Mr DeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
   
Finding Interview Videos and Picture Galleries
#1
For my example, I basically found the celebrity on Instagram. 
This was the first place to head for to find video and pictures  
Hence the shareable link to there was included in the request. 

Then I did a search on YouTube for interview videos. But after copypasting a few links, one hit on their own channel.
And realises that the fakers can probably do that much of a YouTube search as well.

The key seems to be to check and make sure there is plenty of Material, both picture and video, to feed the AI with.

Would be glad if those more experienced in the field could elaborate which guidelines data should meet. From a laymen view one could say the more the better
#2
I find social media pics almost useless, they usually make my data worse. Too many filters, bad lighting, facial obstruction. IMO the best data comes from professional media events, or high def interviews.

I look for natural lighting, a clear face with no wild makeup or bangs covering the eyebrows/side of the face. Make sure to remove blurry pics, and anything where the face is obscured by a hand etc.

My workflow is usually - Find a decent video, run it through ffmpeg to extract ~5 frames per second, extract faces, run those faces through dupeguru at 90-95%, quickly scroll through and delete any bad pics. A bit of waiting around but should only take about 15 mins of effort.
#3
I usually do the following first when building a faceset:
1) Look on youtube for interviews - most end here, if not
2) Search for TV clips, trailers, movies, etc. - if none
3) social media, wallpapers, image galleries - rarely do I do this, but the video quality for this method will likely only be a face-on masturbation video
#4
(10-09-2018, 12:23 AM)matritinep Wrote: You are not allowed to view links. Register or Login to view.Would be glad if those more experienced in the field could elaborate which guidelines data should meet. From a laymen view one could say the more the better

I'm wondering about this too. I mainly use video frames for faceset.

I know that if the faceset don't have enough lighting diversity, it can cause face color difference after conversion. But I also don't want to spend enormous time on collecting faceset.

So, how many videos are enough for one faceset, to not cause face color difference and save time?
#5
(12-03-2018, 03:34 AM)LingCHN19 Wrote: You are not allowed to view links. Register or Login to view.
(10-09-2018, 12:23 AM)matritinep Wrote: You are not allowed to view links. Register or Login to view.Would be glad if those more experienced in the field could elaborate which guidelines data should meet. From a laymen view one could say the more the better

I'm wondering about this too. I mainly use video frames for faceset.

I know that if the faceset don't have enough lighting diversity, it can cause face color difference after conversion. But I also don't want to spend enormous time on collecting faceset.

So, how many videos are enough for one faceset, to not cause face color difference and save time?

depend of the video lenght, the lighting source (if it change all the time or not), but usually one will do the job great Smile

Try to not collect thousand of face_set for one model. just extract you "src" with 5 fps (by this way you gain weight and many time for training.) 
cause if you collect src face at 30 fps or 60...it will be a waste of time. it will extract 60 frames per sec for all the video, finally you will get thousand and thousand of face that will be closely the same.
(not a good idea) juste keep hummm......2500 Faces for one src. not too many.

(10-10-2018, 04:09 PM)HippieJ Wrote: You are not allowed to view links. Register or Login to view.I find social media pics almost useless, they usually make my data worse. Too many filters, bad lighting, facial obstruction. IMO the best data comes from professional media events, or high def interviews.

I look for natural lighting, a clear face with no wild makeup or bangs covering the eyebrows/side of the face. Make sure to remove blurry pics, and anything where the face is obscured by a hand etc.

My workflow is usually - Find a decent video, run it through ffmpeg to extract ~5 frames per second, extract faces, run those faces through dupeguru at 90-95%, quickly scroll through and delete any bad pics. A bit of waiting around but should only take about 15 mins of effort.

no need to use ffmpeg DFL already do that :/
And it sort picture too by blur histogram etc... Too.
#6
Is it a problem if the faceset is really big?
I thought MT extraction also adds jaw data/view direction, so while training, the model automatically ignores tons of images as they don't suit the the target image.

Am i wrong with this?

Forum Jump:

Users browsing this thread: 1 Guest(s)