The selected order of Processors is IMPORTANT. In your example, you have face_enhancer, face_swapper, and expression_restorer. That means you're running the face_enhancer first, then face_swapper, then expression_restorer. You should instead select this order: face_swapper, face_enhancer, and expression_restorer. Also, the best face_swapper model is still inswapper_128. Use the pixelboost in 3.0 to upscale that. Pixelboost does need more processing time as it tries to simulate more pixel resolution at runtime, but it's still the best compared to the other swapper models right now.
Isn't the backend already doing that though? If I'm not wrong the UI is just the UI, order of selection shouldn't matter there because in the background the code will be running it in the correct order
@@Dhruv1223 No, backend respects the order selected by the user. You can verify yourself by changing the order and watching the processor progress bars on the console and the order they execute in.
@@OlivioSarikas it might JUST because the processor is flipped, it should be face swap inswapper first (it's low res), then you use face restore (restoring fidelity of the lower res inswapper).
they're just bad examples to choose from. celebrities are best because you know and recognize their faces. but, you want to be PC so you took two similar faces nobody knows or recognizes, so you end up with a botched video that fails to deliver the wow-effect that everyone expects and that the software surely can deliver upon. long story short, next time, at least grab a male video from envato and substitute it with your own face that we all recognize and love.
Thanks for posting this :) I'm also on 4 GB VRAM still, and always wonder, when I see videos like this, if I'm one of those who can use these things or not.
Some people don't listen but you definitely mentioned it. This is a face swapper not a head swapper. The comment above about the order being important is valid though. I'm going to try this out later, looks like fun.
@@OlivioSarikas Then shouldn't be called head swapper and not face swapper. Seriously, if you removed the input face no one would have noticed. The stretched face looks 10000000000000% the same face you tried to swap. Swapping faces means you replace the face of the target you want to replace, not slightly alter the head. If this is the improvement, give me the old version that actually works and swaps faces.
LOVE FaceFusion. It's a part of Visions of Chaos. FYI, the big deal isnt the quality of the final product, its that you only need 1 picture to create the deepfakes
Caution. YOU SHOULD WARN USERS FIRST. This AI know-it-all fails to mention that 'FaceFusion' forces install on C:Drive and installs masses and masses of software packages that you don't need for most AI applications, like the C++ SDK, Anaconda, another installation of Python. For most face swap stuff you can run those scripts locally in your Forge or A1111 folders without the huge install.
I was just looking at this yesterday, with the intention of installing it this week. The person was using it with Animatediff files, along with upscaling, to change the face as AD messed it up a bit.
I’m sorry but tensor RT wins every time over Cuda and any comparison tensor RT is the latest performance increase the way that they both optimize in the pipeline are different
The MAIN difference is that with Roop unleashed when something moves in front of the subjects face the thing disappears under the face. So if their hand passes across their face its like the face is imprinted on the hand as well. "Occlusion" in face fusion fixes this issue. Roop-unleashed is great for faceswapping a huge batch of photos. FaceFusion is better for video. I've only been able to get FF to do a single photo at a time. Maybe now the "queue' option fixes that? I cant tell since I cant actually get FF3 to work. yet. :(
Roop unleased has a masking feature so that the object that passes the face will not be covered by the new face.. BUT it works very horrible almost unusable... If you say FF do work best for this situation, maybe I will try d version 3 again.. my last test of FF is 2.6.0 and I did not like the accuracy, roop unleashed gives perfect accuracy... I mean almost perfecr because sometimes, I get inaccurate face on SOME face (Not all face though)
@@ripleyhrgiger4669 No, we are talking about "Roop Unleashed" not just roop. Of course you are right, roop is dead, but "Roop Unleashed" is a fork version of Roop and is still receives a constant update. By the way, upon testing Facefusion latest version 3.0, I can conclude that Roop Unleashed still gives me better output faceswap (If you use a dataset - a multiple input image at once which facefusion has no feature like this)
There is in fact available facefusion extension on a1111 and forge... I am using it last time but since I much prefer the accuracy of roop unleashed, i uninstalled all my FF, the standalone and the forge extension...
