Easy Deepfake Tutorial: DeepFaceLab 2.0 Quick96

[@Deepfakery]

📌 DeepFaceLab 2.0 Faceset Extract Tutorial: kzbin.info/www/bejne/p2WXfYOvnMmArrc 📌

[@MS-gu3gj]

Hello please I'm having problem with training How do I train?

[@UandMeDasame]

is there an option where I can extract images from a jpeg or png for my src?

[@Deepfakery]

See this deepfake at 1 Million Iterations! - kzbin.info/www/bejne/op-4k3iGe9GagaM DeepFaceLab 2.0 Quick96 Deepfake Video Example

[@Deepfakery]

UPDATE: DFL Installation Tutorial 2021 - kzbin.info/www/bejne/boicpqhjpbuCf6c

[@Akotski-ys9rr]

is there really much of a difference? i dont think so

[@poikazal]

This is probably the best walkthrough explanation on deepfakes I've ever seen. Quick question, how long does it take to merge? Will it depend on the type of PC your using?

[@Deepfakery]

Merging should go pretty quickly, at least a few FPS. Adding extra options like color transfer and super resolution will slow it down, sometimes dramatically. I believe the process is highly CPU.

[@ruski171]

@@amj76425 shush

[@MMAUniversityTime]

Don't understand why other videos have such a hard time explaining the merge part, Thank you!

[@CerebralRiches]

right? now i understood it at once. others are making my brain commit suicide.

[@Deepfakery]

I don't get it either. They just skip right over really important and basic stuff while adding in all sorts of useless info. Its like they expect you to know how to do it before even watching the tutorial.

[@LexlutherVII]

you described the school education system in one comment!!!😂

[@LexlutherVII]

@@Deepfakery They are naive "ASF" "SMH"wanna make a simple thing into rocket science!!!🤢

[@LuigiLuppo2021]

Excellent... I followed the tutorial faithfully and got the same result as you ... very cool ... really amazing ... thanks!

[@Deepfakery]

Great to hear!

[@mandiraganguli7814]

@@Deepfakery Is this software free of cost?? or I have to pay?? Awaiting your reply..

[@JohnWhite-iz4nl]

@@mandiraganguli7814 its free

[@dreamsanimationstechnology3363]

hi... how can i download soft

[@The404Studios]

This is dangerous, and super freaking realistic what the hell.

[@HuyHoang-gq4kz]

why this is dangerous

[@JoeMitchell868]

@@HuyHoang-gq4kz In the wrong hands of somebody with bad intentions

[@hayleeeeeee]

Not dangerous at all as wont work for most people, takes to long and the results are very poor to be honest, need professional software and knowledge to pull off anything remotely good really.

[@hohohappyday77]

​@@HuyHoang-gq4kz Because the tool will waste your time.

[@K3v114]

@@hayleeeeeee dude, in a matter of a year it went from needing thousands of high quality pictures of a person in all different kinds of lighting and days of training to a single video and 1 hour of training. Soon it will be one picture of subject A to perfectly replace the face in video B in 1 minutes. Then we’ll have audio deepfakes and bam.

[@binbinn28]

I'm stucked at the "Starting, Press "Enter" to stop training and save model" The trainer doesn't show up.

[@Deepfakery]

What is your GPU?

[@electricanimation3379]

@@Deepfakery I have a 2080 super and it won't work I'm stuck at that point

[@AppyTheApe]

@@electricanimation3379 will this work properly on an integrated graphic card? I have Radeon Vega 8.

[@keyopipipi4298]

@@AppyTheApe work for radeon vega 8?

[@AppyTheApe]

@@keyopipipi4298 I asked the same buddy

[@Deepfakery]

*IMPORTANT: CHOOSING THE RIGHT BUILD* Download Here: mega.nz/folder/Po0nGQrA#dbbttiNWojCt8jzD4xYaPw DeepFaceLab is designed to run on Windows 10 and Linux. DFL 2.0 NVIDIA RTX3000 series build - NVIDIA 3000 series GPU required. DFL 2.0 NVIDIA up to RTX2080Ti build - NVIDIA GPU with CUDA 3.5, 5.0, 6.0, 7.0, 7.5, 8.0 and higher. Check your GPU here: developer.nvidia.com/cuda-gpus CPU training requires AVX instruction set. DFL 2.0 DirectX12 build - AMD, Intel, and NVIDIA devices with DirectX12. DFL 1.0 OpenCL build - Devices supporting OpenCL. This version is no longer maintained.

