Skip to content

4xHFA2k

Compare
Choose a tag to compare
@Phhofm Phhofm released this 25 Apr 07:31
· 39 commits to main since this release
c664eb5

Name: 4xHFA2k
Author: Philip Hofmann
Release Date: 07.05.2023
License: CC BY 4.0
Network: ESRGAN (RRDBNet)
Scale: 4
Purpose: An anime images 4x upscaling model based on musl's HFA2k dataset
Iterations: 91,000
batch_size: 9
HR_size: 256
Epoch: 318 (require iter number per epoch: 286)
Dataset: HFA2k
Number of train images: 2568
OTF Training: Yes
Pretrained_Model_G: RealESRGAN_x4plus.pth
Training time: 36h on a P100 GPU

Description: 4x anime image upscaler with a bit of otf jpg compression and blur. Trained on musl's hfa2k dataset release for Anime SISR, which has been extracted from modern anime films, where the selection criteria was high SNR, no DOF and high frequency information.

Examples: https://imgsli.com/MTc2NDgx

Example1
Example2
Example3
Example4
Example5