Fast JPEG Illust v1

Scale: 1
Architecture: SPAN

Author: Rundev
License: CC-BY-4.0
Purpose: Compression Removal, JPEG
Subject: Anime, Cartoon, CGI, Manga
Input Type: Images
Date: 25-04-2024

Size: 48nf (default)
I/O Channels: 3(RGB)->3(RGB)

Dataset: Custom dataset of Danbooru images. I downloaded 38k images (around 70GB, ID 6000000 to 6039999), downscaled JPEG images 50% or 25% depending on the severity of the JPEG artifacts in the image, discarded images that were too small (leaving 35k images at 54GB), and split large images into tiles (creating a total of 110k images). I then sampled 10% of those images (with a bias for images with high-frequency details) to create the dataset of GT images.
Dataset Size: 10988
OTF (on the fly augmentations): No
Pretrained Model: none
Iterations: 184000
Batch Size: 16
GT Size: 64

Description: A SPAN models for blind JPEG artifact removal (quality ≥50%, any chroma subsampling).

LR images were generated using OpenCV and use random quality (between 50% and 97%) and chroma subsampling settings. So the model can be used for images with light and heavy artifacts.

Despite the large range, the model can clean up JPEG artifacts without introducing blur, color shift, or other artifacts. However, it consistently fails to restore high-frequency color detail lost during compression. It also fails to clean up some artifacts near regions with lots of high-frequency detail in heavily compressed images (quality <75%).

Showcase: https://imgsli.com/MjU5MDc2/0/1