DeepFaceLab deepfake tutorial, using generic XSeg
ฝัง
- เผยแพร่เมื่อ 9 ก.ย. 2024
- This video was made to show the current workflow to follow when you want to create a deepfake with DeepFaceLab.
When the face is clear enough, you don't need to do manual masking, you can apply Generic XSeg and get auto mask, otherwise you have to proceed with the manual creation of XSeg masks.
The quality of the final deepfake, of course can be much more better but for this video I trained the model for few iteration because the video is only meant to explain the workflow.
CHAPTERS:
0:00 Frames extraction
1:26 Faces extraction
4:00 XSeg masks
5:35 Training
10:20 Merging
14:00 Final result
GitHub link: github.com/ipe...
_You, the uploader of this video are a GOD!!!! Unfortunately, I find this software very intimidating and I know I'll get lost trying to follow the steps. Good job for showing this to the world though!!_
I've just installed it now. Super easy to follow the steps. All you have to do is basically play the video, pause, do the same thing, repeat until done. The only thing is, it takes FOREVER to process each step.
@@RazorRamonElJefe _There is a website where you can do this now for image or video and it's open source and it's free. :)_
Awesome video! :) Thank you for this and God bless!
Step 1:
extract images from video data_src
extract images from video data_dst FULL FPS
Step 2:
data_src faceset extract
data_dst faceset extract
Step 3 (XSEG MASKS):
.XSeg Generic) data_dst whole_face mask - apply
.XSeg) data_dst mask - edit
.XSeg) data_src mask - edit
Step 4:
train SAEHD
Step 5:
Train SAEHD
Final:
Merged to mp4
I don't have the folder with the files that are in the video. This is a scam.
It's not a scam, it's you that you don't know how to use software. Learn before talk.
@@deepfakeengita2967 i found video with full tutorial...you need to create full tutorial
@@ethnsec please share the link
@@deepfakeengita2967 Hello man, is this still a relevant tool? I am so much in need of suchj a tool. Are there better alternatives now? Anyway what I would like to know is: do you think your project work faster with BIGGER VRAM? I need a tool that work fast, what do you think?
I like your choice of music
SUPER DUPER RESULT!! Please kindly advise how long does it take to do this kind of video ? As i could notice on the computer watch - couple of days?
1 night
@@deepfakeengita2967 Does this work well for swapping a single face within a video containing multiple faces?
any replacement app/workflow for whole head swap?
Do I have to draw the outline of the mask to every frame?
Edit: you don't have to, you only need to draw a few to give the ai an outline and it will do the rest.
how many would you do? maybe once every 50 frames? or just whenever there’s noticeable movements?
The content is likely great, but every time I open a tutorial video and there is no voice narration, I am lost. I can't concentrate on what to do if there is only a mouse pointer circling in some area. I appreciate your work, but can't utilize it. Maybe it is just me, but I guess this is true for many auditive learners. Best wishes, though.
Exist a voice tutorial. I'll do a new one.
In the first step I got this error: ''failed to load the native TensorFlow runtime''. What shall I do?
Nice tutorial!! Does this work with any picture or video I desire? I want to use my friends' faces and mine in an action scene of a movie!
Yes of course
My process -
2) extract images from video data_src
3) extract images from video data_dst FULL FPS
4) data_src faceset extract
5) data_dst faceset extract
5.XSeg) data_dst mask - edit
5.XSeg) data_src mask - edit
5.XSeg) train
5.XSeg) data_dst trained mask - apply
5.XSeg) data_src trained mask - apply
7) merge SAEHD
This is exactly what I did. Now while I started to merge, the model seems to apply the destination face itself to data_dst. What am I doing wrong?? It's not applying source face for some reason.
When I run 6) train SAEHD the process stopped after «Model Summary» - it doesn't saving - and in the «model» map only one file (SAEHD_defaul_options). Any suggestions?
I used 3000_series (GeForce RTX 3080)
Try to increase paging file size.
@@deepfakeengita2967 Thanx, I will try.
@@deepfakeengita2967 How to do that?
@@pittstone Windows + R to bring up run, type "sysdm.cpl", Advanced tab > under Performance click Settings > Advanced tab > you'll see Virtual Memory section to adjust at bottom.
Future actors will have enough childhood footage on the internet that they'll be able to deepfake their childhood faces on childhood actors for flashback scenes.
Hello !!!
How long for 174000 frames with your config ?
