Faceswap Forum

I'm curious if anyone has noticed if Warp-to-Landmark is more susceptible to "googley eyes."

I recently used Warp To Landmark on a new model and noticed the eyes are way more googley than usual. This is the second time I've used it so not a lot of experience on this, but it would make sense in my pea-brain because if your eye landmarks are off even a little (which is easy to do,) and the program is directly aiming towards those landmarks (rather than random) it's going to be more susceptible to learn the model wrong.

At first I thought this could be caused by bad data because I added more B-faces which appeared to have wonky eyes (because of weird angles of the face.) But that seems unlikely because out of 6,000+ faces the wonky ones are maybe 25-ish. This made me wonder about the possibility it could be Warp-to-Landmark.

Has anyone else noticed this?

I personally don't put much stock in warp-to-landmarks, but your mileage may vary. IMHO it artificially limits the amount of data you are training on, so I have never looked into it beyond making sure that it is working as intended.

I was just about to edit this post when I saw your reply.

I noticed that I for some reason had "Split G-Block" ticked off on this model(I usually don't use it.) I'm thinking that's the culprit for the googly-eyes. I'm glad I noticed as I was just about to upload my settings for a "Stojo-512." Otherwise this model turned out fantastic in around 700k I.t.s/1.5M e.g.s W/efficienetv2 L