[@wyngreece7]

for nvidia GTX super 1660?

[@CerebralRiches]

This worked well for my 1660 ti. It took 1 hour for 35k Iter's. I'm also happy with the result. Thanks for the tutorial!

[@choksm9079]

whats your cpu and ram?

[@parishna4882]

@@choksm9079 lol why? The gpu does the work.

[@IzludeTingel]

12 hours for me to reach 49k iterations lol, 9900k underclocked to 2.8 (speed steps to 3.6)

[@karenslivesmatter2186]

Guys how do you not keep pressing p

[@karenslivesmatter2186]

Nvm I asked a dumb question I just realized they keep running and p is just for preview

[@SkyShazad]

This is the BEST straight Forward Video out there for learning this ... THANK YOU

[@I77AGIC]

this is perfect for first timers. thanks for *acutally* keeping it short and sweet

[@toxicshot4528]

short and straight to the point, thanks!

[@worldmasterpiececartoonsli133]

Hi there! I really enjoyed the video but I came across a little issue... I get the following error when doing the step at 3:09: Error: No training data provided. Did I miss a step? Please could you help me? Thanks.

[@Deepfakery]

Make sure you extracted the images and then the facesets from both videos.

[@worldmasterpiececartoonsli133]

@@Deepfakery Okay thanks!

[@XMrRex]

Wow really good and straight to the point. Thanks for the video man I've been interested in this for a while. Is there some sort of guidelines, for example if I want to attach my own face to a video and I'm recording a source vid. What would be the fastest way to capture all my face angles and have the source be as small as possible? Just look around in all directions?

[@Deepfakery]

That sounds like a good start for a general faceset, but you also want many facial expressions, so maybe try reciting something during the recording or have someone interview you. If you’re doing a specific video then anything you can do to match the angles and lighting will be a great help. You can delete unnecessary src images during the View Facesets step.

[@1GTX1]

I was about to give up but than i used the rtx2080ti zip build for my 1660 super and it works, thanks for the video!

[@Deepfakery]

Yeah there’s 2 different builds now

[@Luciferaorg]

when I click on train Quick96, it iniciate models at 100%, then loading samples at 100%, then the second loading samples do not start, I have a Rizen Tyhreadripper 1920 with 12 cores and 24 threads and an Nvidia GeForce RTX 2070, do you think the problem is because I am using feepface lab from an external hard disc? I will copy it to my internal hard disc and see what happen

[@teriosshadow17]

Here is the issue i have with it. It says error "Failed to load native Tensorflow run time" What am i doing wrong?

[@Deepfakery]

Could be a file writing error. Try unzipping the package again. Sometimes one of the little python files gets corrupted during extraction.

[@Gaurav-by5bb]

Wow man u are replying to each comments hatsoff bro 👍🏻

[@Deepfakery]

I'm trying to help all the people who are in the same situation I was. It took a long time to figure out how all this works and I'm still learning each day.

[@vodkavodka8903]

@@Deepfakery thank you sir

[@AlexDelgado328]

Time To make a dame da ne meme

[@xvx_j_xvx5255]

You don't need to use this just search up dame da ne tutorial

[@everettsalmans104]

Just use python and Kawping

[@Deepfakery]

They're right, it's a different software called First Order Motion. You can make a kind of Dame Da Ne meme using DeepFaceLab but it would look weird. Best bet is to use the other method...

[@parishna4882]

honestly, that janky method is better left out of the deep fake side of things. More like the obviously bad fake to annoy your friends with, side of things... lol

[@methadonmanfred2787]

@@parishna4882 quality of the result depends on the user, you have to know how to use it properly, then you can get some damn good looking deepfakes

[@hagemaru96]

i get phyton has stopped working during training

[@danielwillems9114]

Thanks for the detailed instructions. Approximately how long should one expect this process to take from start to finish?

[@pasatorman8294]

what do "f", "wf" and "head" mean in the faceset extraction? what changes depending on what you choose?

[@Deepfakery]

They are different sizes/areas of the face that can be trained. Check out my faceset extraction tutorial for a full explanation

[@fakeshemptaboo]

@1:42 all I get is “failed to load the native tensorflow runtime”… how do I fix this?

