Scale Control | ComfyUI YouTube Tutorial

[sedetweiler]

This works well, but the images are always tiny and when combined with the input image, a resolution change box is seen. Is there a recommended way to make this work with images that are closer to 1024 on a side?

[Gourieff]

The non-commercial inswapper model has only a 128x128 resolution
To get better results you can use facerestores and upscalers
If you pick up the right combination, you can achieve an excellent quality up to a resolution of 12MP in the final result

[Gourieff]

Also you can use an original image (made before a swap) and take some details from there via Photoshop (or another similar App)

[Gourieff]

Here's the 12MP demo:
https://github.com/Gourieff/Assets/blob/main/comfyui-reactor-node/DEMO_KIRA_12MP.jpg

Not a bad result I think for 12MP image with a swapped face, like a mobile phone shot
And if you postprocess it with inpainting tools - it will be much better

Source face:

[sedetweiler]

It would be ideal if we could upscale the replacement face and then merge it back. I don't see nodes for doing that.

[sedetweiler]

I just did a video on your node yesterday, so you should see some traffic. Cheers!

[Gourieff]

It would be ideal if we could upscale the replacement face and then merge it back. I don't see nodes for doing that.

It hardly can be done like this, because a detected face is not an image, it's smth like a mini-model of a face with a lot of parameters
And when we swap faces with help of Insightface library - we don't blend them, we recreate a new face (with a resolution of the inswapper model) in the target (input) image
Then we can work with this result and apply a facerestorer/upscaler

[Gourieff]

I just did a video on your node yesterday, so you should see some traffic. Cheers!

Thank you my friend!
Can you share a link?

[sedetweiler]

https://www.youtube.com/watch?v=ekofgf9T-9c

It would be ideal if we could pipe this into the facedetailer in the Impact Pack, so perhaps we can upscale it and try and keep the corrected face. I guess we need to start looking to some of the other open source models so we can get some resolution going. Your node works perfectly, we just need a model that can scale. Keep up the great work!

[Gourieff]

https://www.youtube.com/watch?v=ekofgf9T-9c

It would be ideal if we could pipe this into the facedetailer in the Impact Pack, so perhaps we can upscale it and try and keep the corrected face. I guess we need to start looking to some of the other open source models so we can get some resolution going. Your node works perfectly, we just need a model that can scale. Keep up the great work!

Thanks!
Great tutorial you did!
Yeah, we will look for an opportunity to get more high res model, it’s not yet clear how to train such models, but I think it’s only a matter of time

[sedetweiler]

It would be amazing if you could output a mask. that way we could have some additional ability to upscale and replace if needed. thanks again for an amazing node!