MrDeepFakes Forums

Some content may not be available to Guests. Consider registering an account to enjoy unrestricted access to guides, support and tools

  • We are looking for community members who are intested in helping out. See our HELP WANTED post.

Blurry Image, any idea ?

ceilan12

DF Pleb
Hello, i am woking on a project, the first face i'v worked on is pretty good. but the 2nd, the face is really blurry.

i use H128 train. DeepFaceLabCUDA9.2SSE .
[font=Tahoma,Verdana,Arial,Sans-Serif]i got few hundred images.  really good face and profil all in HD.  (1100+ aligned images)
525,000 iter
[/font]

I tried to convert with : seamless then Hist match. Différents tests but no good results
Model Options : https://ibb.co/SR8YLQw
Autobackup : true
write_preview_history : True
batch_size : 4
sort_by_yawn : False
random_flip : False
lighter_ae : Frue
pixel loss : False
running on : Geforce GTX 980M

important information : my src model has glasses, maybe the blur problem come from here ?
[video=youtube]
 

TMBDF

Moderator | Deepfake Creator | Guide maintainer
Staff member
Moderator
Verified Video Creator
This is an abysmal result for 525 thousand iterations so your settings are probably set up wrong and yes, glasses are not helping.
Since you have so many iteration and blurry output you probably trained with low batch size, make sure you read the guides about DFL.

In this case (if you want glasses) you should train it on src faceset without glasses, use SAE full face mode with DF architecture and when converting use FAN-DST or FAN-DST conversion method because it is trained on various obstacles and it can technically mask these parts of the faked face out, leaving glasses of dst beneath visible, it may not work for glasses that well and you may end up having to manually mask out glasses from the DST video and then put them in video editing software over the fake clip.

If you have any questions you'd like to get answers post them in this thread:
https://mrdeepfakes.com/forums/thread-faq-creation-thread
We are making a new DFL guide and FQ/tips thread for such issues so make sure to post (one entry, please try to ask all your questions in one post) what you'd like to see answers to.
 

ceilan12

DF Pleb
I see, thanks a lot for the help Tust ! 
Do i have to clear workspace and restart from scratch if i train SAE ?
i just edited and posted my model option. Yes i got 4 batch size, but with more i got some error, and i made a ton of test.
Anyway, i'm gonna keep working on it, but batch size affect speed not quality ? so i don't understand
 

TMBDF

Moderator | Deepfake Creator | Guide maintainer
Staff member
Moderator
Verified Video Creator
ceilan12 said:
I see, thanks a lot for the help Tust ! 
Do i have to clear workspace and restart from scratch if i train SAE ?
i just edited and posted my model option. Yes i got 4 batch size, but with more i got some error, and i made a ton of test.
Anyway, i'm gonna keep working on it, but batch size affect speed not quality ? so i don't understand

Thats why we need a new guide, people don't understand how basic things work because they are not explained well in existing guides. Batch size does affect quality as having higher one causes model to be more accurate, it does also increase iteration time and vram usage co if you are getting OOM errors it means you are running out of vram. You can get similar results with lower batch like 8-10 compared to 16 but there is no comparison between a model trained to 60-80k iteration a batch size of 36 (which you can get to easily with SAE on Goole Colab as long as you get Tesla T4 assigned) and a batch size 4 one. I recommend you giving DFL on Colab a try or running SAE at at least a batch of 4 for first 20k iterations and then increasing it to 8-12 till results become clear, use optimizer mode 3 to utilise CPU and its RAM additionally and to get rid of OOM errors.

What's your gpu?
 

ceilan12

DF Pleb
tutsmybarreh said:
ceilan12 said:
I see, thanks a lot for the help Tust ! 
Do i have to clear workspace and restart from scratch if i train SAE ?
i just edited and posted my model option. Yes i got 4 batch size, but with more i got some error, and i made a ton of test.
Anyway, i'm gonna keep working on it, but batch size affect speed not quality ? so i don't understand

Thats why we need a new guide, people don't understand how basic things work because they are not explained well in existing guides. Batch size does affect quality as having higher one causes model to be more accurate, it does also increase iteration time and vram usage co if you are getting OOM errors it means you are running out of vram. You can get similar results with lower batch like 8-10 compared to 16 but there is no comparison between a model trained to 60-80k iteration a batch size of 36 (which you can get to easily with SAE on Goole Colab as long as you get Tesla T4 assigned) and a batch size 4 one. I recommend you giving DFL on Colab a try or running SAE at at least a batch of 4 for first 20k iterations and then increasing it to 8-12 till results become clear, use optimizer mode 3 to utilise CPU and its RAM additionally and to get rid of OOM errors.

What's your gpu?

