DeepFaceLab

from Wikipedia, the free encyclopedia

DeepFaceLab (also DFL ) is one of the best-known open source software for creating so-called DeepFakes , i.e. H. Replacing a person's face with another face in a video. The software was programmed in the Python programming language and is designed for Nvidia graphics cards . DeepFaceLab is currently compatible with Windows , Google Colab , CentOS Linux and Linux .

history

DeepFaceLab split from Faceswap in 2019 and was primarily developed by one person. The developer of DeepFaceLab is called "iperov" on the GitHub platform . A total of 17 “contributors” work on DeepFaceLab. Since DeepFaceLab was released, versions for Google Colab, CentOS and Linux have also been released.

Process / function

DeepFaceLab uses the machine learning framework open source library " TensorFlow " from Google , which is often used for artificial intelligence or machine learning. For a good deepfake, the software needs ten to twenty seconds of video material from similar scenarios from both people. It is also helpful to use two people who are as similar as possible. DeepFaceLab does not consist of a single program , but mainly contains many .bat files and pre-trained models.

To create a deepfake with DeepFaceLab you need a video that includes the person whose face you want to use. A second video includes the person whose facial expressions are broadcast. Images are generated from these videos, on which the software then recognizes the outlines of the face and important points such as eyes, mouth and nose.

As soon as this preparation is completed, the Artificial Intelligence begins training the artificial neural networks by disassembling the individual images of both videos in an encoder and then reassembling them in two decoders , one for each video or each face of the people . At the end, the artificial intelligence compares the original with the output of the decoder for each image in the two videos and then modifies the weightings of the artificial neural network. This process is repeated countless times until the result of the decoder is very similar to the original image. This procedure is called supervised learning . In the next step, the encoder of one person is connected to the decoder of the other person in order to transfer the facial expressions from one person to the other person.

Finally, these data are merged into a video and saved in a selectable file format.

Detailed instructions on how to use this software are freely available.

distribution

According to the website Github, DeepFaceLab is used for 95% of all deepfake videos and DeepFaceLab is the "leading software for creating deepfakes". There is now a large community around DeepFaceLab, which is represented on platforms such as Telegram , Reddit or MrDeepFakes. You can also find pre-trained models from the community.

Videos that were created with DeepFaceLab continue to achieve great popularity and achieve views in the millions on the YouTube video platform . In some forums , questions about the DeepFaceLab often reach well over 100,000 views and several hundred posts.

criticism

Criticism of DeepFaceLab is, among other things, that hardly any additional functions have been added since Faceswap was split off. Functions such as “ MultiGPU ” or “Color Augmentation” are cited as examples .

DeepFaceLab does not offer a user-friendly interface because of the many .bat files, already trained models and other files.

In addition, for a long time you needed a powerful PC with a good GPU to use DeepFaceLab, but you can now use DeepFaceLab via Google Colab.

There is also an increased risk of fraud through deepfakes , authentication using videos is less secure and can be misused in various ways (e.g. in so-called revenge porn ).

Web links

Individual evidence

  1. ↑ Synced: Chinese “DeepFake” App Goes Viral, Renewing Concerns About Potential Misuse of Face-Swapping Tech. September 3, 2019, accessed May 23, 2020 (American English).
  2. ↑ a b c d iperov / DeepFaceLab. In: Github. Retrieved May 23, 2020 (English).
  3. ↑ a b r / SFWdeepfakes - Looking for background reading about DeepFaceLab. In: Reddit. Retrieved May 23, 2020 (American English).
  4. ↑ iperov - Overview. In: Github. Retrieved May 23, 2020 (English).
  5. ↑ The UNGE BIFI FAKE: Did Unge really eat meat? In: YouTube. October 13, 2019, accessed May 23, 2020 .
  6. ↑ DeepFaceLab 2.0 Guide / Tutorial. In: MrDeepFakes. May 2, 2020, accessed on May 23, 2020 .
  7. ↑ Deepfake videos: can you spot the fake? | reporter. In: YouTube. reporter, May 13, 2020, accessed on May 23, 2020 .
  8. ↑ chervonij: chervonij / DFL-Colab. In: Github. May 22, 2020, accessed on May 23, 2020 .
  9. ↑ Ctrl Shift Face. In: YouTube. Retrieved May 23, 2020 (English).
  10. ↑ MrDeepFakes Forums - Guides and Tutorials. In: MrDeepFakes. Retrieved May 23, 2020 (English).
  11. ↑ Create deepfakes with DeepFaceLab. In: Ress. October 15, 2019, accessed on May 23, 2020 .