[@KaelumKrispr]

Having issues with the samples loading part in the training, just stops after all the samples load then says press any key to continue, which causes nothing to happen

[@FJPkagami]

i have the same problem too

[@kalevteener3380]

same

[@paulgeorge9228]

are you using the right build for your nvidia or amd card?

[@MCMIXING]

@@paulgeorge9228 what build do i need for 7900xtx? thanks

[@yasayanpan]

Hey man, this is a great video! I just have one question. After I did everything in the video the result I got was just a still image. What can I do to fix this?

[@Deepfakery]

Did you do both final steps of the merger, apply settings to all frames (shift + /) then render all frames (shift + >)?

[@yasayanpan]

@@Deepfakery Thank you so much for answering, I must have missed those steps.

[@charan1969]

@@Deepfakery Faceset extraction is not working properly on command prompt... Extracting faces ... Error.. I didn't get any options to type CPU or GPU . . . Unable to start subprocess . Press any key to continue 😞

[@diegorobles100able]

@@Deepfakery when im merging im not getting any response when the merger keys pop up, i cant set any of the settings

[@lewisjaygomes1617]

@@Deepfakery i did this but the merging doesnt happen. it never starts. what is going on?

[@ym17191]

amazing, clear explanation! worked like a charm on my gtx 1080 and 2700x

[@Deepfakery]

Excellent! You should be able to use the SAEHD trainer with that setup. I'm working on a tutorial but there's alot more details to cover.

[@worldmasterpiececartoonsli133]

Hi again. I'm almost at 2000 iterations in my training but I still have just weird pixels next to the faces. There is no progress. Is this to do with my drivers?

[@Deepfakery]

2000 iterations isn't much when using Quick96. Let it go overnight at least. If you're just seeing weird colors and no shape face its possible you're using the wrong build for your card.

[@kimotee1081]

Whenever I start training quick96, it pops up a window saying python has stopped working. How do I fix this?

[@pockydew9467]

I tried training quick96 with my GPU(RTX3060) but it stops at "Press any key to continue". I had to use my igpu(AMD Radeon Graphic) to render but do you have any solutions to fix this?

[@ChristopherMikrowelle]

On Step 5 when i do the training it, my training preview window doesn’t open. Please help me I’ve been looking for the solution and for help everywhere and noone can help me:/

[@SlimeCore_]

if you still have the problem just close it and start again it should open after some tries

[@exZact]

If you remove upside down photos of face in the dst will it cause the mask to cut out in the video? Or getting rid of unwanted faces? I know it won't affect src but does it affect dst?

[@Deepfakery]

Upside down photos are usually a result of false face detection. You will need to remove them and they will be skipped in the merger. There can be more than one face per person sometimes, so those upside down phots might have good ones in the previous or next file. Check the filenames for possible duplicates. If you have to remove those faces then you can either fix it in post by duplicating a nearby frame, or skip that section of the video altogether.

[@anonymousviper5075]

hey! just wanted to ask can i select source(data_scr) as image and destination(data_dst) as video and is the graphics card important for it

[@Deepfakery]

If you want to do a single image (like the Baka Mitai meme) look for First Order Motion Model.

[@ea-fc-mobile-goals]

5:07 how many times do we have to press the shift + / key? just once or till the frames/video ends?

[@Deepfakery]

Just once, it will apply the settings from the current frame to the remaining frames, then shift + > will process the frames

[@ea-fc-mobile-goals]

@@Deepfakery thanks man. With your tutorial I was able to make a pretty good deepfake. Need to start learning the adv settings now.

[@ChuckstaGaming]

OMG, this is gonna be fun! Gonna stick myself in my favourite shows and film 😂🤣

[@armitx9]

🤣😂

[@XXLV-]

Hahahah

[@brockphillips6411]

Can you use a picture and insert that face into the video or do you need two videoos?

[@amoghvarshhattigoudar5832]

It takes more than a day just too create 10min video

[@markkocsicska2590]

@@brockphillips6411 you need 2 videos, but any free video editor software can put pictures after each other using a user friendly interface and extract it as mp4

[@mapiles26]

Hi just a question how you got so many it/s per second mines running at 1.20 is it depends on the hard ware? or the software i downloaded, i use the DX12 something

[@Deepfakery]

I have 1080ti's so I use the NVIDIA version. It is faster.

