I have a conversion problem

Discussions about research, Faceswapping and things that don't fit in the other categories here.


Locked
User avatar
Sconoro
Posts: 8
Joined: Mon Feb 10, 2020 12:23 pm
Contact:

I have a conversion problem

Post by Sconoro »

I have a number of questions please help.

I have a conversion problem.

I did not touch the initial settings and tried to convert after training in 42059 iterations.On preview I was shown a good option.
But when I start looking at the final version, it’s blurry What to do in this case?

Image
Image
Image

The second question is why the conversion takes place in photos and does not immediately convert to video?And accordingly, how to output directly to the video?

The third question is whether it is possible to use the tool somewhere as in the cmd shell conversion, that is, there were pressed tools that can be applied to the mask?
Image

Where are they in the program here before doing the conversion to video.

User avatar
torzdf
Posts: 2636
Joined: Fri Jul 12, 2019 12:53 am
Answers: 155
Has thanked: 128 times
Been thanked: 614 times

Re: I have a conversion problem

Post by torzdf »

Sconoro wrote: Mon Feb 10, 2020 12:44 pm

I did not touch the initial settings and tried to convert after training in 42059 iterations.On preview I was shown a good option.
But when I start looking at the final version, it’s blurry What to do in this case?

This can be for a number of reasons. The Original model (which you appear to be training on) is not particularly highly detailed so it will always be blurry to an extent.

Other reasons for blurry swaps:

  1. Poor quality training images

  2. Undertraining (i.e. you need to train more). It's been a while since I trained Original, but I'm pretty sure that 42k is a low figure, so you probably need to train more

Sconoro wrote: Mon Feb 10, 2020 12:44 pm

The second question is why the conversion takes place in photos and does not immediately convert to video?And accordingly, how to output directly to the video?

Use the FFMPEG writer plugin to convert straight to video.

Sconoro wrote: Mon Feb 10, 2020 12:44 pm

The third question is whether it is possible to use the tool somewhere as in the cmd shell conversion, that is, there were pressed tools that can be applied to the mask?

Where are they in the program here before doing the conversion to video.

That screen grab is from DFL. Use the Preview tool to tweak your convert settings (Tools tab).

My word is final

User avatar
Sconoro
Posts: 8
Joined: Mon Feb 10, 2020 12:23 pm
Contact:

Re: I have a conversion problem

Post by Sconoro »

Thanks for the answers we’ll try but there’s one question why does this guy have a couple of days and his model is ready? Here is an example video. I want to achieve the same result and with what tool this result comes from?
What do me need the computer’s characteristics for their, GeForce 1080 Ti 11GB of memory or what does the conversion time depend on and how to achieve the same result as on video?


User avatar
djandg
Posts: 43
Joined: Mon Dec 09, 2019 7:00 pm
Has thanked: 4 times
Been thanked: 2 times

Re: I have a conversion problem

Post by djandg »

42k Iters is low, more like up to 10 times that number needed.
If you have a 1080Ti and using "orginal" Trainer you should be able to increase Batch size to about 180 or even 200 (when running no other graphics programe), which will speed things up quite a bit ! Read the instructions on optimising batch size for training.

User avatar
deephomage
Posts: 33
Joined: Fri Jul 12, 2019 6:09 pm
Answers: 1
Has thanked: 2 times
Been thanked: 8 times

Re: I have a conversion problem

Post by deephomage »

The Faceswap forum really isn't the place to find answers about conversion issues with DeepFaceLab. You should post your problems with the software on the DFL developer's github page. Sometimes other DFL users will offer advice. You could also try contacting the creator of the video you posted and asking him questions.

Faceswap has a conversion preview feature that allows you to view the effect of applying various conversion settings without having to convert the entire video.

As a general matter, you should aim to train your model for at least 100 thousand iterations at the highest batch size your GPU can run without getting an "Out Of Memory" (OOM) error. You should also use the highest quality image samples available for each face set that you want to train.

User avatar
Sconoro
Posts: 8
Joined: Mon Feb 10, 2020 12:23 pm
Contact:

Re: I have a conversion problem

Post by Sconoro »

I use faceswap and not github it was a toolkit issue and I found an answer to it.

The question is what quality should be video for example 1080p and then the same quality or 720p

The author of the video does not answer questions, no matter how much I try, he does it purely for himself and shares his tools especially dont intend.

I have to use optimizer saving because it says that there is not enough memory allocated, I have more

Ge Force 1050TI Intel I7 8700 CPU 3.7GHZ

what settings for it optimal in the guide said to use that it is possible to use everything on the processor, but they wrote it is not recommended, it can not withstand and burn out?

User avatar
deephomage
Posts: 33
Joined: Fri Jul 12, 2019 6:09 pm
Answers: 1
Has thanked: 2 times
Been thanked: 8 times

Re: I have a conversion problem

Post by deephomage »

I don't know what guide you're referring to, but your processor and video card will not "burn out" or get damaged by training a model. You do need to have good cooling for your CPU and video card because training a model stresses your system and takes several days of 24/7 computer time.