GPU = GeForce 980m
CPU : I7 4870HQ
8go ram. (i think it's why i got errors quicky if superior to 4 (batch size) )

Thanks a lot for the informations. Well i have to find a way to use at least 20+ batch size then.
 I am a bit lost about the SAE goole colab / DFL on colab, can i follow this guide to begin with it ? 
https://mrdeepfakes.com/forums/thread-deepfacelab-google-colab-tutorial

currently, i am using version : DeepFaceLabCUDA9.2SSE


 
 

TMBDF

Moderator | Deepfake Creator | Guide maintainer
Staff member
Moderator
Verified Video Creator
Again, prepare your src without any glasses, extract frames and extract/align them, do the same for dst, create a new folder called workspace inside of which put copy of data_dst, data_src (with aligned folder and all those frames) and an empty model folder. You can copy only aligned folder as thats all we need for training, you save space on google drive and time uploading by skipping aligned_debug in dst as well as original frames in both. Pack this folder (workspace) into a zip as workspace.zip and upload to google drive. Then follow instruction in the Colab notebook, first run Clone Github repository and install requirements, select clone, wait till it says done, then upload workspace by running Import from Drive, click the link that appears, select the google account on which you have the workspace uploaded, then copy the code and paste back into colab. Lastly start training, select sae and do everything as it says in the guide.
 

ceilan12

DF Pleb
Sorry i know you detailled everything, im not a pro and i don"t understand everything.
I understand you told me to try SAE, problem is when i launch he SAE training, after put the options, it just say : saving …

so i guess my only last and 2nd option is to try SAE on Goole Colab like you said ?
 
How many SRC images do you have? Batch size of 4 will reduce your quality as well as having lighter AE enabled. @dpfks has a great guide on this forum and using Colab is really easy and you will get much better results with the Tesla T4. If you get a K80 I would restart it until you get a T4. With 128 resolution on a SAE model, I can usually get a batch size of 24 which will significantly help your result as long as your SRC images are good.
 

ceilan12

DF Pleb
i have 1004 src image all HD. I used lighter ae for bypissing the memory errors.
with H64 it was working without it, but not with H128.
I have read the good guide of dpfks, all i know is from there. I just tested H64 and H128 for now

I'm curently reading the guide for colab and uploading my workspace right now
Idon't understand yet the tesla T4 / K80 thing (don't know what it is ) so i can't understand all you said after that for now, but i appreciate your help, thanks
 

ceilan12

DF Pleb
I just switched to my "big" pc instead of laptop. (1080 / i7 8700k)
I have a question, is batch size 8 mean = 8gb for the GPU ?
so for exemple with a 1080 8gb i can't get over 8 batch size ?

i'm trying to custom a bit the optimizer mode, but it's not dangerous ?
 

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
ceilan12 said:
I just switched to my "big" pc instead of laptop. (1080 / i7 8700k)
I have  a question, is batch size 8 mean = 8gb for the GPU ?
so for exemple with a 1080 8gb i can't get over 8 batch size ?

i'm trying to custom a bit the optimizer mode, but it's not dangerous ?

No it doesn't mean that. It's how many images are in your batch while training. A higher number will generally give you a better model. You only have to use a higher optimizer mode if you are having OOM errors on the settings you want.
 
ceilan12 said:
i have 1004 src image all HD. I used lighter ae for bypissing the memory errors.
with H64 it was working without it, but not with H128.
I have read the good guide of dpfks, all i know is from there. I just tested H64 and H128 for now

I'm curently reading the guide for colab and uploading my workspace right now
Idon't understand yet the tesla T4 / K80 thing (don't know what it is )  so i can't understand all you said after that for now, but i appreciate your help, thanks

1004 source images should be good. Did you get Colab working? I am sorry about not being clear wth the T4 and the K80.  Google will give you a Tesla T4 or a K80 and it will tell you once you start training which one it gives you.  I have found the T4 significantly better, and I won't train with a K80.  After a few restarts, I'm always able to get one. 

Hope this helps
 

ceilan12

DF Pleb
Thank you Both for your replys. I learned a lot since yesterday and understand better now what you were trying to explain.

I made my first try with colab, but i failed. I'm gonna retest. Meanwhile, i have tested SAE on my PC, crazy quality comparing to H128 ...

I got many OOM errors before i was finally able to start the training.
Here are my options : do you think i can push it further with my gtx 1080 8gb ?

