Different samplers and other thoughts and questions

[osi1880vr]

First of all, thank you very much for sharing this beast, it is fast as light and produces loads of good results. On my 4090 I was able to run upscale to 7000x7000 which no other upscaler managed to get done, at least not in time :D.

I wonder if it would be possible to add that we can choose different samplers to play with, by now it is using hardcoded euler.
Or is it that only euler would work?

Other question is unloding the model when doing the VAE step.
Is see it takes a big chunk of extra memory when running the VAE especial when you try to get around tiled VAE.
I do this in other workflows where I do unload the gen model when running the VAE step as this would free some more memory.
So would it be possible to allow unloading the DiT model when doing the VAE step and also unload the VAE model when doing the DiT step?
It will add some time but for sure the free extra memory can be used. I would guess it could invite the smaller cards for even bigger adventures.

I did try running the compiled model and it came out slower which was a surprise after it took me a while to get it working at all, including hacking the code a little. it seemed to use way more memory then the non compiled which I have been told is totaly normal but then moving around that extra memory added so much time that the speed gain during DiT was eaten away :D is that me or is that just part of the process?

sorry for mixing so many topics at once.

Keep up your great work. I'm already impressed beyond Awe :D

[adrientoupet]

Hi there,

To your questions:

• Offloading dit/vae, is that not the same as setting the offload_device on the dit/vae loader, or you're looking for something different?
• Different samplers: If you're keen for this, can you create a new issue in https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler/issues and explain which one you want to add, and what value that would bring, then we'll try to get to it when we can.
• Compiled uses a bit more memory but should be faster, not in the first batch, but in the subsequent ones. If you have a use case where its not the case, please share debug log and machine specs and create a new issue for it please.

Thank you!

[adrientoupet]

Thank you - yes I tried to promote anything that was useful and leave the rest out. But I'm not perfect so appreciate the community playing with stuff and making suggestions. One request, please open a new issue here https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler/issues with example image showing the seams artifact so I can look into it. Thank you

[osi1880vr]

Well I would not do that ;) We got to find a different way since they are NSFW ;)

[adrientoupet]

Ah yes please dont do that. But if you can find all public images that show the same artifact, please share them. Thanks you for being mindful.

[noland81]

I upscale large files so I often encounter seams. The trick here is to split the title decode into 512 blocks. The problem will be much less

[adrientoupet]

@osi1880vr - LAB shouldnt show artifact anymore in v2.5.11 - thank you for reporting this issue.

[osi1880vr]

I can confirm LAB is now not showing the artifact anymore on images where I could reproduce it with the old version, thank you very much =)

[adrientoupet]

Awesome, thank you very much for confirming!