Accepted by ICCV 2025

Jiakai Zhang, Shouchen Zhou, Haizhao Dai, Xinhang Liu, Peihao Wang, Zhiwen Fan, Yuan Pei, Jingyi Yu

1 ShanghaiTech University, 2 Cellverse Co., Ltd., 3 HKUST, 4 UT Austin

1. Clone repo

   cd CryoFastAR

2. create conda envionment and activate

conda create -n cryofastar python=3.11
conda activate cryofastar
pip install -r requirement.txt

Due to the limited experimental data in public, our model only supports four types of experimental data from cryoDRGN's datasets. We split them into the training data (30,000 for each) and unseen remained test data. To play with them, please first download the cryofastar-ckpts.zip, and unzip it to checkpoints/, then we provide one example for real data and one example for synthetic data to reproduce our results in our paper, please download data.zip, and unzip it to data/

Simply running below commands can generate results under results/ including the reconstructed volume as well as the angluar histograms.

bash evaluate_model_real.sh
bash evaluate_model_synthetic.sh

To inspect how the 3D volume grows while the model consumes new views, use the helper below:

python visualize_reconstruction_progress.py \
    --ckpt-path checkpoints/cryofastar.pth \
    --root-path data/your_dataset \
    --num-views 64 \
    --out-dir results/progress_example

This script runs a single evaluation pass but saves intermediate diagnostics in results/progress_example/:

• renders/render_XXXX.png: top-row input thumbnails, volume slices, a max-intensity projection, the pose scatter, an orientation heat map, and the 2D shift error map (disable via --render-mode none).
• progress.csv: timeline metadata (time, processed images vs. total, sampled ids, etc.).
• pose_reference_dirs.npy / pose_inference_dirs.npy: accumulated pose orientations for reference and inference views; optional pose_<metric>_values.npy matches the chosen pose heat-map metric.
• translation_positions.npy / translation_errors.npy: shift statistics gathered during the run.
• progress.mp4: automatically generated 16:9 video (10 fps by default) that strings the renders together; skip via --no-video or adjust speed with --video-fps.

Use --snapshot-interval to control how often a snapshot is written, pass --max-snapshots if you want to cap how many are kept (the default keeps all), and switch to --pose-source gt to visualise the ground-truth pose distribution instead of predictions. --pose-heatmap-metric now supports loss3d, rotation angle (rot), F-norm rotation error (rot_fro), 2D shift error (shift), or plain density; combine with --heatmap-vmin/--heatmap-vmax to clamp the color bar. Turn on --snapshot-regularize if you prefer the slower but sharper Fourier-domain filtering for each saved volume.

@inproceedings{zhang2025cryofastar,
  author    = {Zhang,Jiakai and Zhou, Shouchen and Dai, Haizhao and Liu, Xinhang and Wang, Peihao and Fan, Zhiwen and Pei, Yuan and Yu, Jingyi},
  title     = {CryoFastAR: Fast Cryo-EM Ab Initio Reconstruction Made Easy},
  booktitle = {Proc. ICCV},
  year      = {2025},
}