README.md

---
library_name: transformers
tags:
  - image-restoration
  - lensless-imaging
license: mit
---

# Lensless Computational Imaging

10 models for reconstructing images from a lensless camera.
Trained on [DigiCam-Mirflickr-MultiMask-10K](https://huggingface.co/datasets/bezzam/DigiCam-Mirflickr-MultiMask-10K).
## Models

**Classical**

| Subfolder | Description |
|---|---|
| `admm100` | ADMM, 100 iterations, fixed hyperparameters. |
| `fista100` | FISTA, 100 iterations, fixed hyperparameters. |

**Unrolled**

| Subfolder | Description |
|---|---|
| `le_admm20` | ADMM, 20 unrolled steps with learnable per-step hyperparameters. |
| `le_fista20` | FISTA, 20 unrolled steps with learnable per-step hyperparameters. |

**Modular**

| Subfolder | Description |
|---|---|
| `modular_le_admm5_prepost` | Pre-UNet → 5 ADMM steps → post-UNet. |
| `modular_le_admm5_pre` | Pre-UNet → 5 ADMM steps. |
| `modular_le_admm5_post` | 5 ADMM steps → post-UNet. |
| `modular_le_fista5_prepost` | Pre-UNet → 5 FISTA steps → post-UNet. |

**GAN super-resolution**

| Subfolder | Description |
|---|---|
| `admm100_bsrgan` | ADMM-100 + pretrained BSRGAN x4. |
| `admm100_bsrgan_finetune` | ADMM-100 + BSRGAN x4 fine-tuned on DigiCam. |

## Usage

```python
from transformers import AutoModel

model = AutoModel.from_pretrained(
    "timofeiiz/lensless-computational-imaging",
    subfolder="modular_le_admm5_prepost",
    trust_remote_code=True,
)
recon = model(lensless, psf)
```