-
Notifications
You must be signed in to change notification settings - Fork 8
DLISR
AkarinVS edited this page Aug 27, 2021
·
8 revisions
akarin.DLISR(clip clip, [, int scale=2])
DLISR (Deep Learning Image Super Resolution) is part of NVidia NGX Technology SDK.
It's a high quality image upscaler (2x/3x/4x), but it's very slow, even on a RTX GPU.
Suitable for natural images, CGI and anime style line drawings. For anime content, compared to the well-known waifu2x, it can preserve the detail textures better, but it does not have anti-alias effects, so the lines might appear jagged after upscaling. (Hint: merge waifu2x output with DLISR output with a line mask.)
DLISR does not tolerate noisy inputs, so it's only suitable for upscaling clean inputs. (Hint: if you use bm3d, use bm3dcpu, not bm3dcuda.)
It's very slow. e.g. when 2x scaling, we have:
- On RTX 3070 Ti: 1.2fps for 1080p inputs; 2.66fps for 720p; 7.7fps for 480p.
- On Tesla T4 (~RTX 2070): 0.34 fps for 1080p inputs; 0.75 fps for 720p; 2.20 fps for 480p.
- On GTX 1660S: 0.07 fps for 1080p inputs; 0.16 fps for 720p; 0.35fps for 720x480.
- DLISR does not work well with other GPU filters, if you use DLISR in your script, make sure it's the only GPU filter. (tracked as #4). Also, as DLISR can easily saturate the entire GPU, if you only have one GPU, using other GPU filters will only slow things down.
- DLISR always pick the best GPU to use, so it's impossible to specify a specific GPU device to use, even though there is
device_id
parameter, it does NOT work.