Search options (Page 12 of 18)

The new technique used for the "lite" versions is potentially a little more accurate than the normal versions but I suspect that GPUs with low internal memory bandwidth will have issues with v14 "lite". I'm keeping it for now and will see what happens further down the line.

Edit:
The figures from @aloola and @RickyAstle98 show that this is a heavy model. But now everything has kind of bedded in, Rife v1.14 "lite" (v2) is using a little less GPU resources than before, down from 94-100% with Rife v1.14 to between 86-92% with "lite" while "so far" retaining the same quality as before. It's also very smooth. YMMV on GPUs with limited memory bandwidth.

I've updated and I don't see any obvious differences. But I have thought about updating for the latest fixes to TensorRT and Cuda so thanks

Yep. That's where I got the last copy from too. It's configured for Catmull Rom. The ones I saw configured for Mitchell must have been local variations by other users.

Thanks. If you see a particular issue that you want fixing then this does make sense. Especially if you are saying that v14 fixes this issue for you. I still do testing of new models, changing code etc but I think SVP is stable and works well now so I would not recommend most people do this. Why fix something when it ain't broken ?

I use a number of videos highlighting different artefacts. But Alita might be the best because it has different types of fast movement and exposes a couple of minor artefacts. One of which has been fixed by Rife 4.14.

It's a change to the code in base.py. I got this from a previous post by @blackmickey1007 for completely disabling SC.

SVP4\script\base.py

input_m = input_m.misc.SCDetect(threshold= 0.1 or 0.2)

The threshold values are just suggestions. You will need to experiment to see what works for you. You won't find a pefect setting because when you fix double images from fast movement it will break slow panning. When you reset for slow panning you get double images in fast movement. But you can get a compromise which improves fast movement a little without breaking slow panning. I couldn't get even this minor improvement in 4.12/4.13 but 4.14 does seem to be better overall for fast movement so here it was possible.

I'm not going to comment on config for OLED. It's best you experiment for yourself. Here are some points to think about.
dither-depth=auto and spirv-compiler=auto is the default so not necessary
fbo-format is disabled in gpu-next so is not needed and won't work anyway.

KrigBilateral isn't really necessary and although I've never seen it myself it can apparently cause mistakes.
Is FSRCNNX_x2_16-0-4-1.glsl that much better than FSRCNNX_x2_8-0-4-1.glsl considering the extra processing power needed?
Do you need to use both FSRCNNX_x2_16-0-4-1.glsl and SSimDownscaler.glsl? That's a lot of processing and sharpening.
SSimDownscaler.glsl is configured by default for Mitchell but you are using ewa_lanczossharp. So that is not optimal.

You are sharpening in upscale, shader and downscale. Although FSRCNNX might not be doing anything because by default it only works when a x2 or greater upscale is needed. This might be why you can get away with running FSRCNNX_x2_16 and SSimDownscaler which are very GPU heavy.
Some shaders overwrite others so it's best to use append when you want multiple shaders to work at the same time.

Hope this helps a little.

Now that you have said it is obvious. I already knew what multi does in theory. But what I didn't realise was that in practice it directly relates to the fps which again is now obvious. Thanks

Thanks but that's not what I'm trying to do. Or at least I don't think that's what I'm trying to do. Does changing the code in the way I mentioned do this?

I'm using my GPU heavily and can easily hit 100% usage with 4.12.4.13 but max out at 80% with 4.9. So maybe you don't see it under lighter loads?

Make sure you have a 4.13 onnx file in the "v2" folder. If you do then delete it and any files created by it, download the v2 onnx file again and put the new onnx file into the "v2" folder.

This is not necessary. In the SVP rife folder, just replace the onnx file for the model you want updated then delete the generated model and engine files. This will trigger regeneration. But if you are using model names that are different to the actual model being used then I don't know since I stopped doing this some time ago.