[@bigbanana531]

So I’m having a problem with step 4 when I go to check the faceset extracts from data_dst and data_src it doesn’t show me anything it just pops up with some sort of browser and and only shows images extracted from the video and they aren’t aligned I followed all the step correctly do you have a solution?

[@Deepfakery]

You should be seeing a file browser with square images of all the faces that have been pulled. Those are the aligned faces, meaning they’ve been rotated upright and cropped to a square. Are you seeing something else?

[@bigbanana531]

Deepfakery I solved the problem

[@bigbanana531]

I ran into another problem, when I go to convert the face to the video after I finished training it says no faces found for 00001.png, copying without faces did I miss a step after the training

[@Deepfakery]

Did you run the interactive merger? You need to run Merge Quick96 then Merge to mp4

[@bigbanana531]

Deepfakery I’m using an older version of the software for AMD.

[@IEnjoyDarkSouls2PVP]

when i get to the merger part of the vid the face is nowhere near as clear as urs mine is really blurry did i do something wrong maybe

[@ronanm4418]

followed the tutorial and got a pretty good result, it was actually pretty easy!

[@kaoe145]

what specs do you have and how long did it take

[@ronanm4418]

@@kaoe145 I have a nvidia 1060 6GB max q, I only did about 3000 iterations and it took probably less than an hour. Let me know if you want to know the other specs but I'm pretty sure gpu is the most important

[@kaoe145]

@@ronanm4418 thanks for replying I have a gtx 970 I been training a 5 min video it’s been running for 6 days the preview is still Blurry how long was your video

[@ronanm4418]

@@kaoe145 I used the default one that came with it, the elon musk/tony stark one. Mine wasn't perfect but I didn't leave it running that long

[@MrEmotional33]

May i ask which cpu you have?

[@notallpolitical]

I am running with a 4080 super on the 30 series executable as thats the most modern, however, my iterations per second is only 1.46it/s when a 1080ti in the video is getting 3.71it/s , how do I increase the iterations, my GPU isnt being utilised at all, thanks

[@se960]

Anything more recent thats better than this for deepfakes?

[@SongStudios]

The same software has a more capable model called SAEHD Quick96 is more so meant for tests or fun. Warning: I recommend you to learn everything you can about deepfakes to get good results. You ***NEED*** a powerful computer, more importantly, a powerful GPU If you want to go for highly realistic models, I cannot recommend anything less than a 3090

[@BeyondAnabolics]

@@SongStudios how can you use saehd? Do i have to press "train saehd" instead of "train quickxxx" following the same steps as this tutorial, or there are some other steps to do?

[@matu4pc488]

I have gtx 1070 and me didnt open training preview (i must go with cpu but I do not want go with cpu) 3:15

[@intheworld9578]

I have a problem. when i merge it just shows me the video with the original face without the face that I chose

[@anc5275]

Me too 😢

[@TREXHUNTERX]

have you solve this problem?

[@a846810]

CATFISHING TO A WHOLE NEW LEVEL

[@SECourses]

would these pop up windows work on ubuntu linux desktop? those GUI windows?

[@BnymnSntrkNu]

best deepfacelab tutorial without a doubt. thanks man!

[@Deepfakery]

Thanks for the support! Try the SAEHD model next, its a similar process and you can even use the same files from this tutorial. Only thing I would mention is turn off 'flip faces randomly', and use random warp for a good while then turn it off. There's alot more options to SAEHD but it doesn't have to be complicated either.

[@fakerdeeper9329]

awesome tutorial thank you. just started uploading my own deepfakes today... hopefully I can get better and better

[@abrahamnunez9761]

i followed the instructions but when i play the merged video, keep feezing the video in 3 secs while audio is running up to end of video. help me

[@Deepfakery]

Did you do both final steps of the merger, apply settings to all frames (shift + /) then render all frames (shift + >)?

[@abrahamnunez9761]

@@Deepfakery i did that right now, still get only 3 secs video. i will try now more frames

[@SuperFinGuy]

@@abrahamnunez9761 In my case the interactive merger is not working properly, I use it only to play around with the settings, after that I rerun the merger and just say no to the interactive merger, and everything goes smoothly.

[@hans983]

I also only get 3 seconds. Could you solve it?

[@stoney819]

can I re use the source for multiple destinations? or do I have to clear the workspace and re train every time

[@onkarjadhav38]

I have gtx710 and 24 gb of ram. Which one would be better to use?

