SceneAligner: 3D-Grounded Floorplan Localization in the Wild
Paper • 2605.22581 • Published • 5
LoRA layers of SceneAligner: 3D-Grounded Floorplan Localization in the Wild (code, project page). Built with DINOv3.
SceneAligner matches the density map of a gravity-aligned 3D reconstruction with a floorplan, using DINOv3 ViT-B/16
with our LoRA layers (rank 16). This repository contains only the LoRA layers (adapter_config.json,
adapter_model.safetensors); the DINOv3 weights are not included. The code repository loads them with
scenealigner.inference.load_model, from timm's copy
of Meta's DINOv3 ViT-B/16 weights or from a local copy of Meta's weights.
Our LoRA layers are a derivative work of the DINOv3 weights and are distributed under the DINOv3 License.
@inproceedings{cho2026scenealigner,
title={SceneAligner: 3D-Grounded Floorplan Localization in the Wild},
author={Junhyeong Cho and Ruojin Cai and Hadar Averbuch-Elor},
booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
year={2026}
}
Base model
facebook/dinov3-vit7b16-pretrain-lvd1689m