LeafSR

Demo

Overview

LeafSR is a PyTorch super-resolution pipeline for plant leaf images. It enlarges low-resolution leaf photos 4x by predicting the fine detail that bicubic interpolation misses, and it turns a Kaggle-style notebook into a reproducible project with training, inference and comparison scripts.

The model is an ESRGAN-lite conditional GAN. A generator built from Residual-in-Residual Dense Blocks predicts a residual on top of the bicubic image, and it is trained with a mix of pixel, perceptual, frequency and adversarial losses.

On held-out images it lowers the mean absolute error from 18.20 (bicubic) to 16.89, a 7.2% reduction. The demo below is a recording of the project's own live site, which shows three real paired samples and predictions from the trained checkpoint.

What it does

• ESRGAN-lite generator. Residual-in-Residual Dense Blocks with pixel-shuffle upsampling, predicting a residual over the bicubic image.
• Conditional discriminator. It sees the bicubic image together with the real or generated high-resolution image.
• Hybrid objective. Charbonnier, residual, VGG perceptual, FFT and adversarial losses combined.
• Careful training. CUDA mixed precision, a cosine schedule, gradient clipping, EMA-smoothed weights and early stopping.
• Inference. Eight-way test-time augmentation and an optional top-k checkpoint ensemble, writing PNGs or a Kaggle-style submission CSV.
• Reproducible. YAML config, a Docker image, smoke tests and CI, plus scripts to prepare data and build comparison grids.

Screenshots

How it works

1. Low- and high-resolution leaf images are matched by filename and cropped into aligned patches with random flips and rotations.
2. Each patch is upsampled 4x with bicubic interpolation, and the RRDB generator predicts a residual that is added on top.
3. A hybrid loss keeps the output faithful, and a low-weight adversarial term from a conditional discriminator sharpens texture.
4. Mixed-precision training with a cosine schedule keeps EMA-smoothed checkpoints and stops early when validation stops improving.
5. At inference, eight flipped and rotated variants are averaged and written out as images or a submission file.

Install and try it

LeafSR needs Python 3 and PyTorch. The competition dataset is not included in the repository, so a full run needs the original data in data/.

1. Set it up

   git clone https://github.com/BriskAM/leafsr.git
   cd leafsr
   python3 -m venv .venv
   source .venv/bin/activate
   pip install -r requirements.txt

   Copy
2. Prepare the data

   python scripts/prepare_data.py plant-leaves-super-resolution-challenge.zip --out-dir data

   Copy
3. Train

   python scripts/train.py --config configs/default.yaml

   Copy
4. Enhance images

   Use scripts/make_comparison.py to build low-res / bicubic / prediction / target grids like the one above.

   python scripts/infer.py \
     --checkpoint runs/leafsr/best_leafsr.pt \
     --input-dir data/test_Low_Resolution \
     --output-dir runs/leafsr/predictions

   Copy

Notes

• The dataset, VGG weights and checkpoints are kept out of git, so the demo uses three committed paired samples and predictions precomputed from the trained checkpoint.
• Validation error is the mean absolute error on held-out training images. The original run used a Tesla T4 with mixed precision.
• The project site (linked below) is a live, interactive version of this demo.