[@Deepfakery]

Not sure what you’re asking. Your 710 has 2gb of vram correct? You should be able to run Quick96 on that, but 4gb is recommended for the SAEHD model

[@onkarjadhav38]

@@Deepfakery I mean to say, Should I use CPU or GPU?

[@Deepfakery]

I think you will have to use CPU. I have just checked that DeepFaceLab requires your GPU to have CUDA Compatibility 3.0 and above. Yours appears to be version 2.1

[@onkarjadhav38]

@@Deepfakery Yes! So is the process same for CPU?

[@Domi-lb2pj]

Dude wtf why would you have so much ram and that gpu

[@BhavyaGandhiViolin]

heyy! i am getting the error in the first step itself where it says, the system cannot find the specified path. ' "" ' is not a recognized as an internal or external command

[@kalel_with_a_hyphen4812]

did you fix it?

[@Eclair_Visual]

Fatal Python error: Py_Initialize: unable to load the file system codec ModuleNotFoundError: No module named 'encodings' ??????????

[@h3xagon0001]

use 7zip to extract it

[@plixplop]

SO cool, thanks for the tutorial. Will this method handle applying the fake face onto a target face that is turning to a full-on side profile? What about being partially in shadow? Or does it need to be a pretty well-lit target face? Thanks!

[@Deepfakery]

Yes, it will work for almost any face with enough effort. However it will require using an XSeg mask to really get the tough angles and dark faces. Also a good faceset of course. Profiles are kind of difficult to do well; you can see some examples in my latest videos.

[@FemboyLucky]

Hello! I was following your tutorial and man! It's very detailed but I do have a question/problem. When using the face extraction, Some faces are either upside down, blurred, out of placement or not even there. Do I just delete it and re-run the extraction or Can I just delete and continue from there?

[@Deepfakery]

Check out this Faceset Extraction Tutorial - kzbin.info/www/bejne/p2WXfYOvnMmArrc If you still have questions feel free to ask!

[@kalvinwin59]

Is there any way to streamline the deepfake ? I just want to be able to pass in a video and have it go through and choose a face from a dataset. I would like to be able to pass in a meeting with 3/4 individuals and have the deepfake be generated. Is this possible ?

[@JustAGuyProduction]

First of all, great video. You're a great teacher. Second, I know deepfake was invented for creating "adult" content and I shouldn't do it, but I can't resist the urge to put some of my favorite celebrities into my favorite "adult" videos.

[@skymoore3177]

Username checks out

[@_Looft]

bruhh what I would've never even thought of that. Seems like a lot of effort just for a nut lmao

[@Akotski-ys9rr]

@@_Looft lmao thats kinda weird

[@elsaylli2161]

weird ahh

[@seanplew6548]

lol u didnt need to say this out loud my guy

[@felixbo]

Thanks for the great explanation. Everything worked fine so far. The problem is, i am stuck at the merging process. It tells me merging is 100% but the lines "Session is saved to model/xy.dat. Done. Press any key to continue" (5:17) are not showing up. It doesent save the session, there is no file inside of the folder. I waited for half an hour after it was 100% but seems like its stuck forever. Tried a second time, same problem. Do you have any idea what could be causing the problem and may you help me?

[@felixbo]

Hey i "solved" the problem by ignoring the missing .dat file and converting the merged frames. I didn't know these were created.

[@Deepfakery]

Yeah once it reaches 100 and the hourglass goes away it should be done. As you've probably seen the files are in data_dst/merged

[@boondox270]

I love the Arnold Schwarzenegger Sylvester Stallone deepfakes. I wish someone would do John Candy and Chris Farley!

[@deeber35]

If I have a frame where the face is at a weird ngle, do I delete in Aligned Results, will I still get to manully extract it? Not sure what deleting a frame means for masking in that frame.

[@lendur2743]

Never thought deepfaking would be so easy

[@JainamSutaria777]

Exactly. it's not an easy

[@nukmuk]

@@JainamSutaria777 the video is literally 6 minutes long soo .......

[@JainamSutaria777]

@@nukmuk Yeah! 🌝

[@jj987987987]

You think its easy because you dont know why it works. You are just following a video and pressing buttons. Someone already did the hard part for you.

