SR Flaw Scry: Artifact Detection for Super-Resolution

🤗 Demo | 📦 Model | 🗂️ Dataset

🏅 Overview

Find where super-resolution goes wrong and which artifacts appear

🤗 Demo

Try our hosted demo on Hugging Face. The demo compares RealESRGAN and SwinIR on examples with texture hallucinations, stripe-like pattern artifacts, and blur-like surface defects.

🖥️ Run the UI

To run the project locally, install the package first:

pip3 install git+https://github.com/msu-video-group/srflawscry

Then launch the interface:

python3 -m srflawscryui

It will automatically download our demo dataset and model. To use your own images, put them into assets/data folder with this structure:

assets/data/
  MyMethod/
    sources/<item_id>
    results/<item_id>

where the sources folder contains the original low-resolution inputs and the results folder contains the 4x-upscaled images with matching file names. Metric computation is supported on CPUs, Apple Silicon Macs, and CUDA-capable GPUs. At least 3 GB of VRAM is recommended for GPU inference.

🚀 Using WASD in Your Project

Our artifact detection metric is called WASD. You can easily add it to your project with the following snippet:

from pathlib import Path

from srflawscry import WASD

model = WASD.from_pretrained("egorchistov/sr-artifact-detection-wasd").eval()
model.run(
    image_path=Path("result.png"),
    source_path=Path("source.png"),
    output_path=Path("artifact_mask.png"),
)

🛠️ Extending This Repository

For local development, you need the full installation:

git clone https://github.com/msu-video-group/srflawscry.git
cd srflawscry
pip3 install --editable .[dev]
pre-commit install

You can extend this repository in many ways. For example, if you want to add a simple scalar metric such as LPIPS or MANIQA, create a metric class that returns one number per image and add it to default_config(). If the metric also produces a heatmap or mask, save it in outputs/ so the UI can display it beside the compared images.

FolderWithSourcesLayout
        │
        ▼
Sample(item_id, result_path, source_path)
        │
        ▼
Metric.evaluate(sample) ──► numbers in metrics.csv
        │
        └───────────────► optional masks in outputs/

❓ Need Help?

Feel free to open an issue if you have any questions.