CJ_Sucks_At_Life
DF Vagrant
Hey, I'm very new to deepfaking and I have some questions on collecting datasets.
I'm currently using Faceswap(Because its one of the few programs that supports AMD GPUs) Dfl-H128 (Since its a decent trainer that doesn't take a week to train) with mostly default settings (enabled sharpen)
But i'm kind of having some confusion when it regards how the datasets effect how my initial deepfakes will look.
My first deepfake was trying to swap jerma (Since he has 6 hours of greenscreen footage) with the newsman who can pronounce Llanfairpwllgwyngyllgogerychwyrndrobwllllantysiliogogogoch perfectly. I trained the AI on the first minute of the Jerma greenscreen and the respective clip (1700 frames and 500 frames)
They turned out alright, albiet a bit glitchy when any turned to the side. With also being a bit blury. For some reason the newsman one did better but that might have been from the large difference in skin tone.
However. When I try with a larger and what I think is better dataset. Specifically attempting to swap Jerma (using around 5000 frames of the shots when he is close in the greenscreen video as opposed to when he is just standing in the background) and Linus Tech Tips (4000 frames of him sitting and talking from "Red's Overpriced "Mini Mag" Cards - The Real Story"). They turn out horrendously blurry.
So i'm just here to ask, what should I look for in datasets? The only thing I think could be screwing up this comparison is that the linus dataset had some issues with it thinking some profile pictures were faces. However extraction seemingly already labeled those different faces (Like "Linus_000275_0" VS "Linus_000275_1") so I don't know how badly it effected the data.
Sorry if I sound like a dumbass, new to this stuff
I'm currently using Faceswap(Because its one of the few programs that supports AMD GPUs) Dfl-H128 (Since its a decent trainer that doesn't take a week to train) with mostly default settings (enabled sharpen)
But i'm kind of having some confusion when it regards how the datasets effect how my initial deepfakes will look.
My first deepfake was trying to swap jerma (Since he has 6 hours of greenscreen footage) with the newsman who can pronounce Llanfairpwllgwyngyllgogerychwyrndrobwllllantysiliogogogoch perfectly. I trained the AI on the first minute of the Jerma greenscreen and the respective clip (1700 frames and 500 frames)
They turned out alright, albiet a bit glitchy when any turned to the side. With also being a bit blury. For some reason the newsman one did better but that might have been from the large difference in skin tone.
However. When I try with a larger and what I think is better dataset. Specifically attempting to swap Jerma (using around 5000 frames of the shots when he is close in the greenscreen video as opposed to when he is just standing in the background) and Linus Tech Tips (4000 frames of him sitting and talking from "Red's Overpriced "Mini Mag" Cards - The Real Story"). They turn out horrendously blurry.
So i'm just here to ask, what should I look for in datasets? The only thing I think could be screwing up this comparison is that the linus dataset had some issues with it thinking some profile pictures were faces. However extraction seemingly already labeled those different faces (Like "Linus_000275_0" VS "Linus_000275_1") so I don't know how badly it effected the data.
Sorry if I sound like a dumbass, new to this stuff