JOI Bathtub Hayley Atwell wants you to stroke for her (DFL2 test #3)
I trained on higher res but default dimms to test out, which unfortunately meant smoother skin texture in the end. A lot of trial-and-error also with both GAN and TrueFace, neither of which I was fully satisfied with.
Two biggest issues with this one:
– Level of face details – I’m explaining it with the low dimms and an average src faceset. I’ll try lowering res and increasing dimms to see what depth changes. Also will try working on higher resolution src facesets.
– Overall blend-in – Mostly color issues around the mask area (especially forehead) but also the faceshape wasn’t ideal for the src. Not sure if the size of masking could be improved, since this was WF but finished product looks closer to FF.
Here’s more of the process breakdown for those curious:
Face extraction/alignment is generally frustrating, but even moreso when the tip of a dildo is right on the chin. This is a problem I’m still facing and still trying to figure out – it’s very difficult to properly align face detection around the jawline/mouth position when anything is over the bottom half of the face.
For XSeg, I used a lot of manual masks at the beginning and tons of iter on XSeg itself. There’s still some issues with teeth and the dst index finger occasionally protruding over the face. Blurry fast-moving obstructions in front of dst face made it tough to really narrow details.
SAEHD Training – WF/512 res/DF-UD/256 AED/64 ED/64 DD/22DMD
As mentioned, training with high res but low/default dimms. Not the best combo, however this allowed me some testing flexibility. Lots and lots of pre-training before any real start.
0-15k iter – Default settings with Random Wrap and mask training (at that point the face was relatively shaped thanks to pretraining)
15k-100k – Defaults but no RW, with mask, and specific facial feature priorities
100k-140k – Defaults no RW, with mask, no specific facial feature priorities
140k-220k – Defaults no RW, with mask and Uniform Yaw
220k-300k – Defaults no RW, with Uniform Yaw, no mask priority
At that point loss values (and face) had plateaud so I tried two parallel versions to add and verify face details. Both with GAN, and one version with TF.
300-340k – GAN 0.1 + TF 00004 V1 / GAN 0.1 V2
Ultimately settled for V2 that you see here. TF gave me even more color issues than the current version for no improved face details (probably due to faceset used).
Tried my hands at SOT-M without super res, and a bit of blurring around the mask to finish the blend. I’m satisfied with most of the overall color, although SOT-M may not be worth the processing power compared to other color modes (like RCT or IDT/IDT-M).