dejpeg

JPEG artifact removal that runs entirely in your browser. Photos never leave the device: inference runs client-side via ONNX Runtime Web (WebGPU primary, WASM fallback), backed by a compact custom-trained model.

Try it: https://dejpeg.ludwina.top

Models

Model	Size	Best for
DeJPEGNet v2 (default)	15.4 MB	Staying close to the original: beats FBCNN on PSNR, LPIPS and DISTS at every tested quality
DeJPEGNet v2 + Texture	15.3 MB	v2 with a slider for added grain, which is synthetic and not recovered detail
DeJPEGNet v1	5.4 MB	Smaller download, tuned for perceptual quality

LIVE1, PSNR (dB) / LPIPS / DISTS, one pass, quality unknown to the model:

	QF10	QF20	QF30	QF40
FBCNN	27.77 / 0.250 / 0.179	30.11 / 0.158 / 0.132	31.43 / 0.119 / 0.108	32.34 / 0.097 / 0.094
v2	28.05 / 0.239 / 0.174	30.37 / 0.152 / 0.129	31.70 / 0.113 / 0.106	32.61 / 0.092 / 0.091
v2 + Texture at 0.5	28.04 / 0.228 / 0.158	30.33 / 0.142 / 0.112	31.65 / 0.104 / 0.089	32.54 / 0.083 / 0.075
v2 + Texture at 1.0	27.97 / 0.210 / 0.137	30.23 / 0.125 / 0.091	31.50 / 0.088 / 0.069	32.35 / 0.068 / 0.057

v2 takes about 55 ms per 512x512 tile in Edge on WebGPU on one NVIDIA GPU. Weights are in the releases.

What's here

Path	What
apps/web/	Astro + Svelte 5 web app (the tool itself)
packages/inference-core/	Framework-free TypeScript inference library: worker, codecs, adaptive tiling
packages/fbcnn-py/	Modern PyTorch port of FBCNN
packages/dejpeg-train/	Research training pipeline: WebDataset corpora, synthetic degradation, dedup manifest
tools/convert-fbcnn/	FBCNN PyTorch → ONNX FP16 conversion tool
models/dejpeg/	DeJPEGNet v1: training, inference, ONNX export (weights via GitHub Releases). The v2 training code is not in this repository

Quickstart

Web app (bun):

bun install
bun run dev          # http://localhost:4321

Python packages (uv):

uv sync                       # fbcnn-py + tools (workspace)
pip install -e models/dejpeg  # shipped model package

Model inference:

from dejpeg import load_model, restore_image

model = load_model()                  # v1 release weights (auto-downloaded to ~/.cache/dejpeg)
restore_image("input.jpg", "out.png", model)

Training from your own images:

python models/dejpeg/scripts/train.py --data ~/datasets/clean-images --out runs/v1

The full research pipeline (packages/dejpeg-train/) builds WebDataset shard corpora from DIV2K/Flickr2K-class sources and synthesizes JPEG degradations on the fly; see its README.

Development

bun test               # inference-core unit tests
uv run pytest packages/fbcnn-py
pytest models/dejpeg   # model architecture/inference tests

License

Apache-2.0. FBCNN-derived code (packages/fbcnn-py/) originates from an Apache-2.0 upstream. Benchmark images (Classic5, LIVE1) are not distributed.