For those who have tried, do you feel the 512 version is an improvement? #171

New Issue

2021-12-07T22:22:04+01:00

paulescobar commented

2021-12-07 22:22:04 +01:00

(Migrated from github.com)

I am converting on mostly SD or sub-HD (720p) target content. I find the new "--crop_size 512" looks very strange. You can see borders on the edges of the new face & sometimes there are strange fluctuations (like wind is blowing beneath the face). The key facial features like eyes, nose, and mouth look a little more "darker" or "contrasty"...and I think the facial accuracy is lesser as a result.

Could this be the result of increased sharpness in this version?

Honestly, I find setting it to "--crop_size 224" produces far better results. It blends properly & looks more accurate. I think this is because it has a "softer" look - which may help produce better accuracy.

I am converting on mostly SD or sub-HD (720p) target content. I find the new "--crop_size 512" looks very strange. You can see borders on the edges of the new face & sometimes there are strange fluctuations (like wind is blowing beneath the face). The key facial features like eyes, nose, and mouth look a little more "darker" or "contrasty"...and I think the facial accuracy is lesser as a result. Could this be the result of increased sharpness in this version? Honestly, I find setting it to "--crop_size 224" produces far better results. It blends properly & looks more accurate. I think this is because it has a "softer" look - which may help produce better accuracy.

AaronPorts commented

2021-12-08 15:00:55 +01:00

(Migrated from github.com)

I think the problem is in the VGGFace2-HQ dataset, which is more contrasting and a little unnatural due to the resolution upscale.

👍 1

nonlin commented

2021-12-08 16:38:02 +01:00

(Migrated from github.com)

Also partial? Didn't they say training on that is in beta or incomplete?

> Also partial? Didn't they say training on that is in beta or incomplete?

lhucklen commented

2021-12-16 04:14:21 +01:00

(Migrated from github.com)

also curious as to if this will be improved upon.

illtellyoulater commented

2022-01-06 17:07:36 +01:00

(Migrated from github.com)

I also agree that "--crop_size 224" produces better results. The new model seems to produce more detailed results, but at the same time it will look a bit off and more unnatural.

acSpock commented

2024-01-27 00:12:57 +01:00

(Migrated from github.com)

I'm pretty sure that one needs to train the 512 model themselves on top of the base 512. Don't believe there exists a complete model available, unless someone could correct me.

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: CalvinBackup/neuralchen-SimSwap#171