Lower pretrained model resolution possible?

[instant-high]

Is it possible to train, or has someone, willing to share, a lower resolution trained model? e.g. 256x256 pixel ? I would need it for speed of inference reason and limited GPU memory. I can run codeformer on my GTX 1050 / 2GB, but it's a bit slow (1.6 it/sec.)

On the same GPU I can run GPEN 256 togehter with faceparsing and thin plate spline motion model combined in one inference script.

Regards Thomas

Edit: The project mentioned above is a rewritten wav2lip using insightface/onnx for face crop and alignment, faceparsing for masking the face, tpsmm for optional animating the or swapping another face, and gpen for final face enhancment. Depending on torch version and graphic driver version, I can run all together in one in inference.

[day-dreaming-guy]

Have you managed to get this working with Wav2Lip?

[instant-high]

@day-dreaming-guy https://github.com/instant-high/wav2lip-onnx-HQ

[day-dreaming-guy]

You are really Instant @instant-high ! Haha <3 We are trying to run this on local iPhones. Do you think it's possible? What's the RAM requirements?

[instant-high]

On windows selecting xseg, face enhancer gfpgan and frame enhancer ~ 12gb Without frame enhancer ~ 4gb Only pure wav2lip no enhancer or else ~ 1,5 gb for target video 1080p

[day-dreaming-guy]

Got it. Are you available for paid consultation by any chance? We really need help, stuck on this issue

[instant-high]

Sorry, I don't think that I can help you. I'm doing all that stuff for fun, no professional I know nothing about iOS or iPhone.

Have you tried to run the minimum version of wav2lip-onnx first?

Btw. we should switch over to my repo, sczhou will kick us out...