Quote:1. What does Stepped Resize do ? Does it upsclale using the above model again ?
Stepped resize is a sub-option for the selected resizer.
What is does is instead of directly resizing to the target resolution XXXX x YYYYY it applied the resizer multiple times / steps.
In the intermediate steps not the target but an intermedia resolution is used as target, for example:
Code:
# Step 1: current: 640x352, target: 1280x704
# Resizing using fmtconv - spline64
clip = core.fmtc.resample(clip=clip, kernel="spline64", w=1280, h=704, interlaced=False, interlacedd=False) # resolution 1280x704 before YUV420P8 after YUV420P16
# Step 2: current: 1280x704, target: 1920x1056
# Resizing using fmtconv - spline64
clip = core.fmtc.resample(clip=clip, kernel="spline64", w=1920, h=1056, interlaced=False, interlacedd=False) # resolution 1920x1056 before YUV420P16 after YUV420P16
Usually it does not make sense to use this on a machine learning based resizer, but for conventional resiziers.
Quote:2. Why does it say 48x48 to 852x480 in the corner when the input is 720x480 ?
48x48 is used while the input resolution is not known (no source loaded), should not happen when a source is loaded
The label indicates the resolution before and after the resizing.
If you for example additionally used "Filtering->Vapoursynth->Misc->Script->Lower res. before resize" it might indicate: "640x352 -> 320x176 -> 1920x1056" First is the base resolution for the resizing (after crop). Second is the resolution adjusted by 'Lower res. before resize' and the last is the target resolution for the resizing.
Quote:3. Is it possible to get the video to be rescaled first to 852x480 (16x9) then upscaled, because now its upscaling DVD (720x480) which would ruin the Aspect Ratio, since DVD starts at more of a square, what would be the best course of action here if i want a DVD times x4 upscaled, rescale to 852x480 then run Hybrid again ?
That makes no sense.
If your source PAR isn't set correctly, let Hybrid know the correct source PAR. (check the Vapoursynth Preview, to see whether you selected the right PAR)
It will then make sure that source par is kept during resizing.
If you want to change the output PAR i.e. to 1:1 let Hybrid know and it will make sure that you can only resize so that no distortion of the image takes place. (or is kept as low as possible, keeping the mod in mind)
Quote:4. Why cant i use Tensorrt Cores for RealCugan but i can for ESERGAN whats up with that. I would be really nice if it was present on CUGAN.
Hybrid will only offer options the used filter offers.
Quote:I understand VSGAN is present but it throws an error when selecting it, Picture below.
That is a bug, I didn't take stepped resizing into account when adding colorfix. => will look at it after work
Quote:5. Is VGAN not using Tensorrt only vsmlrt ?
Hybrid will only and can only offer options the used filter offers.
Quote:, No duplicate aware in VSGAN, assume since its not vsmlrt VSGAN like i used before ?
The duplicate aware option is something I implemented, due to a user requires for VSGAN, which uses
https://github.com/Selur/VapoursynthScri...esizing.py. It probably could be extended for VSLMRT, but I haven't written code for it or tested it so far.
Quote:5. I would want to apply RealCugan x2 Model twice, which i know is available in VSGAN (Maybe i want VSMLRT instead?) not possible in RealCugan Section?
You could use stepped resizing, or use the model (assuming it is in a compatible format) in VSGAN or VSMLRT multiple times.
(Usually applying a machine learning multiple times makes no sense.)
Quote:6. What is Sync Gap in REALCugan and would that reduce quality if switched to VSGAN/vsmlrt ? (I just tried VSMRT with cugan x2 onnix slim fp16 clamp and results seem better)
see:
https://github.com/Kiyamou/VapourSynth-R...cnn-Vulkan
Yes, that option is not available when using CUDAN through VSMLRT or VSGAN and thus can influence the output quality.
Quote:7. What is the order of the Multi Upscale since i want denoise x3 first then no denoise mode to go through, i assume it goes in order of selection when browsing for it ?
Yes, it should use the order the models are selected.
Quote:8. You cant load Custom Engine Models ? Because i am not sure what the log is cmd is blank, but from what i learned if u use the same min opt max Shapes as the video upon engine creation it would speed up the upscale process. How would this program know the shapes it needs to make when i give it two x2 models to upscale, unless it makes high res shape engine which would worsen the FPS Encoding, there is a case also with RealCugan x3 where it needs to be the exact shape of the video (same min opt max shapes) otherwise the engine wont create.
You can see the code Hybrid creates in the "Vapoursynth Script View". Hybrid does not change shapes parameters or other stuff.
Quote:9. I guess above problem is solved by resizing the video back to orignal state, but than wouldnt i be losing tons of detail by cramping 2 UPSCLAES in a 480p video lmao.
If you apply multiple resizers, Hybrid will simply apply one after the other and add the end make sure that the 'resize output resolution' meats the target resolution.
So if you apply three x2 resizers on a 640x320 source while aiming for 1920x1080, the first would resize to 1280x640, the second to 2560x1280 and the third would resize to 5120x2560 and then at the end Hybrid would resize that to 1920x1080.
Quote:10. Why are there 2 VSGAN in Frame and Others ?
In the Others-tab each time a model is applied, the output is resized to the input resolution before applying the model.
It is mainly meant for 1x_... models.
Quote:and which model is Cugan Pro Denoise LIght ? x1 Denoise x2 Denoise or x3 Denoise ?
No clue. Don't think Hybrid comes with those models.
Cu Selur
Ps.: Like I wrote, I'll look into color fix and stepped resizing after work.