I have the same issue with the Pinokio install version. Tried it 5 times and cuda and Tensort do not appear. So if you dont mind your videos being swapped at 1.2 frames per second then great. Otherwise a 5 minute video takes.... I have no clue I got annoyed at 3 minutes when it only had about 200 frames done.
Same here... I had the previous version installed through Pinokio (2.6.1) which worked perfectly and could access CUDA. Upgraded to 3.0, which didn't work and needed to do a fresh install of 3.0 completely. That's when it only used the CPU with no CUDA option! Just wondering if it's a 3.0 issue on Pinokio and may get fixed in an update?
@@pancake7035Yeah, it's annoying! The thing it's, I had a working copy of the 2.6.1 version with CUDA working, then upgraded to 3.0 and now it doesn't use CUDA! Plus, every time I upgrade, it see to download the same massive files (I'm sure I have multiple copies of the same huge downloads scattered all over my machine!) But when it all works, it's great! And that makes all the hassle worth it, in the end...
@@OlivioSarikas Thanks, I assumed that might be the case. Anyway, I've heard that it's not really reliable at the moment and results might vary. Interesting stuff nevertheless.
Nice video Olivio but have to say was a little disappointed as I was expecting you to go through all the options and give a brief explanation as to what they do etc. :)
Seems to be mostly -NONCOMMERCIAL- in its use of 3rd-party processing. _PSA: Check the license of every processor you may end up using commercially._ Much love.
'Bare mode' is explained on the Github repo. If he'd made a video on that, most people probably wouldn't get it to work and this comment section would get flooded with Python questions.
It still looks awful. Nothing like real life at all and obviously AI generated. I was hoping they'd be building on top of something other than the low-res insight-face model but seems not. Until there is a better model to build on top of, every version of facefusion is going to look baaad.
Agree, especially when u use faceset, it give much more accurate result.. but I havent tried facefusion 3, my lasg version I used facefusion was, 2.6.0 maybe I will try again d version 3 and see if roof unleashed is still better...
@@PredictAnythingSoftware Just ONE feature that Fusion not have yet, and that's the MAIN reason because ROOP it's better and it will keep doing better a job, it's the build in feature to CREATE FACE SETS, so you don't just supply ONE image as reference to do the swap, you can supply any amount of different angles and expressions images from the same person as source for the swap . So, ALWAYS will be better results on ROOP because of that, because of having a lot more data as reference to do the swap, NO other face swap app have that feature build in.
With James Cameron joining the board of Stability AI, it's over. The age of AI creativity that's open for the masses is dead. Done. Finished. You might as well start thinking about some other kind of RUclips channel or an entirely different hobby.
The selected order of Processors is IMPORTANT. In your example, you have face_enhancer, face_swapper, and expression_restorer. That means you're running the face_enhancer first, then face_swapper, then expression_restorer. You should instead select this order: face_swapper, face_enhancer, and expression_restorer. Also, the best face_swapper model is still inswapper_128. Use the pixelboost in 3.0 to upscale that. Pixelboost does need more processing time as it tries to simulate more pixel resolution at runtime, but it's still the best compared to the other swapper models right now.
yes, good point! :)
This is the way!
Isn't the backend already doing that though? If I'm not wrong the UI is just the UI, order of selection shouldn't matter there because in the background the code will be running it in the correct order
@@Dhruv1223 No, backend respects the order selected by the user. You can verify yourself by changing the order and watching the processor progress bars on the console and the order they execute in.
Do you need nvidia
Ah, my favorite tool. Have had so much fun with that, it's amazing!
the installation is so simple, i remember facefusion 1 was a pain to install
the face looks nothing like the woman in the photo. still looks like the original face in the video
Agreed. If you told me it was two takes of the same woman, I'd believe you
because the head shape is different. I said it in the video.