Only 21000 with my RTX 3070 since 1Hour..
Tks and very good job !
Do u mean to extract aligned image?
That's a WHOLE movie? Damn!
thanks
subscribed you🙌🙌
Thank you for the video... but I found difficult to read what options that you are using due the font size, too little... Do you have a blog or website with more info, just asking?
1) Full Screen
2) Play in Slow Motion
3) Zoom = Ctrl + Mouse Wheel Up or Down.
4) Or screenshot and then zoom on the screenshot.
5) He actually just presses ENTER all the time and then it uses the default value. Maybe 1 or 2 times where you manually have to write "0".
4 not working. Not finding any faces
When i run train SAEHD, after it loads samples, initializes models etc. it says "Press any key to continue" and it would just sit there and do nothing.
I already incresed the page file size to 16gb, I have 16gb physical ram and a 2070 Super on Windows 11
Same, only works with CPU. I don't know if it's because of the CUDA version
Had the same problem, changed build from DeepFaceLab_DirectX12 to DeepFaceLab_NVIDIA_up_to_RTX2080Ti solved problem
Can anybody tell me how to use a pre-trained model?
Can you stop and resume training or is training one long wait? Lets say i did 10,000 iterations and stopped, can i add to that at a later time?
you can stop it and resume everytime u want by just launching the train bat again and selecting your model
I accidentaly close the training preview window, how to open it again
What if you can't automatically or manually extract the majority of faces? Do you have to go with XSeg and draw polygons around them?
How did you get your keyboard on your screen like I saw instead of hitting tab?
Great vid. What track/ artist is that first (song)?
When i run 6) train SAEHD it often just sits there after loading samples. one time i got an error about importing numpy core. no clue why it get this message. any suggestions i am running up to 2080ti and i have a 2060
Try to enable paging file.
@@deepfakeengita2967 windows paging file?
is this resolved? rtx 3080 ti having the same issue
resolved after customizing page size to 16-32gb
Where is the pre-trained models?
You can find the pre-trained model on the forum (link in the main page of DFL GitHub page) or even create it by yourself.
I got the error when do [train SAEHD] :ImportError: DLL load failed: The paging file is too small for this operation to complete.
from . import specfun
I have 12GB 3060 card , 16GB RAM and a 12th Gen i5-12400F
Any help pls ?
It's even written. Enable/increase your paging file.
Computing motion vectors: 0it [00:00, ?it/s]
No frames to merge in input_dir.
Done.???
You don't have any frames in dsta_dst/
awesome man can you make a tutorial on real time deep fake using using low end pc broo
I had the problem that when i was running train SAEHD and after loading samples, initializing models etc. it says : press any key and it would do nothing
Try to increase paging file size
@@deepfakeengita2967 I also serched a little bit and for me it worked to make a quick Quick96 train and the SAEHD train Worked
having the same issue, have you by any chance found a solution for this?
Failed to allocate 2764800 bytes in function 'cv::OutOfMemoryError'
You prolly finished available RAM
when i try to train SAEHD, it says internal compiler error, idk what to do
Google.
hi, how do you fix an alignment that looked weird because of face obstruction like "mic" obstructing the chin, making the aligned face look big
The generic XSeg model should be enough, otherwise you can use the XSeg editor.
@@deepfakeengita2967 im really new to facelab, thank you for answering my question.
how do you do head replacement including hair?
Hi! You have to use an head model, and manual XSegging the aligned images, creating masks that include also hair and ears
whats a head model?@@deepfakeengita2967
If I download the pre-trained xseg, where should I put it and how to use it?
Put it inside model folder, and start XSeg training
After applying xseg, am I supposed to use SAEHD? Actually I wanna use Quick96
Yes
Where can i download the mp4 files?
06:21 this is moment where I press enter to continue and nothing happens :///
did u solve it?
@@asenjogonsa3868 no sorry 😢
I didn't find those .bat files like you are using ! Where to find those? Please help
Prolly you didn't download the package from github. Go here: mega.nz/folder/Po0nGQrA#dbbttiNWojCt8jzD4xYaPw or here: disk.yandex.ru/d/7i5XTKIKVg5UUg/DeepFaceLab
i5 9th gen
Nvidia 1650
8gb ram
will it work for me?
yup
the merger 10:48 doesnt come out
help
Maybe you didn't choose interactive merger.
@@deepfakeengita2967 i have same problem
help pls come not out
hello, is it important to use xseg? what is the benefits to use it?
