daminovkamil/lensless-reconstruction

Lensless Reconstruction

PyTorch checkpoints for lensless image reconstruction on bezzam/DigiCam-Mirflickr-MultiMask-10K. The models reconstruct RGB scenes from coded-mask sensor measurements and PSFs.

Code and demo: daminovkamil/lensless-reconstruction

Checkpoints

File	Model	Test PSNR	Notes
model_best.pth	pre+U5+post	16.47 dB	Best model, 4M DRUNet pre + learned ADMM-5 + 4M DRUNet post
pre4_u5_post4.pth	pre+U5+post	16.47 dB	Same architecture as the best model
u5_post8.pth	U5+post	15.30 dB	Learned ADMM-5 + 8M DRUNet post
fista_prepost5.pth	FISTA pre+post	15.41 dB	Bonus model, 4M DRUNet pre + learned FISTA-5 + 4M DRUNet post
admm.pth	Le-ADMM-20	12.11 dB	Unrolled ADMM with learned per-iteration parameters
admm100.pth	ADMM-100	6.69 dB	Classical fixed-parameter ADMM baseline

Each .pth checkpoint stores its own Hydra config, so no separate config file is needed for inference. sample.zip contains a small custom-layout sample for the Colab/demo flow.

Metrics

All metrics are computed on the ROI of the DigiCam test split.

Model	PSNR ↑	SSIM ↑	MSE ↓	LPIPS ↓
pre+U5+post	16.47	0.463	0.0233	0.537
FISTA pre+post	15.41	0.387	0.0292	0.593
U5+post	15.30	0.385	0.0304	0.581
Le-ADMM-20	12.11	0.366	0.0629	0.777
ADMM-100	6.69	0.292	0.2155	0.807

Usage

Clone the project repo and run inference with the hosted weights:

git clone https://github.com/daminovkamil/lensless-reconstruction
cd lensless-reconstruction
pip install -r requirements.txt

python inference.py \
  inferencer.from_pretrained=daminovkamil/lensless-reconstruction \
  inferencer.pretrained_filename=model_best.pth \
  datasets.test.data_dir=demo_sample

Use the FISTA checkpoint by changing only the filename:

python inference.py \
  inferencer.from_pretrained=daminovkamil/lensless-reconstruction \
  inferencer.pretrained_filename=fista_prepost5.pth \
  datasets.test.data_dir=demo_sample

Reconstructions are written to data/saved/reconstructions/<split>/<ImageID>.png.

Citation

This homework implementation follows the ADMM formulation from Monakhova et al. 2019 and the modular learned reconstruction setup from Bezzam et al. 2025.