Official implementation of Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Pass, CVPR 2025
Jianing Yang, Alexander Sax, Kevin J. Liang, Mikael Henaff, Hao Tang, Ang Cao, Joyce Chai, Franziska Meier, Matt Feiszli
# clone project
git clone https://github.com/facebookresearch/fast3r
cd fast3r
# create conda environment
conda create -n fast3r python=3.11 cmake=3.14.0 -y
conda activate fast3r
# install PyTorch (adjust cuda version according to your system)
conda install pytorch torchvision torchaudio pytorch-cuda=12.4 nvidia/label/cuda-12.4.0::cuda-toolkit -c pytorch -c nvidia
# install requirements
pip install -r requirements.txt
# install fast3r as a package (so you can import fast3r and use it in your own project)
pip install -e .Note: Please make sure to NOT install the cuROPE module like in DUSt3R - it would mess up Fast3R's prediction.
Use the following command to run the demo:
python fast3r/viz/demo.pyThis will automatically download the pre-trained model weights and config from Hugging Face Model.
The demo is a Gradio interface where you can upload images or a video and visualize the 3D reconstruction and camera pose estimation.
fast3r/viz/demo.py also serves as an example of how to use the model for inference.
Left: Upload a video. Right: Visualize the 3D Reconstruction
Click here to see example of: visualize confidence heatmap + play frame by frame + render a GIF
To use Fast3R in your own project, you can import the Fast3R class from fast3r.models.fast3r and use it as a regular PyTorch model.
import torch
from fast3r.dust3r.utils.image import load_images
from fast3r.dust3r.inference_multiview import inference
from fast3r.models.fast3r import Fast3R
from fast3r.models.multiview_dust3r_module import MultiViewDUSt3RLitModule
# --- Setup ---
# Load the model from Hugging Face
model = Fast3R.from_pretrained("jedyang97/Fast3R_ViT_Large_512") # If you have networking issues, try pre-download the HF checkpoint dir and change the path here to a local directory
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)
# Create a lightweight lightning module wrapper for the model.
# This provides functions to estimate camera poses, evaluate 3D reconstruction, etc.
lit_module = MultiViewDUSt3RLitModule.load_for_inference(model)
# Set model to evaluation mode
model.eval()
lit_module.eval()
# --- Load Images ---
# Provide a list of image file paths. Images can come from different cameras and aspect ratios.
filelist = ["path/to/image1.jpg", "path/to/image2.jpg", "path/to/image3.jpg"]
images = load_images(filelist, size=512, verbose=True)
# --- Run Inference ---
# The inference function returns a dictionary with predictions and view information.
output_dict, profiling_info = inference(
images,
model,
device,
dtype=torch.float32, # or use torch.bfloat16 if supported
verbose=True,
profiling=True,
)
# --- Estimate Camera Poses ---
# This step estimates the camera-to-world (c2w) poses for each view using PnP.
poses_c2w_batch, estimated_focals = MultiViewDUSt3RLitModule.estimate_camera_poses(
output_dict['preds'],
niter_PnP=100,
focal_length_estimation_method='first_view_from_global_head'
)
# poses_c2w_batch is a list; the first element contains the estimated poses for each view.
camera_poses = poses_c2w_batch[0]
# Print camera poses for all views.
for view_idx, pose in enumerate(camera_poses):
print(f"Camera Pose for view {view_idx}:")
print(pose.shape) # np.array of shape (4, 4), the camera-to-world transformation matrix
# --- Extract 3D Point Clouds for Each View ---
# Each element in output_dict['preds'] corresponds to a view's point map.
for view_idx, pred in enumerate(output_dict['preds']):
point_cloud = pred['pts3d_in_other_view'].cpu().numpy()
print(f"Point Cloud Shape for view {view_idx}: {point_cloud.shape}") # shape: (1, 368, 512, 3), i.e., (1, Height, Width, XYZ)Train model with chosen experiment configuration from configs/experiment/
python fast3r/train.py experiment=super_long_training/super_long_trainingYou can override any parameter from command line following Hydra override syntax:
python fast3r/train.py experiment=super_long_training/super_long_training trainer.max_epochs=20 trainer.num_nodes=2To submit a multi-node training job with Slurm, use the following command:
python scripts/slurm/submit_train.py --nodes=<NODES> --experiment=<EXPERIMENT>After training, you can run the demo with a lightning checkpoint with the following command:
python fast3r/viz/demo.py --is_lightning_checkpoint --checkpoint_dir=/path/to/super_long_training_999999To evaluate on 3D reconstruction or camera pose estimation tasks, run:
python fast3r/eval.py eval=<eval_config><eval_config> can be any of the evaluation configurations in configs/eval/. For example:
ablation_recon_better_inference_hp/ablation_recon_better_inference_hpevaluates the 3D reconstruction on DTU, 7-Scenes and Neural-RGBD datasets.eval_cam_pose/eval_cam_pose_10viewsevaluates the camera pose estimation on 10 views on CO3D dataset.
To evaluate camera poses on RealEstate10K dataset, run:
python scripts/fast3r_re10k_pose_eval.py --subset_file scripts/re10k_test_1800.txtTo evaluate multi-view depth estimation on Tanks and Temples, ETH-3D, DTU, and ScanNet datasets, follow the data download and preparation guide of robustmvd, install that repo's requirements.txt into the current conda environment, and run:
python scripts/robustmvd_eval.pyPlease follow DUSt3R's data preprocessing instructions to prepare the data for training and evaluation. The pre-processed data is compatible with the multi-view dataloaders in this repo.
For preprocessing the DTU, 7-Scene, and NRGBD datasets for evaluation, we follow Spann3r's data processing instructions.
- Q:
httpcore.ConnectError: All connection attempts failedwhen launching the demo?- See #34. Download the example videos into a local directory.
- Q: Data pre-processing for BlendedMVS,
train_list.txtis missing?- See #33.
- Q: Loading checkpoint to fine-tune Fast3R?
- See #25
- Q: Running demo on Windows? (TypeError: cannot pickle '_thread.RLock' object)
- See #28. It seems that some more work is needed to make the demo compatible with Windows - we hope the community could contribute a PR!
- Q: Completely messed-up point cloud output?
- See #21. Please make sure the cuROPE module is NOT installed.
- Q: My GPU doesn't support FlashAttention /
No available kernel. Aborting execution?- See #17. Use
attn_implementation=pytorch_autooption instead.
- See #17. Use
- Q:
TypeError: Fast3R.__init__() missing 3 required positional arguments: 'encoder_args', 'decoder_args', and 'head_args'- See See #7. It is caused by a networking issue with downloading the model from Huggingface in some countries (e.g., China) - please pre-download the model checkpoint with a working networking configuration, and use a local path to load the model instead.
The code and models are licensed under the FAIR NC Research License.
See contributing and the code of conduct.
@InProceedings{Yang_2025_Fast3R,
title={Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Pass},
author={Jianing Yang and Alexander Sax and Kevin J. Liang and Mikael Henaff and Hao Tang and Ang Cao and Joyce Chai and Franziska Meier and Matt Feiszli},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month={June},
year={2025},
}
Fast3R is built upon a foundation of remarkable open-source projects. We deeply appreciate the contributions of these projects and their communities, whose efforts have significantly advanced the field and made this work possible.