Whole image upscale

[Pete11]

Hello,

First of all, you did a great job with SeedVR2.5. it really is amazing and I appreciate the sharing. I want to ask you, it is mainly focussing on faces?! And, if it is, there is a way to improve the whole image?! 2 questions I supposed :)
I tried many times and no matter what it let's the others things in the image without upscaling while the faces are perfect, it looks wierd.
Thank you in advance!

[adrientoupet]

Thank you for the kind words. It shouldn't just focus on the faces - do you want to share some input examples with the workflow you're using? It should be upscaling the whole picture.

[Junchien999]

I found during testing that it tends to focus easily on faces. I suspect this is either an inherent feature or due to the model being trained on a larger volume of facial data.

[Pete11]

Hi, the workflow is basic. I mainly use it for image restoration with Qwen Image Edit 2509 and Flux Kontext, after input the image to SeedVR with 2k upscale using 3b parameter model.
Before upscale it with SeedVR I downscale the image to about the size the one you have it in the youtube video

[adrientoupet]

Make sure there is at least 2 to 3 upscale factor between the input resolution and output resolution. You can try the 7b fp16 model. Make sure you're using the latest version (v2.5.10). Outside of that, if you want more help, share some example input/output so we can look into it further. Good luck.

[Pete11]

[Pete11]

The first image is before restoring, second after upscale and the third is after restoring

[shodan5000]

Tried myself but couldn't get SeedVR2 to restore this in any sense of the word. This is just Qwen-edit.

[urixl]

Whoa, that's good.
Did you use Qwen in ComfyUI or elsewhere?

[shodan5000]

It was 4 step nunchaku qwen-edit 2509. The provided example workflow in ComfyUI.

[Pete11]

:) This is what I'm saying. Don't know why but it kills me! :). The thing is Qwen Edit is messing up the facial features, I compare the images with a face compare node and didn't get pass 52%. Flux Kontext does better but the rest of the image is kinda loose. I managed to send a mask just with the background and either way, SeedVR2.5.10 doesn't help a lot. I'm still trying but all I got is a very beautiful face and loose background. I don't give up though! :)
Thank you for your time!

[urixl]

Just for fun: Topaz Wonder+Face recovery afterwards.