scunet_color_real_gan changes levels

[NSQY]

Python 3.10.9
R61
GTX 1080
Driver Version: 525.60.11
CUDA Version: 12.0
Arch Linux

Might be PEBCAK.

https://slow.pics/c/38RmCJ0C

from vstools import vapoursynth, core, vs, depth
from vsscunet import scunet

def scunet_int(clip: vs.VideoNode, model=None) -> vs.VideoNode:
    _scunet = scunet(clip, model=model)
    return depth(_scunet, _scunet.format.bits_per_sample, sample_type=vs.SampleType.INTEGER)


rgb = core.imwri.Read(
    '/images data/Lenna.png'
    ).std.Crop(right=15)

ref = rgb.text.Text('ref')

rgb = depth(rgb, 16, sample_type=vs.SampleType.FLOAT)
format = 'RGBH' if rgb.format.bits_per_sample == 16 else 'RGBS'

ref.set_output()
images = [scunet_int(rgb, model=i).text.Text(f"model={i}, format={format}").set_output(i + 1) for i in range(5)]

[HolyWu]

Unfortunately that's a problem of the pretrained model itself. cszn/SCUNet#13

[AIisCool]

@NSQY If you don't mind my asking, how were you able to clean the image but retain the details? I've tried all the scunet models but while it cleans, it removes almost all the details.

[NSQY]

@NSQY If you don't mind my asking, how were you able to clean the image but retain the details? I've tried all the scunet models but while it cleans, it removes almost all the details.

You should be able to reproduce the comparison with my provided code. scunet seems to perform particularly well on Lena, however it looks worse on more complex samples such as Lichtenstein.png

If you're looking for image restoration functions I suggest trying https://github.com/HolyWu/vs-dpir