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.

Combining training modes? plz help

havok

DF Vagrant
Im pretty new to deepfacelabs so forgive me if this is a dumb question. To my understanding each mode like H126 and DF have different strengths and weaknesses. like some are better at side shots or with shadows. So if my fake requires different angels, face obstructions ect is there a way to convert my trained data with multiple modes? Or would i just have to splice my footage into sections every time theres a head turn or obstruction?

Also an unrelated question. How large are your files? I have a  209 mb minute clip and the extracted pngs alone are 80gbs with of pictures. are file sizes usually this large?
 

easy40

DF Admirer
Verified Video Creator
havok said:
Im pretty new to deepfacelabs so forgive me if this is a dumb question. To my understanding each mode like H126 and DF have different strengths and weaknesses. like some are better at side shots or with shadows. So if my fake requires different angels, face obstructions ect is there a way to convert my trained data with multiple modes? Or would i just have to splice my footage into sections every time theres a head turn or obstruction?

Also an unrelated question. How large are your files? I have a  209 mb minute clip and the extracted pngs alone are 80gbs with of pictures. are file sizes usually this large?

Your question is not very clear. Regarding models, SAE is supposed to be the best one at reconstructing obstructed faces. H128 is good at handling obstructions too, while DF is crap at doing that.

If you want to use more than one model for one scene, you'll have to train each model from scratch. Since this is very time consuming, it is just better to stick to using a single model for a certain scene.

Your second part of the question is not clear. Is it a one minute clip that weights 209 Mb? Or a 209 minutes video?

If it's a very long video with a very high bitrate then yes, just the extracted PNGs can take tons of space.
 

havok

DF Vagrant
easy40 said:
havok said:
Im pretty new to deepfacelabs so forgive me if this is a dumb question. To my understanding each mode like H126 and DF have different strengths and weaknesses. like some are better at side shots or with shadows. So if my fake requires different angels, face obstructions ect is there a way to convert my trained data with multiple modes? Or would i just have to splice my footage into sections every time theres a head turn or obstruction?

Also an unrelated question. How large are your files? I have a  209 mb minute clip and the extracted pngs alone are 80gbs with of pictures. are file sizes usually this large?

Your question is not very clear. Regarding models, SAE is supposed to be the best one at reconstructing obstructed faces. H128 is good at handling obstructions too, while DF is crap at doing that.

If you want to use more than one model for one scene, you'll have to train each model from scratch. Since this is very time consuming, it is just better to stick to using a single model for a certain scene.

Your second part of the question is not clear. Is it a one minute clip that weights 209 Mb? Or a 209 minutes video?

If it's a very long video with a very high bitrate then yes, just the extracted PNGs can take tons of space.

Seems like you understood the questions :p
OK thanks, yeah I figured id probably have to redo the whole process if I wanted to have multiple training modes in a video. 
The video is 209mb large and around 90 seconds long, its like 4k or something.
 

easy40

DF Admirer
Verified Video Creator
havok said:
easy40 said:
havok said:
Im pretty new to deepfacelabs so forgive me if this is a dumb question. To my understanding each mode like H126 and DF have different strengths and weaknesses. like some are better at side shots or with shadows. So if my fake requires different angels, face obstructions ect is there a way to convert my trained data with multiple modes? Or would i just have to splice my footage into sections every time theres a head turn or obstruction?

Also an unrelated question. How large are your files? I have a  209 mb minute clip and the extracted pngs alone are 80gbs with of pictures. are file sizes usually this large?

Your question is not very clear. Regarding models, SAE is supposed to be the best one at reconstructing obstructed faces. H128 is good at handling obstructions too, while DF is crap at doing that.

If you want to use more than one model for one scene, you'll have to train each model from scratch. Since this is very time consuming, it is just better to stick to using a single model for a certain scene.

Your second part of the question is not clear. Is it a one minute clip that weights 209 Mb? Or a 209 minutes video?

If it's a very long video with a very high bitrate then yes, just the extracted PNGs can take tons of space.

Seems like you understood the questions :p
OK thanks, yeah I figured id probably have to redo the whole process if I wanted to have multiple training modes in a video. 
The video is 209mb large and around 90 seconds long, its like 4k or something.

Yeah, it makes sense that a 4k video with an extremely high rate is so heavy.

No need to train a 4k video in my mind though. Faces are processed with a 128 resolution. Close ups won't be very clear even when using a 720p or 1080p video, I imagine even less clear with a 4k video.
 

Pololo

DF Vagrant
Is there any way to force the resolution to 256? The original PNG is 256x256, but when I train a 1080p video, if the head is not small, it remains as the ass face, is blurred.

I'm trying with photos in a video, and as the face is big, say that the size of 256, are blurred and unreal, but if the face is the size of 128, they are perfect and super realistic. In another topic of the forum I explain, if you put 1000 photos in a video format, you train them perfectly, but only those with a small head, as it is a little bigger, are blurred.

(I'm using google translator, sorry).
 
Top