Teds Pics

[Ted]

I would be very grateful for the workflows. Is this done on a ComfyUI platform?

On the Comfyui platform there's a "Upscale image" node that you can connect between the VAE Decode node and the Save Image node.



Upscale node.jpg

That method of upscaling doesn't resample the image using a model, so any corruption in the original image will exist in the upscaled image.

[BobDarkroom]

I would be very grateful for the workflows. Is this done on a ComfyUI platform?

On the Comfyui platform there's a "Upscale image" node that you can connect between the VAE Decode node and the Save Image node.



Upscale node.jpg

That method of upscaling doesn't resample the image using a model, so any corruption in the original image will exist in the upscaled image.

Thanks for pointing that out Ted.
I will look at the workflows you sent to figure out how to do it properly.
As said, I'm a newbie so I still have a lot to learn

[Max-Plank]

That method of upscaling doesn't resample the image using a model, so any corruption in the original image will exist in the upscaled image.

I think I understand. The method you suggested is called "inpaint" or "refine".

Can I ask you how you make such LoRA? Thank you

[Ted]

That method of upscaling doesn't resample the image using a model, so any corruption in the original image will exist in the upscaled image.

I think I understand. The method you suggested is called "inpaint" or "refine".

Can I ask you how you make such LoRA? Thank you

All these AI terms confuse me a bit. I think I'm just basically doing an image to image.

As for making LoRAs, that's quite a bit of work. You need images of what you want, and descriptions to go with them. I used about 70 images of whip marks from movies and 3D art. Then you use software (kohya_ss) to train the LoRA. You'll need a video card with at least 12 gig of VRAM. There are lots of you tube videos that go into detail.

[xs70]

Image to image after upscaling is (one way of doing) HiRes. Because of the higher resolution, the effect of a HiRes is distinct from either an image to image or an upscale (the model will add extra detail).