With one no.. BUT you can still do it with a couple of images (5 or more) using this method (Xseg and saehd). Here is a tutorial that explains it: ruclips.net/video/e-k-k67gf1o/видео.html Hope this helps ;)
what is pretrained set ? the process we repeat many times with XSeg editor , mask and XSeg data_src/dst mask-apply, didn't create a pretrained set to run trainSAEHD ?
The pretrained set is used to speed up training your model to transfer face A to face B. There are no pretrained models for masking yet as far as I know. The XSeg is used to make a mask for your video, a pre-step before the actual training.
Why do I get "No saved models found. Enter a name of a new model" on train SAEHD? I watched your video and followed it the same way, but it pops up like that! Also, the question window that pops up in the middle is different from yours!
Also, sorry for the multiple posts. but if im using my own videos, what's the best codec to use for the mp4? and do I blast a TIFF sequence with keyframe every frame? I know different codecs can butcher color shifts etc. I can find no information on what to render out my own videos with. im assuming image sequences are needed
why batch size 4 for xseg training? what's the batch size and can I go higher? I know it gobbles up VRAM but I have 24gig. can I up the batch size? is that the amount of images processed in one chunk? thanks
Pardon the pun but my head literally falls off when i hear about pre trained models. Still have no clue what they are? Benifits? Where to place them. Dang. No tut or walkthrough on the interwebs!!!
Pre trained Models are Models that where already Trained with random faces so you are saving time and your ai doesnt have to learn how to Draw simple faces
save a lot of tutorials.... i place these under MUTEorials.... This one feels ok cos process is slow. But some of the more complrx 3d packages are useless.
Make sure you executed step 7 (merge quick96 or saehd depending on which you used). I also sometimes found that either the lossless verson or the regular version seemed to work while the other didnt.
@@tirr0j i done a test on the elon ~ downy default deepfake. I have a 3090. It took about 8/9 hours to do 125k... my cpu aint great its ryzen 7 3700... really happy with the speed.
@Can Vaa i watched a bunch of tutorials. I only tried it once. I dodnt like the idea of leaving my computer on at full capacity for 10 or 20 hours. It was feeling so hot. -Maybe 50k is not enough. -Make sure you got good hi quality videos for both. -Make sure your video has subject with lots of different angles (and him/her talking). -1080p size at least
As helpful as this tutorial was, it, along with every other video on mask training on YT didn't cover a question I've been looking foran answer to since I recently started dabbling with DFL. How do you create a mask of a subject with long hair ? Is it possible ? I mean after facesets are extracted and loaded into Xseg the image is too close up to mask in or out long hair. I have viewed a ton of videos on using DFL with Xseg and nowhere is it covered. To be honest if I see one more video of someone swapping Robert Downey Jr's head or face with Elon Musk's I'll headbutt my monitor. If anybody has an answer for me I'd be grateful. Somebody needs to do a complete timestamped tutorial on using Xseg, explaining all the functions and how to use them. This is probably as informative as any video I've seen on creating masks but it still left me with questions.
@@tirr0j I would love some more detail on how you setup your external gpu farm. I'm considering a similar setup. Can you give any details on how well DeepFaceLab scales across multiple gpus?
thank you for that video man, first i thought i must label all frames....
Is it possible to make a deep fake video with just one image of a person I would like to know how but haven’t found much info one the subject
With one no.. BUT you can still do it with a couple of images (5 or more) using this method (Xseg and saehd). Here is a tutorial that explains it: ruclips.net/video/e-k-k67gf1o/видео.html
Hope this helps ;)
what is pretrained set ? the process we repeat many times with XSeg editor , mask and XSeg data_src/dst mask-apply, didn't create a pretrained set to run trainSAEHD ?
The pretrained set is used to speed up training your model to transfer face A to face B. There are no pretrained models for masking yet as far as I know. The XSeg is used to make a mask for your video, a pre-step before the actual training.
Could you make a tutorial on how to use a pre-train model
just put the retrain model inside workspace/model and set pretrain mode to no
How do you xSeg a face with the hair down or something that doesn't have a clear visible line around the ears-jaw?
Why do I get "No saved models found. Enter a name of a new model" on train SAEHD? I watched your video and followed it the same way, but it pops up like that! Also, the question window that pops up in the middle is different from yours!
Make sure to extract the pretrained model files into the right folders in your workspace if you want to use them.
Also, sorry for the multiple posts. but if im using my own videos, what's the best codec to use for the mp4? and do I blast a TIFF sequence with keyframe every frame? I know different codecs can butcher color shifts etc. I can find no information on what to render out my own videos with. im assuming image sequences are needed
any special trick to catch eyes movements? i see many deepfakes always looking straight , and few with real eyeball move
When i try to merge saehd it says
no faces found for 00001.png, copying without faces
Any solutions guys?
why batch size 4 for xseg training? what's the batch size and can I go higher? I know it gobbles up VRAM but I have 24gig. can I up the batch size? is that the amount of images processed in one chunk? thanks
Pardon the pun but my head literally falls off when i hear about pre trained models. Still have no clue what they are? Benifits? Where to place them. Dang. No tut or walkthrough on the interwebs!!!
Pre trained Models are Models that where already Trained with random faces so you are saving time and your ai doesnt have to learn how to Draw simple faces
why no sound on this tutorial?
save a lot of tutorials.... i place these under MUTEorials.... This one feels ok cos process is slow. But some of the more complrx 3d packages are useless.
when i click "merge to mp4" it says reference_file not found! ... any solution?
Make sure you executed step 7 (merge quick96 or saehd depending on which you used). I also sometimes found that either the lossless verson or the regular version seemed to work while the other didnt.
I assume we need to exclude facial obstructions like hands, sunglasses, visibly thick hairs, and others?
My iterations only go up to 2. PLease help
How long did 10k iterations take and on what GPU ?
A single Tesla V100-PCIE-16GB, which had about 5k/hour with preview writing on.
@@tirr0j i done a test on the elon ~ downy default deepfake. I have a 3090. It took about 8/9 hours to do 125k...
my cpu aint great its ryzen 7 3700... really happy with the speed.
@@armondtanz I think your video was longer as usual
@Can Vaa i watched a bunch of tutorials. I only tried it once. I dodnt like the idea of leaving my computer on at full capacity for 10 or 20 hours. It was feeling so hot.
-Maybe 50k is not enough.
-Make sure you got good hi quality videos for both.
-Make sure your video has subject with lots of different angles (and him/her talking).
-1080p size at least
As helpful as this tutorial was, it, along with every other video on mask training on YT didn't cover a question I've been looking foran answer to since I recently started dabbling with DFL. How do you create a mask of a subject with long hair ? Is it possible ? I mean after facesets are extracted and loaded into Xseg the image is too close up to mask in or out long hair. I have viewed a ton of videos on using DFL with Xseg and nowhere is it covered. To be honest if I see one more video of someone swapping Robert Downey Jr's head or face with Elon Musk's I'll headbutt my monitor.
If anybody has an answer for me I'd be grateful. Somebody needs to do a complete timestamped tutorial on using Xseg, explaining all the functions and how to use them. This is probably as informative as any video I've seen on creating masks but it still left me with questions.
THANK YOU SO MUCH!!!!
What's your system spec ?
I ran everything on my zbook laptop except for the actual training, that was done on an external gpu farm on a Tesla V100-PCIE-16GB
@@tirr0j I would love some more detail on how you setup your external gpu farm. I'm considering a similar setup. Can you give any details on how well DeepFaceLab scales across multiple gpus?
i got blurry on when merge SAEHD
Verticaal filmen