Planning to port to DFL 2.0
Added 2020-05-23 16:36:44 +0000 UTCThe images on the left are 256x256 pixels, the ones on the right are 128x128.
With DFL 1.0 i could never imagine training at over 128x and get good results because i don't have that powerful a GPU, but now with the optimization of DFL 2.0 i might be able to create even more realistic fakes.
With the introduction of several model types, one of which is basically a "head swap" we might soon see fakes (If i can do anything about it) that are practically impossible to tell apart from the real thing.
Anyone who has ever done any kind of video/photo editing, knows that the more you enlarge the image the more artifacts are visible in the enlarged image. Now with the trained images twice as big, the resulting fakes will theoretically have twice the clarity.
I would train at 512x512 (which DFL 2.0 supports) but it would require over $10,000.00 worth of GPUs lol
