Open xllusion-dong opened 1 week ago
You can check the upstream code for comments explaining them a bit: https://github.com/KwaiVGI/LivePortrait/blob/d09527c76275969e99f4d9e29775f02df1b4c48e/src/config/argument_config.py#L16-L39, some are exclusive and will be ignored if retargetting is set for instance.
The 4 first are related to the internal cropping for the driving video, I always left them at default for now
wrong face is because "starting face" in video is too different, im using skip first frames box until it wotks fine, sometimes on 3rd frame sometimes on 24th frame it works fine . With some videos it works from frame 0 but the head needs to be super stable, you have a shot from a movie - this is quite unstable , if it would be aligned with dlib or somethin else, it would work better
Does anyone know how to increase the output resolution? I can't get it higher than 1280. I've been messing with the .py files looking for any reference to 1280, and changing it to 1920. Nothing has worked so far.
I'd also love to know what Dsize and Scale do.
input image size determines resolution
input image size determines resolution
That's not the case in my experience. I was working with 1920x1080 input footage, and it was doing some internal scaling that downsized it to 1280px. To double-check, I did a 2x upscale of all of my input footage, and the output was still 1280 px.
change it , imo it should be done so it checks res of the input image
change it , imo it should be done so it checks res of the input image
Thanks for the reply. I changed all of those before I asked the question. I've been testing this extensively to see if it's a viable animation tool. As you can see, there is a fairly obvious difference in quality between the input and output. I'm hoping there's a way around that.
Hold on.................
For some reason those settings started working after a computer restart. I had previously restarted Comfy and it didn't appear to work. Very strange. There's still a marked difference in quality though.
Maybe that's just the way it is. Here's the output:
change it , imo it should be done so it checks res of the input image
I think we can just disable that whole resize function, if wanted it is better done with other nodes anyway.
It will be appreciated if there is explation for these setting.
So I could find out if it is possible to make result better by adjust setting. Or it is the limit of model. Some output not as expected, wrong mouth shape or others.