RESULTS.txt

ALGORITHM - PSNR(LUMA)/RUNTIME(MS)/PARAMETERS
BILINEAR - 23.03/0/0
BICUBIC - 23.70/0.3/0
NGU-SOFT-VERYHIGH - 23.86/37.7/?
LANCZOS - 23.88/1.1/0
NNEDI3-256 - 24.07/62.4/49.6K
RAVU-R4 - 24.09/3.6/41.4K
FSRCNNX-8 - 24.32^/6.1/2.9K
FSRCNNX-8-LINEART - 24.46^/6.1/2.9K
NGU-SHARP-LOW - 24.46/3.9/?
NGU-SHARP-MEDIUM - 24.51/5.3/?
NGU-STANDARD-VERYHIGH - 24.51/37.7/?
ACNet - 24.55*/7.9/4.7K
FSRCNNX-16 - 24.57^/30.4/10.5K
FSRCNNX-56 - 24.60^/52.4**/15.2K
NGU-SHARP-HIGH - 24.69/11.0/?
ANIME4K-M - 24.73/1.5/1.6K
NGU-SHARP-VERYHIGH - 24.80/37.7/?
ANIME4K-L - 24.94/2.5/2.9K
ANIME4K-UL - 25.14/10.7/15.9K
WAIFU2X-CUNET - 25.61^/>1000/1283.3K

^ hese algorithms have not been trained on the area averaging downscaling degradation[1], this has a negative effect on PSNR.
* ACNet performs joint super resolution with denoising. This has a larger negative impact on PSNR compared to perceptual quality.
** FSRCNNX-56 runtime is estimated based on the number of multiply-add operations compared to FSRCNNX-16, as it fails to run on 4K upscaling tasks.

[1] The area averaging downscaling operation is the correct downscaling operation for all natural images and line art images. Bicubic interpolation and Lanczos introduces gradient overshoot when downscaling if used without a lowpass filter. Lowpass filtering adds blur to the image, which degrades fine detail present in line art.