[@lendur2743]

@@jj987987987 yea, thats the point of programs? they do things for you so you can use them

[@iamdeveloper-xt3it]

hello i like make my own dfm model with SAEHD. i tried based on Instruction but its failed. Can you make explain how to do that. i Want make new dfm model which will work for all relatable face in Deepfacelive like Pretrained Tom cruise dfm model

[@GROTTEZ]

My fantasy it’s gone be real 😎

[@fuzsd1590]

@Paul Mathews bro???

[@bigdj0ntz8]

hello, follow these steps but when try to train the mdel with quick96 it got stuck at "Initializing models: 100%", have a 3060ti, tried with cpu and got the same issue, know how to proceed?

[@b0chen]

lol I mean Elon is basically irl Tony Stark

[@Odolwa2]

does 2060 work? if so, any tutorials how to fully utilize it? else, the one i downloaded from deep face lab's git doesn't recognize it and goes to quick96 :( i don't want to use the CPU...

[@mrfakeybr2510]

Is the result of the final video with just 100,000 interactions? OMG!

[@Deepfakery]

Yes 100k, about 2 hrs time. I have an example of the same video at 1M iterations for comparison

[@gustavkirchoff4633]

yeah just did this 95k interactions looks good, im on a 1050ti btw

[@diab2901]

@@gustavkirchoff4633 I am too but im to impatient to wait expessially if its just a test to learn how to use the program

[@Monero_Monello]

@@gustavkirchoff4633 Do I have to hold down P or is there a way to make it automatic? Or can I just leave the program, come back in two hours, press P and I'll get 90k?

[@Monero_Monello]

@@gustavkirchoff4633 Also, how do you get 90k? I get 50/min (I have a geforce GTX 1050)

[@paulgeorge9228]

i got 603037 iterations and some of my result previews are still blurry, do I have too much src/dst (9 k and 30 respectively), do i need more iterations, better face set, or fix some settings?

[@hans6300]

thanks for the free upvotes man

[@sickojoseph9377]

Haha reddit go brrrrr

[@artfood6508]

Some NVIDIA GPUs, even among the newest ones (but for notebooks), will not do any of the processes. You also can't activate hardware scheduling for these types of GPUs on Windows 10. Because of this I had to do everything with CPU only, slow but worked, thanks a lot.

[@Deepfakery]

It seems like a lot of their notebook GPUs do not support the CUDA Compute version needed for DeepFaceLab, even if they have enough VRAM and other resources

[@Pikelol]

Kind of technology which definitely won't be used for porn.

[@potato22258]

You're way late to the party, my man.

[@hassassinator8858]

@@potato22258 Some of the stuff out there is so realistic it's downright terrifying

[@yeetogami2575]

I can't extract the facesets. If i extraxt even with default settings(src or dst), it extracts just 1 image and then shows a bunch of errors. What do I do?

[@kavya1638]

i'm using this on myself lol, it's just funny to put random stuff on my face. i wonder, can i put my cat's face on my body?

[@Deepfakery]

You can try the cat's face using the manual extractor but you'd probably have better luck with EbSynth, maybe...

[@charan1969]

@@Deepfakery @Deepfakery Faceset extraction is not working properly on command prompt... Extracting faces ... Error.. I didn't get any options to type CPU or GPU . . . Unable to start subprocess . Press any key to continue 😞

[@paulgeorge9228]

after training the xseg do i still need to do saehd training (i already have 1.4 mill iterations byt without xseg) or can i go straight to merging the saehd?

[@Deepfakery]

Well you should really apply to the mask then train the deepfake. Technically you can go straight to merge and use the mask, but the full area of the face won’t be properly trained.

[@paulgeorge9228]

@@Deepfakery about how many more iterations with training do i need, now i have the xseg mask trained? is 30 k enough?

[@Raniaska0306]

"More advanced projects"... hehe

[@randomguy-jo1vq]

I see you are a person of culture as well

[@arnavtete7793]

*insert Lenny face here

[@FSXNOOB]

Wich text 2 speech do you use for this movie?

[@Deepfakery]

Nuance

[@speedhead]

dawg - you just allowed me to have more goddamn fun THANK YOU

[@hackmedia7755]

how long it usually take to train from scratch? should I just load a pretrained model?

[@shrusher2234]

нихуя не понял, но очень интересно

[@jarnekumbruck3843]

hey this is a good video But i have a question. How long should you train him? and you can, for example, let him train too long so that it no longer works.

