproc_per_node=8 inference_ddp.py --config configs/inference.yaml

Citation

If you find this work useful, please cite:

@article{adaptsplat2026,
  title={AdaptSplat: Adapting Vision Foundation Models for Feed-Forward 3D Gaussian Splatting},
  author={Mingwei Xing, Xinliang Wang, Yifeng Shi},
  journal={arXiv preprint arXiv:2605.10239},
  year={2026}
}