- Edited
Vrsumo2017 you don't say
https://m.youtube.com/shorts/3_VWzr9HzWw
Vrsumo2017 you don't say
https://m.youtube.com/shorts/3_VWzr9HzWw
AlexSLRLabs Please dont spend so much time in this "innovation". You cant compare a good lighting scene with a fake greenscreen compositing video. The extraction of the greenscreen or the FG-mattes for the actors are just a small part. After that "Roto-Part" you have to tweek the edges. And what is with greenspill? The color don't look right in the skintones. Ah and Greenscreen-Reflections will be a also a big topic.
And after the forground tweaking you have the match the perspective, lighting, colours, black and white values. of the BG-Plate to the FG. Trust me, at the end of the day you will watch a cheese Video, with no quality standards. Or you spend lots of money in the post-production.
Schrock85 You are absolutely right.
Removing the background is one thing, quite easy.
The hardest part is how to place your segmented actor into a new background keeping lights, colors and all this stuff natural.
The problem is known as Image Matting
https://paperswithcode.com/task/image-matting#datasets
For now, we are trying a naive approach: segment actors -> replace the background, done. This is mostly to make an MVP and collect data.
Later we will probably try to spend some more time on this to make it a bit more natural by solving the Image Matting problem
Schrock85 For a bg removal - the data is currently labelled by our data
team for that task. The existing video matting models often give "blinking" artifacts as seen in forum and at greenscreen videos produced with hep of Capcut+manual corrections.
Such artifacts are caused by out-of domain data and especially seen on certain poses (there is no artifacts on people standing straight, but some on doggystyle, for example).
As for comment the greenscreen is only a way to enable passthrough mode, so "greenscreen reflections" are not a problem. The border refinement is a question of model, so it is solvable at good quality like here
The tone/lights corrections based on passthrough enviroment is definetly a hard task but for now most users have quest2 where the passsthrough is black and white where such problem is not relevant
There were some advances on ai - passthrough. So we want to share some of the models are into the developement for porn-specific passthrough. Here is comparison of our current custom approach for with background-removal vs one proposed by capcut. Despite there is dome work to go, there is significant progress on ai-passthroughand infrastructure for alpha-channel streaming
Examples of capcaut-based bg removal
and current model results on the same videos
AlexSLRLabs i can tell the difference
I noticed vrpornnow recently pass through cut outs have improved a lot around the edges
Vrsumo2017 Hell, practically took out the dudes body
That cap cut was fun to mess with, I tried it again and easily it took dudes body out, like in that pic and it leaves the stunt cock.
We are an o way to release a new ai-generated passthrough:
@AlexSLRLabs woah?? for any video?!?!? what software are you using to create the passthrough version?
sletje Tried some existing software but finally stopped on in-house solution - custom trained neural network specifically for porn matting task. Since its still computationally expensive, the alpha for most videos will be precomputed on our servers and streamed. The limitation here is the resources - so like in case of scripts it may not be released for all videos at once, but likely in batches
AlexSLRLabs Would it be possible to add a request for passthrough videos like the scripts have?
JDubs2121 Yes, that's highly probable way to prioritize the videos for alpha-channel creations
AlexSLRLabs I subbed after reading this post. Very excited for this feature. When is it expected to be available?
Hey guys. I had a chance to get AI generated alpha channel for one of my compilations and you can check the quality of AI generated alpha here https://www.sexlikereal.com/scenes/alpha-passthrough-reverse-cowgirl-laying-down-v1-39136
As far as i know there will be an option to switch between having male body or go without male body.
This one is without the male body.
Such an AI generated alpha passthrough takes around 4 hour per 1 hour scene as for last time we checked, so it would be pretty easy to get passthrough for large amount of scenes.
The feature is almost ready to launch, so you can expect it very soon!
Vrsumo2017 Male body in or out? I like having the option for both
Wow! and looks like we'll have the options for both
Video demonstration toggling with body or no body.
https://www.sexlikereal.com/blog/456-advances-in-ai-background-removal-for-passthrough
SLR
AlexSLRLabs im wondering if you're taking suggestions for the first batch of AI PT scenes that will be released?. Lmk if you have any or if i can share. do you think this feature comes mid jan?
Rakly3 And work on multiple different platforms.