Early Copycat Experimentation | Rotopaint Nose Removal
I cut off my caption at the bottom accidentally and don't care to fix it, read RTX (2060).
This test was mostly for feeling out the limits and parameters to see what CopyCat excels at. I don't think this was a success by any measure. For the final result I actually feed my footage through the Inference node twice, as only doing so once still leaves a fair amount of nose definition in it. I think the subtlety of the paint out in RBG detail is why the effect feels so soft, and kind of just feels like a censor of the nose instead of a full paint out of it. I think it did a fairly good job with interpolating the varying scale of the face, but ultimately this effect falls flat (about as flat as his new face, har har).
CopyCat Settings:
Epochs: 10000
Initial Weights: None
Model Size: Medium
Batch Size: 2 (Auto)
Crop Size: 512
Checkpoint Intervals: 1000
Contact Sheet Interval: 100
Training Length: 3.5 Hours
RTX 2060 | i7-8700 | 32GB 2666MHz
The .cat(nc) files took up about 1.32GB after all the training finished.
Done in NukeX 13.0v4 in the AIR Copy Cat ML node.
Original Source: • Rando Funny 16