Better alignment = cleaner tomograms. GIF shows a denoised reconstruction after patch tracking alignment and miss-alignment.
Installation is limited at the moment to a specific python, CUDA, and torch version. This might be fixed at some point in the future. For now, its easiest to set everything up in a conda environment.
First create an environment called miss-alignment with cuda-toolkit 12.9 and activate it:
conda create -n miss-alignment -c conda-forge python=3.12 cuda-toolkit=12.9 -y
conda activate miss-alignment
We need to fix some GPU dependencies for accelerated reconstruction:
python -m pip install torch==2.8.0
python -m pip install torch-projectors --index-url https://warpem.github.io/torch-projectors/cu129/simple/
Important
If your GPU's have the Blackwell-architecture make sure to install at least v0.11 of torch-projectors.
Finally install miss-alignment with this command:
python -m pip install miss-alignment
Check that the CLI shows up with:
miss-alignment --help
This lists the available commands (train and infer); run
miss-alignment train --help or miss-alignment infer --help for details on
each.
See docs/usage.md for instructions.
A full list of changes per release is available on the GitHub Releases page.
Chaillet, M.L., van Loenhout, J., Leung, M.R., Burt, A., and Tegunov, D. (2026) MissAlignment Teaches Itself Better Cryo-ET Tilt-Series Alignment by Making It Worse. bioRxiv. https://doi.org/10.64898/2026.04.29.721716