A photo of Olivio would have been better :-)
@@OlivioSarikas it might JUST because the processor is flipped, it should be face swap inswapper first (it's low res), then you use face restore (restoring fidelity of the lower res inswapper).
they're just bad examples to choose from. celebrities are best because you know and recognize their faces. but, you want to be PC so you took two similar faces nobody knows or recognizes, so you end up with a botched video that fails to deliver the wow-effect that everyone expects and that the software surely can deliver upon.
long story short, next time, at least grab a male video from envato and substitute it with your own face that we all recognize and love.
Merci ! Super tutoriel !
Also works on low GPUs... I have been a long time 4Gb Vram user with FaceFusion
Thanks for posting this :) I'm also on 4 GB VRAM still, and always wonder, when I see videos like this, if I'm one of those who can use these things or not.
Did I miss something? The ending video looks exactly like the woman on the right and nothing like the woman on the left.
do you guys listen? I said the head shape is different and that is why the face is streached onto the head shape
Some people don't listen but you definitely mentioned it. This is a face swapper not a head swapper. The comment above about the order being important is valid though. I'm going to try this out later, looks like fun.
@@OlivioSarikas Then shouldn't be called head swapper and not face swapper. Seriously, if you removed the input face no one would have noticed. The stretched face looks 10000000000000% the same face you tried to swap. Swapping faces means you replace the face of the target you want to replace, not slightly alter the head. If this is the improvement, give me the old version that actually works and swaps faces.
@@Tempestan dude, be happy you get to play with FREE tools that are cutting edge.
@@gohan2091 *face swapper not a head swapper... made my day haha
LOVE FaceFusion. It's a part of Visions of Chaos. FYI, the big deal isnt the quality of the final product, its that you only need 1 picture to create the deepfakes
Try rope-opal. Its so much better and faster than this
I have issues installing the face fusion it stucked on 5/8 of the installation
When i download the final video/image; it shows the green box around the face. How to remove that green box?
great tutorial video as alaways...
What about expression restorer?... how does it work?... 100% = restore to the source? or the target?
I have a beast of a computer running windows 11 and I keep getting a blue screen whenever I start to use FaceFusion. Any suggestions?
anyone know how to uncensor this yet
Can anyone tell me why my project gets to 100% and the output box is blank, just has a camera icon
cuda option not showing can you help me
is the pinnokio UI default still local? Or do I need to click on local URL
how do you get the dark mode interface?
when i push START it doesnt process the video....
Working for flutist instrument player ?
Thanks!!!!
I’m getting 100% work on my m1 mac, how can i limit the cpu usage? What settings should I change? Thanks
hello . what are computer specs that I should have at least to work on
Nice video!
Can you review other faceswap software such as Rope Pearl and Swapface?
why is the live webcam faceswap very slow?
Modul not found error: no modul named ' numpy ' problem please solve it. One vedio for this topic plaese 🙏🙏🙏🙏
how do i open facefusion after install?
Can you do a video on using this to change black and white photos to color?
tensorrt is going to give you a ~50% speed bost fyi
Caution. YOU SHOULD WARN USERS FIRST. This AI know-it-all fails to mention that 'FaceFusion' forces install on C:Drive and installs masses and masses of software packages that you don't need for most AI applications, like the C++ SDK, Anaconda, another installation of Python. For most face swap stuff you can run those scripts locally in your Forge or A1111 folders without the huge install.
@@EarmWermChannel can you run the newest version in Forge tho?
@@undoriel My older version works just fine, plus there's forked versions lying around.
It can be installed anywhere not just on C drive.. as long as theres no spaces on your path.
@@PredictAnythingSoftware How do you install these on a certain hard drive?
mind doesnt have cuda it says directml, should i use it or CPU
maybe you don't have a Nvidia Card. Try both or ask in the Pinokio Discord :)
Can you use animated character faces?
which os you are using Mac or window
I get no output after hitting start… it runs for a couple seconds then stops. Has anyone fixed this?
they have a discord that is linked in the app. better ask there
I was just looking at this yesterday, with the intention of installing it this week. The person was using it with Animatediff files, along with upscaling, to change the face as AD messed it up a bit.
