- Categories:
- Masturbation
- Striptease
- Celebrities:
-
Chloe Grace Moretz
Chloe Grace Moretz427255525178
- Original Pornstar:
- Mary Queen
This is my first ever deep fake. Not perfect, but considering the low quality of dst video I'm relatively pleased with the results (with a couple exceptions below).
Two biggest issues I ended up with:
- Some masks/faces are missing here and there - I think there was an issue originally in the masking process.
- Some deep fake masks were outright unusable so had to remove them (e.g. SAEHD produced a giant mouth or complete blur as the frame's mask). Not sure what's going on there - probably issues with face detection on those particular frames.
I wish there was a way to copy/paste one SAEHD src mask onto another dst frame/mask during the merging process - that would be a great bandaid solution when there's only a couple frames missing here and there.
Here's more of my process for this video:
Faceset
Mary Queen/dst - WF/256/85% JPG for the dst images from source video - original footage was relatively low definition and aligned images weren't fully there
Chloe Grace Moretz/src - WF/high-quality aligned images
Masking
dst - Very tricky to mask, not only because of blurry/low-def, but also because her hair was constantly getting in the way, plus she seemed to constantly be at odd angles. Ended up initially doing nearly 100 manual masks on Xseg. The machine took care of the rest, but it still ended up missing a few images and bad recognition
src - masks were done essentially automatically thanks to high quality faceset
SAEHD Training (128 res - batch 4 - 400ms average)
Pre-trained for 135k iter with provided dataset (both genders)
0-45k iter - default values with face flip and random wrap
45k-155k iter - default values with eyes priority and face flip, no random wrap
155k-165k iter - default values with yaw and face flip, no random wrap, or eyes priority
Ultimately got to 0.1850/0.0639
Merging
As mentioned above, the biggest issue was discovering that training generated some horror-esque giant mouths/blurry faces instead of useable masks. Still not sure if the issue is with the original dst masking, the dst faceset, or something else.
Color-wise, it was a weird match but ended up settling on IDT/IDT-M for most of the video (5 C presses from start). Chloe/src feels almost like she's wearing a wig for half the performance - probably because I cut around Mary/dst's hair, so it adds a sense of depth by default. I mitigated some of it by occasionally going with hist-match, but you lose the richness of skintone. I also occasionally blurred the mask, notably to avoid aliasing on edges.
Two biggest issues I ended up with:
- Some masks/faces are missing here and there - I think there was an issue originally in the masking process.
- Some deep fake masks were outright unusable so had to remove them (e.g. SAEHD produced a giant mouth or complete blur as the frame's mask). Not sure what's going on there - probably issues with face detection on those particular frames.
I wish there was a way to copy/paste one SAEHD src mask onto another dst frame/mask during the merging process - that would be a great bandaid solution when there's only a couple frames missing here and there.
Here's more of my process for this video:
Faceset
Mary Queen/dst - WF/256/85% JPG for the dst images from source video - original footage was relatively low definition and aligned images weren't fully there
Chloe Grace Moretz/src - WF/high-quality aligned images
Masking
dst - Very tricky to mask, not only because of blurry/low-def, but also because her hair was constantly getting in the way, plus she seemed to constantly be at odd angles. Ended up initially doing nearly 100 manual masks on Xseg. The machine took care of the rest, but it still ended up missing a few images and bad recognition
src - masks were done essentially automatically thanks to high quality faceset
SAEHD Training (128 res - batch 4 - 400ms average)
Pre-trained for 135k iter with provided dataset (both genders)
0-45k iter - default values with face flip and random wrap
45k-155k iter - default values with eyes priority and face flip, no random wrap
155k-165k iter - default values with yaw and face flip, no random wrap, or eyes priority
Ultimately got to 0.1850/0.0639
Merging
As mentioned above, the biggest issue was discovering that training generated some horror-esque giant mouths/blurry faces instead of useable masks. Still not sure if the issue is with the original dst masking, the dst faceset, or something else.
Color-wise, it was a weird match but ended up settling on IDT/IDT-M for most of the video (5 C presses from start). Chloe/src feels almost like she's wearing a wig for half the performance - probably because I cut around Mary/dst's hair, so it adds a sense of depth by default. I mitigated some of it by occasionally going with hist-match, but you lose the richness of skintone. I also occasionally blurred the mask, notably to avoid aliasing on edges.