{
  "id": 124130,
  "title": "StyleGAN2 - Dec 2019 paper",
  "url": "/competitions/deepfake-detection-challenge/discussion/124130",
  "author_name": "B. Allabadi",
  "post_date": "2020-01-02T05:12:42.378000",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>demo: <a href=\"https://thispersondoesnotexist.com\">https://thispersondoesnotexist.com</a> \npaper <a href=\"https://arxiv.org/abs/1912.04958\">https://arxiv.org/abs/1912.04958</a> </p>\n\n<p>seems quiet challenging to detect fakeness of those faces :) </p>",
  "messages": [
    {
      "id": 708225,
      "postDate": "2020-01-02T05:12:42.377Z",
      "content": "<p>demo: <a href=\"https://thispersondoesnotexist.com\">https://thispersondoesnotexist.com</a> \npaper <a href=\"https://arxiv.org/abs/1912.04958\">https://arxiv.org/abs/1912.04958</a> </p>\n\n<p>seems quiet challenging to detect fakeness of those faces :) </p>",
      "rawMarkdown": "demo: https://thispersondoesnotexist.com \npaper https://arxiv.org/abs/1912.04958 \n\nseems quiet challenging to detect fakeness of those faces :) ",
      "votes": 2
    },
    {
      "id": 710753,
      "postDate": "2020-01-05T07:04:17.177Z",
      "content": "<p>For a human, it looks pretty tough. </p>\n\n<p>There is a whole section of the paper on identifying stylegan fakes. If you have the same weights or know the dataset used to train on, their projection method (the back projection code is included in that repo) is really interesting and quite powerful. </p>\n\n<p>I’ve not yet tried it on a large variety of different kinds of images, but on faces and some other non-face images I’ve been training it on - the difference between a back projected synthetic and real image is quite pronounced - even for totally convincing fakes. </p>",
      "rawMarkdown": "For a human, it looks pretty tough. \n\nThere is a whole section of the paper on identifying stylegan fakes. If you have the same weights or know the dataset used to train on, their projection method (the back projection code is included in that repo) is really interesting and quite powerful. \n\nI’ve not yet tried it on a large variety of different kinds of images, but on faces and some other non-face images I’ve been training it on - the difference between a back projected synthetic and real image is quite pronounced - even for totally convincing fakes. "
    },
    {
      "id": 708480,
      "postDate": "2020-01-02T10:41:03.757Z",
      "content": "<p>By themselves it's probably impossible to spot whether they are fake or not. But in the context of a video, it might be. ;-)</p>",
      "rawMarkdown": "By themselves it's probably impossible to spot whether they are fake or not. But in the context of a video, it might be. ;-)"
    },
    {
      "id": 710793,
      "postDate": "2020-01-05T08:15:53.907Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 710753,
      "author_name": "interneuron",
      "author_url": "",
      "post_date": "2020-01-05T07:04:17.177000",
      "content": "<p>For a human, it looks pretty tough. </p>\n\n<p>There is a whole section of the paper on identifying stylegan fakes. If you have the same weights or know the dataset used to train on, their projection method (the back projection code is included in that repo) is really interesting and quite powerful. </p>\n\n<p>I’ve not yet tried it on a large variety of different kinds of images, but on faces and some other non-face images I’ve been training it on - the difference between a back projected synthetic and real image is quite pronounced - even for totally convincing fakes. </p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 708480,
      "author_name": "Human Analog",
      "author_url": "",
      "post_date": "2020-01-02T10:41:03.757000",
      "content": "<p>By themselves it's probably impossible to spot whether they are fake or not. But in the context of a video, it might be. ;-)</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 710793,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-01-05T08:15:53.907000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "708225": "demo: https://thispersondoesnotexist.com \npaper https://arxiv.org/abs/1912.04958 \n\nseems quiet challenging to detect fakeness of those faces :) ",
    "710753": "For a human, it looks pretty tough. \n\nThere is a whole section of the paper on identifying stylegan fakes. If you have the same weights or know the dataset used to train on, their projection method (the back projection code is included in that repo) is really interesting and quite powerful. \n\nI’ve not yet tried it on a large variety of different kinds of images, but on faces and some other non-face images I’ve been training it on - the difference between a back projected synthetic and real image is quite pronounced - even for totally convincing fakes. ",
    "708480": "By themselves it's probably impossible to spot whether they are fake or not. But in the context of a video, it might be. ;-)",
    "710793": ""
  }
}