Huge Disappointment!

[Icon67]

[naxci1]

My GPU actually has 16 GB of VRAM and works well, it's just that the purpose matters.

SeedVR2 is unrivaled in restoring old and corrupted videos, meaning you can convert a 240p or 360p video to 720p or 1080p. However, your situation is different; your video is already 1080p, meaning it's new and in good condition, so there's no need to restore it. Only enhancements might be needed, which you'd be better off doing with other applications.

[Icon67]

I didn't say it was in good condition it has artifacts that need removing and I wanted to upscale it which is the purpose of the app.

[marcogva-hub]

As @naxci1 said, to remove artifacts, improve 1080p videos, there are better choices; @naxci1 was talking about 240p videos, I went even further, and obtained amazing results, even with 144p results. The only issue is the inference time, mainly taken by the decode phase... I wish I was able to find a way to speed it up, been at it for a couple months without particular success.

[urixl]

I didn't say it was in good condition it has artifacts that need removing and I wanted to upscale it which is the purpose of the app.

This model is not supposed to upscale 1080p or even 720p videos with a consumer-grade cards.
It was trained at 96Gb cards and it's a miracle that it works at 16Gb cards at all.

My workflow for video upscale is to downscale input to 484x272, then upscale it at 3x magnification. Anything more creates memory overflow and slows down processing to a halt.

[Icon67]

I watched 2 tutorials by Adrien Toupet he never mentioned anything about a resolution limit for upload.

[Fenixdg3duy]

@Icon67 I used it daily for video, but I'm not using it for that right now. However, it's unbeatable for still images, and I can tell you that I'm using it every day to achieve incredible results. Your source material is already very high-resolution and is surely in good condition.

[Icon67]

All this assumption to try and negate my honest experience and opinion!

[FurkanGozukara]

i have an amazing app that auto tunes according to your gpu and your target resolution

i already added pretty much all common target resolutions for all gpus

https://youtu.be/_WT4C78j5-c

here app screenshot

[naxci1]

Merhaba @FurkanGozukara

torch.compile windows ta çalıştımı ki? ben çalıştırdım ama daha da yavaşladı, eğer daha hızlı işlem yapıyorsa ve doğru çalışıyorsa bende denemek isterim.

[urixl]

i have an amazing app that auto tunes according to your gpu and your target resolution

Is it paid?

[FurkanGozukara]

hi @naxci1 i just made a test for you on RTX 5090

currently there is a massive leak for VAE encode decode passes when torch compile enabled

therefore I had to seperate vae encode decode and dit tile into sub process

the speed gain happens when you dont use block swap

for example with torch compile it is 2.56 second / it for batch size 9 and target res = 1296x1920

without torch compile it is 3.4 second / it for dit upscale part

torch compile also increasing vram usage slightly so you have to do lesser batch size

[naxci1]

Hi @FurkanGozukara

Thanks for the tests.

Yes, I gave up on that because I think I experienced the same problems. torch.compile works more effectively in Linux and doesn't actually provide a huge speed increase; the main bottleneck is in the VAE decoding process. If we can find a way to speed that up, the process will be much faster. I actually tried several different methods with Claude, I even integrated WAN VAE, but none of them worked properly because SeedVR2's DIT and VAE model files are very interdependent, so other VAEs don't work correctly. Therefore, the best solution is to rewrite SeedVR3 from scratch, because the current code and DNA structure are very complex; neither Claude nor MiniMax could handle this properly, unfortunately.