How can I make it work with Windows?

[moreAImore]

I see there is a "inference_sr.sh" inside (video_super_resolution)
I think this is for linux systems right?
How can I make it work on windows?

2. the models are to be added to "STAR\pretrained_weight" and this folder to be created manually right?

[moreAImore]

I used chatgpt and made it

[tin2tin]

@moreAImore Please share.

[moreAImore]

@moreAImore Please share.

I will try again, because it failed in the end!
It said error TRITON (That thing usually work on linux best, and is difficult to make it work on windows) I will need to try again

[tin2tin]

On Windows, missing Triton isn't a show-stopping error.

[LucienJacotey]

I have Ubuntu wsl on windows 11 and the script just ends, the video is not processed

[nitinmukesh]

Anyone looking for Triton wheels for Windows, I uploaded it here.
https://huggingface.co/datasets/newgenai79/Windows_wheels/tree/main/TRITON

I installed the 3.0.0 cp310 and no more triton missing warning. Can't guarantee it will work for everyone, but these always worked for me.

[Kwisss]

I think you just need to select the right versions from the install matrix
because otherwise you will encounter issue after issue. I will try it tomorrow and will report back.

[moreAImore]

Anyone looking for Triton wheels for Windows, I uploaded it here. https://huggingface.co/datasets/newgenai79/Windows_wheels/tree/main/TRITON

I installed the 3.0.0 cp310 and no more triton missing warning. Can't guarantee it will work for everyone, but these always worked for me.

Can you share your full install, do a python.exe pip list ot python.exe -m pip list? (or whatever method works but make sure it is the python used in your install , not the whole computer, unless you did not use an env/conda)

[pointave]

Couldnt get it working with windows, when running command it wouldnt find the environment it was in and it gave the unix path instead also saying its missing torch. Working on wsl2 but it uses a crazy amount of memory. 4090 full and using 60gb of ram when doing the 4x upscale with the sample 240p video. 15 gib less sys ram when doing 2x upscale and each step takes like 10 minutes+. I dont think this is meant to run on consumer gpus, maybe only the 5090. edit: saw post on vram requirements and it uses 40gb, it works if you drop from 32 to 12 for framelength. Still takes a minute per step for the sample vid.

[moreAImore]

Couldnt get it working with windows, when running command it wouldnt find the environment it was in and it gave the unix path instead also saying its missing torch. Working on wsl2 but it uses a crazy amount of memory. 4090 full and using 60gb of ram when doing the 4x upscale with the sample 240p video. 15 gib less sys ram when doing 2x upscale and each step takes like 10 minutes+. I dont think this is meant to run on consumer gpus, maybe only the 5090. edit: saw post on vram requirements and it uses 40gb, it works if you drop from 32 to 12 for framelength. Still takes a minute per step for the sample vid.

can you make a step to step working method with wsl?

[pointave]

Do you have wsl and miniconda/anaconda already? I just followed the directions and all I had to change was model.pt to light_deg.pt and what I mentioned earlier, lowering of the framelength (I tried all the way to 8). Otherwise the instructions on the page are exactly what you need to do. I did install cudatoolkit but I dont think that matters.

After runing the bash command, it will recognize the model you downloaded it will then download 4 gigs of other files. I tried more vids and they were mediocre and a few looked terrible. They had to remain small or else it would fill up the vram. Like really small, .4 megapixels 8 second clips were too much for it. If you don't have 40gb of vram, I would pass on this one.

[vtrixlaz]

Almost got it running on Windows without the WSL method buy have hit the following error:
import open_clip
ModuleNotFoundError: No module named 'open_clip'
All videos processed successfully.

Event though the correct torch is installed. Tried a bunch to debug to no avail.