[@Deepfakery]

The longer you can train the better it will be. I have an example of this video after 1 million iterations

[@Factbitesx]

at train Quick96 I have a issue can you help me? ( Memory error: Unable to allocate 3.00 MiB for an array with shape (512, 512, 3) and data type float32) my pc has gtx 1650ti ryzen 4600h 8gb ram

[@Factbitesx]

trainer fails every time :(

[@knightshadow393]

I have a 2070 super and amd 3600 but when i try to do training with quick96 or the other 2 it won't run on my gpu. After loading all images it just says press any key and doesn't do anything. Cpu works but its too slow. How do i make gpu work?

[@aag1977]

Is there a way to run multiple instances? So I can 'prep' a video while another one is training. I've noticed Face Extract won't work if there is another process running. Is there a way around this?

[@XristosGiotis]

so, if i want to make a video with photos (from the begining), i must put my images into the folder with the name data_src ???

[@Deepfakery]

Yep, just put them in the data_src folder, where the frame images would be, then extract the faces. I covered this in my faceset extraction tutorial: kzbin.info/www/bejne/p2WXfYOvnMmArrc

[@Psybernetiks]

Once my merging process is complete, the line about this session being saved & press any key doesn't pops up in the prompt. Any leads on what I can do?

[@Deepfakery]

It will stay at 100% until you escape the preview window. Kinda weird it doesn’t end itself but that’s how it is

[@FatToadRecords]

On Step 6 (train Quick96), it loads the first Sample but not the second. How can I fix this?

[@paulgeorge9228]

so is saehd or quick96 better for more accurate faces? i see from this video that quick96 does a very good job already, does the saehd do an even better job? i hae nvidia gtx1060 max-q design with vr ready graphics (6 GB VRAM) and 8th gen intel core i7-8750H processor and 6 cores/12threads, up to 4.1 GHz, would saehd or quick 96 run better on my laptop? the extraction and face detection processses seems pretty slow for me compared to yours (then again i have a 20 minute video to deepfake)

[@Deepfakery]

Quick96 is basically just a proof of concept for the software that can be used to quickly test things. SAEHD offers better models and the ability to tune various dimensions and additional processing. Quick96 is locked at 96px resolution and most model settings, so its not meant for production.

[@DouglasJones-i5d]

Hi... I love this video and thank you for making this... I want to ask, is the training preview fast or slow? Because on the video it goes fast but on my pc its slow.

[@paulgeorge9228]

Can the merge work well on videos where the destination video's angle causes the face to rotate a lot (like a clock, not side to side)?

[@Deepfakery]

Yes, but the face extractor might have a problem detecting if it rotates too far.

[@torindanius9530]

i did exactly what you did.but when i was going to quick train it say no faces found.not with SAEHD too.result is same. on quick train i used video names and trained but when i am at the page of MERGER there is only original footage,and on his face there is square that's have blur white color.please help

[@Deepfakery]

First, do you have face images when you run view aligned result? Also run data_dst view aligned_debug. You will see all of your frames with landmarks from the detection. Maybe some are missing or false alignments?

[@allscars2882]

when arrive to step of merge, i dont see commands. if i click on tab i see first frame than if i click w or s or shift + > etc ... freeze all what is the error?

[@Deepfakery]

Could be a few things. First off you can reset the merger settings by deleting the merger session file or selecting not to use it when you run the merger. If your frames are very high resolution you may have trouble loading them. Also make sure you still have all the destination frames and didn't move or delete any of them.

[@amparo762]

I have exactly the same gpu (1080gtx ti) than you, but training takes a lot of time. Did I miss something? Should I install some CUDA library in order to accelerate the training process? Thanks a lot and congratulations!

[@Deepfakery]

I have 2 x 1080 Ti so with 1 it will be significantly slower. CPU may affect the speed as well; I'm using an i5-9600K at around 4.30GHz. You cannot easily change the speed of training with the Quick96 model. You can remove some data_src/aligned images to lower the overall package size. The SAEHD trainer, while a bit more complicated, allows you to tune various settings to your system.

[@RockyWild]

Thanks! Do you know why I can't train with Nvidia geforce rtx 3070-8gb? it stops and the training window doesn't appear. I can only train with CPU (AMD Ryzen 7 5800H) 😴 And the training end automatically? Or do you have to manually stop it after X hours? Cheers

[@Deepfakery]

First make sure you're using the RTX 3000 build of DFL. In Quick96 you have to end in manually. With SAEHD and others you can set an iteration number to stop at.