720 or 1080 resolution videos should work fine, but videos on YouTube are highly compressed and not very good quality for training purposes. You should use high-bitrate video sources (Blu-ray or DVD) to make your face sets, if available.

Your video card has only 4 Gb. of VRAM, so close all other programs when training a model.

User avatar
Sconoro
Posts: 8
Joined: Mon Feb 10, 2020 12:23 pm
Contact:

Re: I have a conversion problem

Post by Sconoro »

Thanks for the answer. What amount of time does this result of the model take approximately and what result did you have with what computer settings is this interesting?

What iron did you rather train the model with and what result was and how much time did it take?

User avatar
deephomage
Posts: 33
Joined: Fri Jul 12, 2019 6:09 pm
Answers: 1
Has thanked: 2 times
Been thanked: 8 times

Re: I have a conversion problem

Post by deephomage »

I didn't create the Witcher deepfakes, you' ll have to ask the creator. Generally, it takes at least a few days to train a model on a pair of faces. The higher the resolution, the longer it takes to train the model. I use an RTX 2080ti, but an RTX 2060 Super, or a used GTX 10 series card with 8 Gb. is a good choice.

User avatar
djandg
Posts: 43
Joined: Mon Dec 09, 2019 7:00 pm
Has thanked: 4 times
Been thanked: 2 times

Re: I have a conversion problem

Post by djandg »

With a 1050Ti (4Gb VRAM) and maximum batch size on training and the quality you want, you're better off going on holiday for a week or two leaving the training running!

User avatar
Sconoro
Posts: 8
Joined: Mon Feb 10, 2020 12:23 pm
Contact:

Re: I have a conversion problem

Post by Sconoro »

It's easier to upgrade hardware
Will the GTX 1080 TI be fast in time training? One two days?

User avatar
bryanlyon
Site Admin
Posts: 793
Joined: Fri Jul 12, 2019 12:49 am
Answers: 44
Location: San Francisco
Has thanked: 4 times
Been thanked: 215 times
Contact:

Re: I have a conversion problem

Post by bryanlyon »

1080 ti will let you run a larger model, but it wont actually run much faster than a 1050 ti. Maybe 15-20% faster.

User avatar
Sconoro
Posts: 8
Joined: Mon Feb 10, 2020 12:23 pm
Contact:

Re: I have a conversion problem

Post by Sconoro »

Then why did the guy achieve such a result in a couple of days and I did not?I’m still sitting for about two weeks and the result of 74,000 iteration.

User avatar
djandg
Posts: 43
Joined: Mon Dec 09, 2019 7:00 pm
Has thanked: 4 times
Been thanked: 2 times

Re: I have a conversion problem

Post by djandg »

You don't know what GPU he has. Spend a few grand such on a GPU such as a NVidia TItan RTX with many cores and huge memory and then maybe you can do it in a few days.
I have 2 1070ti's in SLI and can, for example, run Dfaker at 128 batch size (90 EG/s) quite easily but it still takes about a hundred hours to get anywhere near to convergance with good "face on" A and B data. Poor A or B data and face movement will take longer.
The 128 batch size may not be optimum, a lower batch size (in multiples of 2 given 2 GPU's) may give a higher EG/s and hence a higher images through the model per hour (images processed = batch X EG/s) but I can't be bothered to optimise for each A/B face run.

User avatar
Sconoro
Posts: 8
Joined: Mon Feb 10, 2020 12:23 pm
Contact:

Re: I have a conversion problem

Post by Sconoro »

I changed the iron now 1080TI I have 11GB of video memory. About 372,000 iterations have passed and the face is also blurred why? How much more time is needed for the Video frames, I don’t understand, a week has passed.

What settings should I set to make this case faster,now with the load of only half the video card and half the processor anyway, in general, things are not going fast for now?

How many iterations should be so that the video is no longer blurred by a million or two million this is clearly not a day or two.

You described the video card above, it is only slightly better than what I have now, but I don’t see the result of the problem.

I’ve been sitting for a week, it’s hopeless.

:cry: :cry: :cry: :cry: :cry:

User avatar
Sconoro
Posts: 8
Joined: Mon Feb 10, 2020 12:23 pm
Contact:

Re: I have a conversion problem

Post by Sconoro »

How many iterations are needed so that the quality is as above in the video and the training settings need to be set to which original or others to make it faster?

User avatar
torzdf
Posts: 2636
Joined: Fri Jul 12, 2019 12:53 am
Answers: 155
Has thanked: 128 times
Been thanked: 614 times

Re: I have a conversion problem

Post by torzdf »

It doesn't really work that way. The final quality will mostly be impacted by the quality of training images

My word is final

Locked