Even though I clicked on 3.0 to download it has me install 261??
Is it based on the same code that other softwares use? i.e Roop.
Does in handle obstacle in front ofnyje face better than Xseg mask in deepfacelab?
THANKS
Im looking for a a way to remove mosaic from videos.. if you know what Im sayin
JAVplayerEZ if you know what I mean😇
@@rocren6246 8 BUCKS!!!
I’m sorry but tensor RT wins every time over Cuda and any comparison tensor RT is the latest performance increase the way that they both optimize in the pipeline are different
Module not found numpy problem please solve it😢😢
please ask in the pinokio discord that is linked right in their software
How does this compare with Roop Unleashed?
The MAIN difference is that with Roop unleashed when something moves in front of the subjects face the thing disappears under the face. So if their hand passes across their face its like the face is imprinted on the hand as well.
"Occlusion" in face fusion fixes this issue. Roop-unleashed is great for faceswapping a huge batch of photos. FaceFusion is better for video. I've only been able to get FF to do a single photo at a time. Maybe now the "queue' option fixes that?
I cant tell since I cant actually get FF3 to work. yet. :(
Roop unleased has a masking feature so that the object that passes the face will not be covered by the new face.. BUT it works very horrible almost unusable... If you say FF do work best for this situation, maybe I will try d version 3 again.. my last test of FF is 2.6.0 and I did not like the accuracy, roop unleashed gives perfect accuracy... I mean almost perfecr because sometimes, I get inaccurate face on SOME face (Not all face though)
Roop is old and hasn't been supported for a long time.
@@ripleyhrgiger4669 No, we are talking about "Roop Unleashed" not just roop. Of course you are right, roop is dead, but "Roop Unleashed" is a fork version of Roop and is still receives a constant update. By the way, upon testing Facefusion latest version 3.0, I can conclude that Roop Unleashed still gives me better output faceswap (If you use a dataset - a multiple input image at once which facefusion has no feature like this)
Can i use this in A1111 or Forge ?
I don't think so. I think this is only stand-alone.
No this is something completely different than Image generation. Its face swapping from one existing image to another existing image.
There is in fact available facefusion extension on a1111 and forge... I am using it last time but since I much prefer the accuracy of roop unleashed, i uninstalled all my FF, the standalone and the forge extension...
@@PredictAnythingSoftware i installed it yesterday, and now my a1111 is not working anymore
Your system specs??
Better than rope pearl?
Much better and easier. At least as of 9 months ago. Ropepearl was so horrible I never went back.
rope pearl works better to me.
Idk about rope pearl, that one seems kinda buggy, but rope opal (the previous one) is leagues and leagues better than this in basically every way
I dont have the Cuda option?
I have the same issue with the Pinokio install version. Tried it 5 times and cuda and Tensort do not appear. So if you dont mind your videos being swapped at 1.2 frames per second then great. Otherwise a 5 minute video takes....
I have no clue I got annoyed at 3 minutes when it only had about 200 frames done.
Hi, the same issue for me with the local installation. I tried 5 times to reinstall but only showing the cpu provider and no the gpu /:
Same here... I had the previous version installed through Pinokio (2.6.1) which worked perfectly and could access CUDA.
Upgraded to 3.0, which didn't work and needed to do a fresh install of 3.0 completely. That's when it only used the CPU with no CUDA option!
Just wondering if it's a 3.0 issue on Pinokio and may get fixed in an update?
@@Hey-Its-Retro im having the same issue and i dont even use pinokio, i installed it by myself
@@pancake7035Yeah, it's annoying!
The thing it's, I had a working copy of the 2.6.1 version with CUDA working, then upgraded to 3.0 and now it doesn't use CUDA!
Plus, every time I upgrade, it see to download the same massive files (I'm sure I have multiple copies of the same huge downloads scattered all over my machine!)