Like 30 batch size, 256 res ?

hq38aHAh.png
 

TMBDF

Moderator | Deepfake Creator | Guide maintainer
Staff member
Moderator
Verified Video Creator
If you don't mind could you write down your question into a short a simple sentence, together with an answer we gave you (try to also write it as short as possible), we will be making an FAQ/Tips thread for new users and we want your help in making it so other new users can get answers quickly and understand it right away (hence the need for a short and simple nature of these questions/answers):
https://mrdeepfakes.com/forums/thread-faq-creation-thread
 

ceilan12

DF Pleb
Ok im gonna do that @"tutsmybarreh"

By the way, tried my first converted MP4 video with SAE after 35k iteration but got this message, any idea ?
data_dst\merged folder. is not empty, i got image inside. (maybe i can post it in the thread too)

N1tP75Lh.png
 

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
ceilan12 said:
Ok im gonna do that @"tutsmybarreh"

By the way, tried my first converted MP4 video with SAE after 35k iteration but got this message, any idea ?
data_dst\merged folder. is not empty, i got image inside. (maybe i can post it in the thread too)

N1tP75Lh.png
Run your conversion fully first. Also do not deleted any images in data_dst
 

ceilan12

DF Pleb
dpfks said:
ceilan12 said:
Ok im gonna do that @"tutsmybarreh"

By the way, tried my first converted MP4 video with SAE after 35k iteration but got this message, any idea ?
data_dst\merged folder. is not empty, i got image inside. (maybe i can post it in the thread too)
Run your conversion fully first. Also do not deleted any images in data_dst

Hello, conversion was finished, and i did not touched anything in data_dst.

The only thing i did was import my workspace folder (i used it with H128 before ) and i launched a new train & new options with it ( with SAE training this time)

Step i did : 1) run Train SAE.bat 2) stoped training. 3) run convert SAE.bat (convert goes to 100%, i close it) 4) run Converted to MP4.bat
 

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
ceilan12 said:
dpfks said:
ceilan12 said:
Ok im gonna do that @"tutsmybarreh"

By the way, tried my first converted MP4 video with SAE after 35k iteration but got this message, any idea ?
data_dst\merged folder. is not empty, i got image inside. (maybe i can post it in the thread too)
Run your conversion fully first. Also do not deleted any images in data_dst

Hello, conversion was finished, and i did not touched anything in data_dst.

The only thing i did was import my workspace folder (i used it with H128 before ) and i launched a new train & new options with it ( with SAE training this time)

Step i did : 1) run Train SAE.bat 2) stoped training. 3) run convert SAE.bat (convert goes to 100%, i close it) 4) run Converted to MP4.bat

This is in colab? Check if you have enough space. 

Or better yet download your workspace to make sure everything is there, including all merged images. The number of merged images should match the extracted frames in data_dst
 

ceilan12

DF Pleb
dpfks said:
ceilan12 said:
dpfks said:
ceilan12 said:
Ok im gonna do that @"tutsmybarreh"

By the way, tried my first converted MP4 video with SAE after 35k iteration but got this message, any idea ?
data_dst\merged folder. is not empty, i got image inside. (maybe i can post it in the thread too)
Run your conversion fully first. Also do not deleted any images in data_dst

Hello, conversion was finished, and i did not touched anything in data_dst.

The only thing i did was import my workspace folder (i used it with H128 before ) and i launched a new train & new options with it ( with SAE training this time)

Step i did : 1) run Train SAE.bat 2) stoped training. 3) run convert SAE.bat (convert goes to 100%, i close it) 4) run Converted to MP4.bat

This is in colab? Check if you have enough space. 

Or better yet download your workspace to make sure everything is there, including all merged images. The number of merged images should match the extracted frames in data_dst

nono it's on my pc (i checked the space, it's ok apparently) 

I just checked again and my merged folder is now empty o_O
(I just copied my workspace from my laptop on usb key and i put it on my new pc. then i did the step i told you before.)

i'm gonna delete everything and start again from scratch to not lose any more time, and see if it work or not
 

dpfks

DF Enthusiast
Staff member
Administrator
Verified Video Creator
ceilan12 said:
dpfks said:
ceilan12 said:
dpfks said:
ceilan12 said:
Ok im gonna do that @"tutsmybarreh"

By the way, tried my first converted MP4 video with SAE after 35k iteration but got this message, any idea ?
data_dst\merged folder. is not empty, i got image inside. (maybe i can post it in the thread too)
Run your conversion fully first. Also do not deleted any images in data_dst

Hello, conversion was finished, and i did not touched anything in data_dst.

The only thing i did was import my workspace folder (i used it with H128 before ) and i launched a new train & new options with it ( with SAE training this time)

Step i did : 1) run Train SAE.bat 2) stoped training. 3) run convert SAE.bat (convert goes to 100%, i close it) 4) run Converted to MP4.bat

This is in colab? Check if you have enough space. 

Or better yet download your workspace to make sure everything is there, including all merged images. The number of merged images should match the extracted frames in data_dst

nono it's on my pc (i checked the space, it's ok apparently) 

I just checked again and my merged folder is now empty o_O
(I just copied my workspace from my laptop on usb key and i put it on my new pc. then i did the step i told you before.)

i'm gonna delete everything and start again from scratch to not lose any more time, and see if it work or not

This is super annoying but I do this sometimes...

So what happens is I run convert SAE -> it takes like 12 hours. Then I come back later and accidently use the SAME bat file... as soon as you double click it it will DELETE your merged files...Yup That's 12 hours lost.

I wonder if you did this too.
 
Top