Yes it is, cause you can create custom mask based on situation. Moreover, if you don't create segmentation masks, it will be used the mask created from the landmarks, which is the area inside the facial landmarks. This area doesn't cover all the face, only the zone between chin and eyebrows, so you can't use a whole face model for example. Instead for head model, xseg is mandatory.
your GPU is GTX-1060, and you have downloaded DF version for RTX-2080? Will DeepFaceLab 2.0 for 2080 work for GTX-1060Ti?
No I downloaded version UP TO 2080ti. It means ALL the video card of nvidia chronologically released before the series RTX 3000.
I have no idea how to even install it. can you help?
Watch my old videos.
Hi, would you be willing to take deepfake commissions? I'm making a parody short film next summer that needs deepfakes scattered throughout small parts in the film and I havent gotten any deepfake artists to respond.
Tell me more in pvt on telegram or discord. For discord search me in dfl official server while for telegram @DeepFakeEngIta
Privet
Hallo
Asus
Core i5 2.53GHz
Ram 6gb - rom 600gb
Nvidia GT 325M 1GB
My computer specs, can deepfakes work with me?
No.
@@deepfakeengita2967GPU: GeForce Nvidia 710 GT
I'm sure the processing part has been sped up by quite a bit. I was trying to run it step by step on a remote datacenter compute instance with a nVidia A40 (48g Gb) vis-a-vis the shitty GTX 1060 here and the processing time for extraction part was atleast 5x compared to the video here.
Of course it's quickened, it's not a trick to hide something and it's all out in the open. The number of iterations per second is indicated in the progress bar.
Here is a slightly better tutorial, where the person talks you throught it, telling what settings can be left on defaults: th-cam.com/video/QSmHho1uHFM/w-d-xo.html
This wouldn't work on a GTX1080 would it?
It works
@@deepfakeengita2967 can do on 1050ti and just use small batches?
How long does training the AI take?
There is not a time. Depends on when you're satisfied of the results
I wanted to play with this but does the source data need to be video or can it be just series of photos (since photos are higher resolution and can provide more details)?
You can put images in data_src folder and after extract faces but must be a lot
Sorry but that's not how things work. When I was first starting out deepfaking, I thought "Higher resolution= better results". First thing you will realize is your GPU will determine ultimately determine the output resolution of your faces. For 4GB GPU, you will be working with 128x128 pixels to 256X256 pixels. So even if you have a 8K ultra sharp image, it will be converted to a tiny image resolution. Once your GPU is full, the program will give you and out of memory error. The program generates images by compiling multiple angles of a subject. That is MOST Important part. Think of it like paper mache. You are taking pictures of different angles and sticking it on a virtual 3D model so the Ai can map it onto a target face. Typically you need around 2000 to 3000 images to do a decent faceset. This is why video interviews are best. You can get away with 500 to 1000 images (different face angles and expressions) but the results will suck. I typically do 4,000 to 6,000 images and get very good results at 256X 256 pixels that I later upscale to 512x512 pixels. If you have a 6GB GPU you can do 512X512 pixels and double that through AI upscaling to native HD resolutions. Remember that you are not generating an entire frame (ei 1920x1080 full HD frame) but only a face within that frame that takes up only a 1/8, to 1/2 screen. Hopefully all this information helps.
@@SSingh-nr8qz What do you use to upscale the image?
where download data_dst and data_src?
You need to create them my friend
your data_src and data_dst is too simple, you should give an example of hand obstruction or other obstruction 🙃
This is just a workflow example. For more specific things watch other tutorials.
It would be nice to see this done on an Apple Silicon machine.
The best hardware for this software is nvidia gpus
MAC OS Forver ! @@deepfakeengita2967
how do i download it ?
Github link
It works today December 13, 2021. But you need some brain and experience with this kind of stuff to make it work. Likely because nobody wants 15-year-old kids to abuse powerful software like this.
I Know Why Everyone Came Here😒😎
clone voice ou no
Nope
a very complex workflow and a horrible result :D
The result is not important in this case.
which folder should I put the "RTM WF faceset/aligned/faceset.pak"
workspace/data_dst/aligned
When I ran "train SAEHD.bat", I kept getting this error message: module 'tensorflow._api.v1.math' has no attribute 'divide_no_nan'. Do you have any idea how to fix this?
Are you using right DFL version? Which video card do you have?
Ciao! ti posso parlare su discord?? mi servirebbe una mano , grazie