Why don't we get into private deepfakes?

KingpinVR

metichemsi i agree

-aoxomoxoa-

So yes we can do some of this already with existing technology. Results will vary in quality. Currently, we can change and mimic voices with models and using the transcript with appropriate inflection you can get some authentic sounding voices cloned. In terms of swapping the face, I use Rope (Github), which is a wrapper for Insightface, which uses a stable diffusion like process and you can import the video into it, it will track the face (you can have multiple faces like in LR VR, or just multiple people in general) and then you choose images of the face you want to swap. It works best to have the face with multiple angles and multiple lighting conditions if possible, I use like 15 images per video. The software then steps through each frame, identifies the face and swaps it for the desired face.

Things that will cause problems include if her face gets occluded (like she is sucking dick) or if she is too close to the top of the camera lens or just getting too much lens distortion on the face. The software doesn't really know how to handle this. Of course, to move the technology forward, we would need a custom solution tailored for VR, either by projecting the video onto a flat rectangle before swapping the face, or training a custom model that can learn how to swap a face that is very distorted by the curvature of the lens. I have some ideas but would need time to play with this to see what works and what does not.

Currently, there really is not a full body swap technology either, but again that can be something that can be worked on in a similar method that the face model was trained on as well. If people want to try this I recommend starting with a trimmed video where the model is mostly in the center of the screen for best results, maybe in missionary or riding as long as she is not too high up.

In terms of how long this takes, I have an i7-10700K CPU @ 3.80GHz with a 3070 and the process tends to take about 3-4x the length of the video, but this will vary a bit depending on resolution and frame rate.

KingpinVR

-aoxomoxoa- can you make a guide on this?

muSLR

I mean VR deepfakes have been a thing for a very long time... Although the quality does vary, but it's definitely getting there.

https://mrdeepfakes.com/videos/virtual-reality

You can face swap a non-vr videos using deepswap.ai

Although the quality of the video isn't all that great given that it is simply an image that they are referencing.

You can try to generate your own with DeepFaceLab, it's an open sourced tool that is capable of achieving some pretty impressive results. You do need a higher end GPU to achieve decent training speeds however.

I think the biggest bottleneck currently to generate a quality piece of VR deepfake with a custom face as a service is computing power. It's not as simple as stick a face on a VR video, and expect it to work flawlessly. I definitely think this is achievable in the near future though, just need the technology to mature a couple of more years at the slowest.

On the subject of ethics, personally speaking I don't really see an issue of people creating deepfakes for private use.. I don't see that being any different than fantasizing about being with someone, which everyone has done at one point or another. I think the ethics comes in play only when the creator starts to distribute these deepfakes without consent, and that's what really should be monitored and prohibited.

Vrsumo2017

muSLR damn that's bad ass, omg hopefully they gonna use the voices too!

Id really love hear Ariana Grande's voice, Selena gomez. 🥵🤯

Vrsumo2017

muSLR how do watch the mr.df in vr?

wafa

This idea is really something id like to see and use

-aoxomoxoa-

OK so there are a few free software involved. You will need a mid-range+ NVIDIA GPU (30 series and up I would think).

To create clips of your VR video you want to face swap, I recommend using LosslessCut. This is extremely easy to use and very fast, as it cuts the video where you want to without re-encoding it. If you are unsure how to do things in Losslesscut, there are lots of YouTube videos explaining and it is a really user friendly program.
https://github.com/mifi/lossless-cut

For the face swapping software, you can use Rope, follow the python installation and guides at the link below. I would practice with the 2D 1080p 'benchmark video" to get the hang of it, and to optimize the speed of swapping for your system by increasing the number of threads. The author has a discord and made some videos showing how to use it as well.
https://github.com/Hillobar/Rope

Now if you want to change the audio as well, you will need to use a software called FreeVC (Voice to Voice conversion). Here, you would take the voice clip of the audio that you want to convert (that can be exported as a *.wav file from LosslessCut alongside the video) and a 6-15 second sample clip of the voice you want to convert it to. Once the voice has been converted, you can substitute the new audio *.wav file back into your video using LosslessCut again.
https://github.com/OlaWod/FreeVC

Now, I have to use FreeVC by means of the HuggingFace Demo or there may also be a Google Colab linked on that page, as my 8 GB VRAM 3070 does not seem to be enough for FreeVC. I think it has more to do with the software not being optimized for memory efficiency tbh.

And that is it. It takes a bit of work to get everything installed but the workflow itself is not overly difficult. I recommend starting with a clip where the model is mostly in the center of the screen and facing forwards for best results, maybe in missionary or riding as long as she is not too high up. Once you get a feel for it, you can see where you can push the boundaries a bit. Results will vary, but they can be quite good.

