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.

Reverse engineering Seamless Nudity Censorship to create a better version of Deepnude

gerantr

DF Vagrant
According to this paper, the researchers created a program that puts bathing suits on nude women (basically a reverse deepnude filter): https://www.researchgate.net/public...lation_Approach_based_on_Adversarial_Training

Based on my very limited understanding, the current version of deepnude breaks the problem down into 3 parts, which is to first process the image by detecting and highlight the swimsuit, generating a rough geometry of the position of the anatomy, then using the processed image to train the network which is finally used create the nude image. 

This is just my opinion but due to the complexity of the process, the program will often incorrectly position the anatomy which somtime causes it to create really odd and/or disturbing pics.

Instead, a better method would be to generate the training sets by using the Seamless Nudity Censorship program to create a pair of nude and clothed image, then use this to train the network. This is just speculation but I believe it will create a much more realistic and robust program.

I'd really much appreciate it if anyone with a degree in computer science or actual knowledge on machine learning could comment on this.
 
Top