primo-stitch 0.1.0

PRIMO: Panoramic Reconstruction with Integrated Microscopy-Specific Optimization

PRIMO: Panoramic Reconstruction with Integrated Microscopy-Specific Optimization

Python library and command-line tool for stitching a set of overlapping tiles into a single 2D panorama.

Installation

pip install primo-stitch

• Requires Python 3.10–3.12.
• PRIMO depends on PyTorch. The default install pulls the CPU build; for GPU use, install the CUDA build of torch/torchvision first (see https://pytorch.org), then install PRIMO.
• Matcher weights are downloaded automatically on first use (Hugging Face Hub / Torch Hub) and cached locally.

After installation you get the primo-stitch command and the importable primo package (install name primo-stitch, import name primo).

Command-line usage

primo-stitch is the main entry point.

# minimal — stitch a folder of tiles into a panorama
primo-stitch --tile_dir path/to/tiles --output_file panorama.png

# advanced
primo-stitch \
  --tile_dir path/to/tiles \
  --output_file panorama.png \
  --matcher "efficient loftr" \
  --device cuda:0 \
  --blending_mode full \
  --inference_size 0.5 \
  --batch_size 8 \
  --cache_dir cache/ \
  --logfile run.log

By default (full blending with --save_alpha_channel on), the panorama is written as .png (RGBA) regardless of the requested extension. Pass --no-save_alpha_channel to save a .jpg.

Options

Flag	Default	Description
--tile_dir	(required)	Directory with the input tiles
--output_file	panorama.jpg	Output panorama path (extension may be adjusted, e.g. .png when alpha is saved)
--cache_dir	cache/	Cache directory (created automatically)
--matcher	xfeat	xfeat | efficient loftr | loftr
--device	cpu	cpu, cuda, cuda:0, ...
--blending_mode	full	collage | mosaic | full
--inference_size	0.3	Matcher input scale relative to the original (0.25, 0.5, 1, ...)
--batch_size	1	Matcher batch size (higher = faster, more memory)
--save_alpha_channel / --no-save_alpha_channel	on	Save the transparency channel; forces .png output in full mode
--logfile	(none)	Write a debug log to this file

Python API

from primo import Matcher, Stitcher

matcher = Matcher(
    model='xfeat',            # 'xfeat' | 'efficient loftr' | 'loftr'
    device='cuda:0',          # or 'cpu'
    inference_size=0.5,
    batch_size=8,
)

# the alignment device is taken from the matcher
stitcher = Stitcher(
    matcher,
    blending_mode='full',     # 'collage' | 'mosaic' | 'full'
    save_alpha_channel=True,
)

stitcher.stitch(
    input_dir='path/to/tiles',
    output_file='panorama.jpg',
    cache_dir='cache',
)

Matcher and Stitcher expose additional keyword arguments (alignment, photometric correction, blending) — see their signatures for the full set.

Demo

The interactive demo (live preview + machine-readable progress) is driven by primo-stitch-online and the run_online.sh wrapper. For normal use they are not needed; primo-stitch can also emit progress by adding --online, which writes a status.json and a preview.jpg while processing.

Authors

• Gleb Nikolaev — Lomonosov Moscow State University
• Savelii Shashkov — Lomonosov Moscow State University
• Dmitriy Korshunov — Geological Institute of the Russian Academy of Sciences
• Andrey Krylov — Lomonosov Moscow State University
• Alexander Khvostikov (corresponding author) — Lomonosov Moscow State University

Citation

If you use PRIMO in your research, please cite (update pages/DOI with the final proceedings entry):

@inproceedings{primo2026,
  title     = {PRIMO: Panoramic Reconstruction with Integrated Microscopy-Specific Optimization},
  author    = {Nikolaev, Gleb and Shashkov, Savelii and Korshunov, Dmitriy and Krylov, Andrey and Khvostikov, Alexander},
  booktitle = {Proceedings of the 34th ACM International Conference on Multimedia (MM '26)},
  year      = {2026},
}

License

PRIMO's source code is licensed under the Apache License 2.0. It bundles third-party components under their own licenses (Apache-2.0, MIT); see THIRD_PARTY_LICENSES.md.

The accompanying paper is published by ACM under CC BY 4.0 (chosen via the ACM eRights process) — this is separate from the software license above.