Kiernan Shipka lookalike POV fuck – model update, test/comparison.

Original Pornstar: Lily Rader

Description: In this video I present a comparison between the original version of my Kiernan Shipka POV scene, done using DFL 1.0 SAE (with SAE+ mod) DF, 128 res, default dims full face model that was masked using the old FANSeg model vs new version done using my current DFL 2.0 SAEHD DF 192 res, default dims full face model masked using my new general XSeg model and also a 3rd version using my new DFL 2.0 SAEHD DF-UD 288 res, default dims whole face model, also masked using the same general XSeg model. Personally I think my old DF 192 holds up and looks most similar to SRC, although it may have something to do that this model previously had true face ran on it, it also had different colors than the DF-UD model, possibly because it had color transfer and style power ran on it before but for this run both DF 192 and DF-UD 288 were trained without the use of true face, GAN, style power or color transfer (except for RCT during merging). Both were trained for the same amount of time (but DF-UD was ran on faster GPU because it just needs more iteration to look good) at batch size 4, first with random warp enabled, then with random warp and lr dropout enabled, and then with random warp disabled and lr dropout enabled. Merged using the same masking method, same blur/erode setting, with upscaling set to 40. Color matched and sharpened in post (same settings), noise added on top, no additional post processing done.

DF-UD does seem to work better on closeups, which is understandable but overall it doesn’t seem to be affected as much by the face enhancer/upscaling algorithm used in the merger, hence faces that aren’t close to camera don’t benefit from it, unlike lower res models which do, effectively making the only point of high resolution model being extreme closeups where native resolution makes faces look more correct than lower res one where upscaling can make it look a bit ugly. Without upscalin higher res model does produce slightly better looking and sharper faces but without GAN it never get’s truly sharp (see my Billie Eilish fakes for an example of a model that had GAN ran on it, although it doesn’t work as well anymore sadly and I can’t simply turn it off as I did on that one and have it keep the nice detail, in particular in the eyebrows).

The other thing I’ve noticed is that whole face can be more tricky to correctly composite and color correct than full face, especially on challenging scenes with changing lighting. Lower res models also train and merge faster, which is important to me, I might in the end come up with a compromise and use a resolution of 224 which should still offer slightly better detail than 192 and still work nicely with the upscaling in the merger. Hence I will probably go back and prepare one more model – 224 DF-D full face.