[@RockyWild]

@@Deepfakery Thanks. I just want to learn the basic. At the moment I only got blurred faces (I also tried SAEHD and the Eyes and Mouth priority option). I'll use Quick 96 until I get something decent. Maybe it's because I have faceset to 1024, I'll try 512 and train with to GPU. With Xseg I haven't tried it yet but I guess it's also important too Regards

[@Deepfakery]

The trainer has to resize the images to fit the model resolution. You can help it by using the image resizer to match them to the current model. Also packing the images will help with the initial loading.

[@denynugraha9099]

Bro, if the skin color is not the same as the one in the video, what settings are you in? I'm using the SAE model with 300000 epoch iterations, please enlighten me bro🙏🙏

[@Deepfakery]

SAE model has more steps to do. I’m getting to that part in my tutorials but here’s some general advice: (1) pretrain model. (2) disable pretrain and enable random warp. (2.1 optional) enable learning rate dropout. (3) disable RW and LRD. (3.1 optional) enable LRD (4 optional) enable gan.

[@denynugraha9099]

@@Deepfakery i use dfl google colab bro

[@moldorm99]

After I launch the training file and it initializes models and loads samples, I hit any key to continue, and nothing happens. You didn't say what this could be caused by or how to fix.

[@ViolinVoid]

hi sir! great tutorial! question: when im doing step 5: training, if i press enter and save my progress... how do i restart progress on the same file later?

[@Deepfakery]

Just run the trainer again. You will be prompted to choose a saved model or start a new one

[@mayurgavali9847]

Hey great tutorial!, I had one question- Suppose I have person A as the source and B as the destination in model 1, and person C as the source and D as the destination in model 2. Can I make a deepfake with A as the source and D as the destination by using my pre-trained model 1?

[@Deepfakery]

You could do that to jumpstart the training, but if B is drastically different from D then you should just start from scratch.

[@mayurgavali9847]

@@Deepfakery Thanks!!!

[@deeber35]

Somehow I'm at the point where images in the debug folder show the proper masks {in green} from manual extraction, but when I merge, those masks are not applied, and so the output is an uncorrected image. I'd rather not re-do all that manual extraction {there are 100s}. Any idea on what to do? Thanks.

[@Deepfakery]

Any face you extract will have a default mask as shown in the debug image. When you train the deepfake it also trains the mask. In the merger you should have mask options such as DST (original) and Learned PRD / DST (learned during training). If you don't have XSeg masks then don't choose any of the modes mentioning XSeg. Were you able to train the model and select one of these mask modes in the merger?

[@deeber35]

Do I hqave to do step 6, train, if I already have 200,000 iterations with the source images?

[@Vagaberry]

Disregard my previous question, I understand now but if I wanted to use pics as a face swap into a video instead of a video can I upload a jpeg instead of mp4 file for data_src? While that mess the script up? What should I do? And can I upload multiple face pics for it to generate a good one for a swap?

[@Deepfakery]

You can bypass the video extraction and place the images directly into the data_src and/or data_dst folder. After that run the face extraction as normal.

[@Vagaberry]

@@Deepfakery okay I’ve done that but it’s not that great, I used about 10 different images of myself and 450k iter’s and it’s not that great. Should I have used only one image?

[@Deepfakery]

Sounds like you're trying to put a deepfake on just a few images of yourself? First off, Quick96 is slow, so you'll want to try SAEHD. Having few DST images is ok but sometimes more images will help the model generalize the face better. Also make sure to remove any angles from SRC that aren't found in the DST images. Again don't expect alot from Quick96 since you can't change the settings.

[@DanielGroff]

Great tutorial, thank you. I have a problem because the video "data_dst" contains 2 faces, how can I process only one? THANKS

[@Deepfakery]

You need to remove the extra faces. Process is the same for SRC and DST: www.deepfakevfx.com/guides/deepfacelab-2-0-guide/#step-4-2-source-faceset-sorting-cleanup

[@nelsonjoseph3673]

ok, that was cool. So we only need to rename the images to this format if we are using custom data for the process. Is there any specific size ratio in which the data should be made?

[@paulgeorge9228]

i dont think so, i took a look at the code and i think it automatically converts the pixels and size ratio for you. the video sources need to be mp4 or something as the video formats though