⚓ NoDrift3R [ECCV 2026]

NoDrift3R: Raymap-Guided Coupling for Drift-Robust Unposed Feed-Forward 3D Reconstruction

Xiangyu Sun¹, Liu Liu², Seungkwon Yang³, Jingbing Han²,
Seungtae Nam³, Zhizhong Su², Eunbyung Park^3
1Sungkyunkwan University, ²Horizon Robotics, ³Yonsei University

The pose-free framework for feed-forward 3D reconstruction. Our method effectively suppresses the pose drift problem, especially in long-sequences.

Installation

Please clone this project with --recursive first.

Then create the environment (example using conda).

conda create -y -n nodrift3r python=3.10
conda activate nodrift3r
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

Checkpoints

We provide two versions of checkpoints on 🤗 Hugging Face.

1. Single dataset version:

• Large Model and Giant Model trained on the dl3dv dataset.

2. Mixed dataset version for zero-shot generalization:

• Large Model and Giant Model trained on the seven mixed dataset, same as E-Rayzer.

Data Preparation

Please download datasets from DL3DV and RE10k.

For mipnerf360 datasets, please download it from Mip-NeRF 360.

Training

To train the giant model on the DL3DV dataset:

python3 -m src.main \
+experiment=dl3dv \
wandb.mode=offline \
wandb.name=dl3dv \
tensorboard.enable=true \
model.encoder.model_type='giant'

To train the large model on the DL3DV dataset:

python3 -m src.main \
+experiment=dl3dv \
wandb.mode=offline \
wandb.name=dl3dv \
tensorboard.enable=true \
model.encoder.model_type='large'

Evaluation

We provide pre-trained weights for NoDrift3R-DL3DV and NoDrift3R-mixed-dataset on Hugging Face.

To evaluate NoDrift3R on the DL3DV dataset:

CUDA_VISIBLE_DEVICES=$gpu_id python3 -m src.main \
+experiment=dl3dv \
mode=test \
wandb.name=dl3dv \
wandb.mode=offline \
dataset/view_sampler@dataset.dl3dv.view_sampler=evaluation \
dataset.dl3dv.view_sampler.index_path=assets/dl3dv_start_0_distance_150_ctx_24v_tgt_8v.json \
dataset.dl3dv.view_sampler.num_context_views=24 \
checkpointing.load=$ckpt_dl3dv \
test.save_image=false \
test.align_pose=true \
test.pose_align_steps=100 \
test.no_pose=true

To evaluate NoDrift3R on the zero-shot MipNeRF360 dataset

CUDA_VISIBLE_DEVICES=$gpu_id python3 -m src.main \
+experiment=mipnerf360 \
model.encoder.model_type=giant \
mode=test \
wandb.name=mipnerf360 \
wandb.mode=offline \
checkpointing.load=$ckpt_mix_dataset \
test.save_image=false \
test.align_pose=true \
test.pose_align_steps=100 \
test.no_pose=true \
test.save_mip360_video=false

Acknowledgement

This work is built on follow amazing research works, thanks a lot to all the authors for sharing!

• Gaussian-Splatting
• Uni3R
• Depth Anything 3
• E-Rayzer
• YoNoSplat

Citation

If you find our work useful, please consider giving a star ⭐ and citing the following paper 📝.

@misc{sun2026nodrift3rraymapguidedcouplingdriftrobust,
      title={NoDrift3R: Raymap-Guided Coupling for Drift-Robust Unposed Feed-Forward 3D Reconstruction}, 
      author={Xiangyu Sun and Liu Liu and Seungkwon Yang and Jingbing Han and Seungtae Nam and Zhizhong Su and Eunbyung Park},
      year={2026},
      eprint={2607.07168},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.07168}, 
}