But when it all works, it's great! And that makes all the hassle worth it, in the end...
Can FaceFusion also be used on images? Or only video? Nice compact video btw!
on images too. :)
Does anyone knows what the expression restorer processor does?
as far as i know it tries to restore the face expression from the original video onto the new face
@@OlivioSarikas
Thanks, I assumed that might be the case.
Anyway, I've heard that it's not really reliable at the moment and results might vary.
Interesting stuff nevertheless.
Is this better than Rope in your opinion Olivio?
dam my windows is 32bit, I can't install pinokio 😭😭 any another ways plz🙏
I dont get the improvements here or why to use it. We already have had better faceswap tools wayyy back for example rope pearl
Can Facefusion change multiple faces in photo/video?
Yes
Nice video Olivio but have to say was a little disappointed as I was expecting you to go through all the options and give a brief explanation as to what they do etc. :)
Facefusion 3.0 is not very AMD Friendly....constantly crashes
only a 10 GB download..
Can you please make a video of any upscale model
Topaz video AI is the best tool for video upscaling, it is the best I tested so far but its not free...
Seems to be mostly -NONCOMMERCIAL- in its use of 3rd-party processing.
_PSA: Check the license of every processor you may end up using commercially._ Much love.
i dont have a GPU to run this
please explain it in Bare Mode not with Pinokio.
'Bare mode' is explained on the Github repo.
If he'd made a video on that, most people probably wouldn't get it to work and this comment section would get flooded with Python questions.
Interesting since ReActor is giving errors.
is Pinokio free?
yes
mostly sick of application installing whereever it want these day shocking set up of pinochio almost made me delete it instantly
So probably only 2-3 more years and it will catch up to where Rope was a year ago!
rope pearl works better to me.
how can i install sd on hard drive ?
rope pearl is better
Yup its all about the settings. I get ghosting allnthe time
try the debugger and face upscaler
👋 hi
Roop is a lot better and very user-friendly
Old wine in new bottle from Microsoft?
It still looks awful. Nothing like real life at all and obviously AI generated. I was hoping they'd be building on top of something other than the low-res insight-face model but seems not. Until there is a better model to build on top of, every version of facefusion is going to look baaad.
Resolution is super low....
that's why i said you need to use the upscaler that is included!
First
second
@@S9universe third
Cuarto
116th! Do people think I'm cool yet? PLEASE LOVE ME
@@RandomPickles yup we love you fellow human friend 🙏
Roop Unleashed still better than this, and a lot faster
Agree, especially when u use faceset, it give much more accurate result.. but I havent tried facefusion 3, my lasg version I used facefusion was, 2.6.0 maybe I will try again d version 3 and see if roof unleashed is still better...
@@PredictAnythingSoftware Just ONE feature that Fusion not have yet, and that's the MAIN reason because ROOP it's better and it will keep doing better a job, it's the build in feature to CREATE FACE SETS, so you don't just supply ONE image as reference to do the swap, you can supply any amount of different angles and expressions images from the same person as source for the swap . So, ALWAYS will be better results on ROOP because of that, because of having a lot more data as reference to do the swap, NO other face swap app have that feature build in.
nsfw?
no
@@happycollapse6348 gotcha
should work since this is only doing the face
It's not hard to unlock but I can't explain here, YT deletes my message.
gotcha, idk why is deleting many things i post also
With James Cameron joining the board of Stability AI, it's over. The age of AI creativity that's open for the masses is dead. Done. Finished. You might as well start thinking about some other kind of RUclips channel or an entirely different hobby.
Cameron is ruining everything and blaming the kids
Rope ftw. Face fusion is blatant ripoff.
6:50 the result is not that great because the processor steps are switched, it should be 1) Swap then 2) Enhance, not the other way around.
🫣 awful
Modul not found error: no modul named ' numpy ' problem please solve it. One vedio for this topic plaese 🙏🙏🙏🙏