KingpinVR

chris5554 any update?

mapratt1771

chris5554 sorry just noticed i had recieved messages. i never get on here anyway to answer your so far questions, no i dont have a tutorial vid, dont really want to make one unless there's, idk, some incentive. I have managed to get 8k. the software I use to put it all together "after effects" idk when or how but seems to be exporting the videos now at an extremely reasonable file size. i just upgraded my pc recently (gpu and ram) and my projects seem to go quite quickly now to a couple hours for a couple min 8k deepfake. my pc specs:
Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 4.00 GHz
64 GB DDR4
RTX 4080
any other question feels free to ask, sorry for delayed responses

KingpinVR

mapratt1771 right on dude, greatly appreciate it mate! What program or software do you use to do this? I am really curious about this and want to learn how to do it myself. The VR community appreciates your help as well im sure!

mapratt1771

KingpinVR the actual deep fake software I use is swap face, can download it for free at swapface.org.. the free version has limitations as per number of deepfakes you can do and includes water mark.. but if you're going to do all the video editing anyway, it shouldnt be an issue. I mask out the watermark in my final draft.. also a trick I came up with to have unlimited videos with the free version, is that I double up the video that is being processed, in other words, the video length is doubled as it loops back one time, that way I can monitor the process and cancel it after it reaches above 50% still rendures what you have processed tho it excludes audio, which again doesnt matter if im doing all the editing anyways. with all the trial and error to get the perfect deepfake, could take up all your free ones in no time. with my method you have infinite. as long as you remember to catch it before it reaches 100% complete. I split the vr video as it is two seperate perspectives.. i do the left eye and then the right. in after effects I put them back together. i've been messing around with this stuff for a couple years now.. I know most the ins and outs to get what you desire with pictures and videos.. can make any picture and any 2d video "3d" not vr unfortunately but at least it give the appearance its popping out at you. there's more tips regarding swapface, for example, you can also deep fake a picture. so I'll take a screen shot of vr video sometimes and deepfake the screenshot with my target picture to get a better match up.. then that deepfake pic I just made will now become my target face for the deepfake video. they currently have 4 deepfake modes. fast, pro, expert, and hybrid... the best to us as of lasted update is expert. that is if you dont have anything going in front of the person's face.. (hands, legs, cock, ect.) its no biggy if it is for a moment, the deepfake will capture the face again once in sight. lack of immersion with it but oh well, if you aint patient, you get what you work for.. if you do happen to have something in front of the face more so a bj scene. I highly recommend useing pro mode.. it does the best job with incorporating the D as well as the lips around it.. not perfect by any means with it tho. a deepfake screen shot of the bj scene will increase your chances at a more realistic deepfake if use the deepfake pic as your source for the replacement in the deepfake video. you can do a lot more touch ups and spend more time to actually get it perfect in a good video editing software.. agian mine is After Effects.. youtube tutorials cover everything you can think of with that software... I did however spend the last 5 years teaching myself how to use it.. if you're just starting fresh with no software or nohow, you're gonna get frustrated.. if you really really want to be able to do this... you are going to have to put in the time to learn and the time to seek out free versions of all the software on the web. I got about $40k worth of software... i'll let all guess what ive paid for it.. anyway.. what does this all accomplish? well.. i can pick just about any vr girl i want. and put the girl of my dreams as the face.. all in 8k. and if you really want to know how to top that Xtacy, pick up one of those full body tpe/silicone sex dolls for a couple grand, heat the inards of it with heating rod, throw on your vr with your deepfake of the girl of your dreams, and go to town on that doll... that feeling can also be multiplied on some recreational stuff.. ;-)

hypertsunaz

mapratt1771 There are already methods to make 2D videos 3D for VR. As in SBS, but not sbs180. I've converted a POV autumn falls scene because she has no vr videos and that'd be the closest thing lol.

GetReal

I think this is a cool idea then you could have scenes that you didn't even like become your favorites.

thatoculusdude

hypertsunaz how good is the result? Mind sharing the name of the app you used? Thanks!

muSLR

Vrsumo2017 There are only a couple of quality VR creators there, I sub to their fan-topia for pretty cheap and download the 4k+ versions of their clips.

hypertsunaz

thatoculusdude Very good. Though takes quite awhile even with my 11th gen i7 and rtx 4080 for long videos. It's Owl3D. You can use it free with limitations.

mapratt1771

thatoculusdude
I use 3DCombine can convert any video or picture to a 3d sbs.. has many parameters you can set for depth. can aslo generate depth maps for both video and pics too.

thatoculusdude

hypertsunaz thank you!

KingpinVR

muSLR Link?

Redhulk50

What about Deepfakes of your favorite Porn star from 1980s 90s etc and porn stars who have retired and never did VR Porn ?

« Previous Page Next Page »