{
  "id": 134295,
  "title": "<CNN-generated images are surprisingly easy to spot...for now>",
  "url": "/competitions/deepfake-detection-challenge/discussion/134295",
  "author_name": "",
  "post_date": "2020-03-07T07:07:41.523895600Z",
  "votes": 2,
  "comment_count": 1,
  "views": 0,
  "content": "<p>A CVPR2020 paper: 'CNN-generated images are surprisingly easy to spot...for now'</p>\n\n<p><a href=\"https://arxiv.org/abs/1912.11035\">paper here</a></p>\n\n<p><a href=\"https://peterwang512.github.io/CNNDetection/\">project page here</a></p>\n\n<p><a href=\"https://github.com/peterwang512/CNNDetection\">github here (and pretrained model)</a></p>\n\n<p>If anyone has enough computing power, you can try what kind of results this paper can get in this competition.</p>",
  "messages": [
    {
      "id": "765833",
      "postDate": "03/07/2020 07:07:41",
      "content": "<p>A CVPR2020 paper: 'CNN-generated images are surprisingly easy to spot...for now'</p>\n\n<p><a href=\"https://arxiv.org/abs/1912.11035\">paper here</a></p>\n\n<p><a href=\"https://peterwang512.github.io/CNNDetection/\">project page here</a></p>\n\n<p><a href=\"https://github.com/peterwang512/CNNDetection\">github here (and pretrained model)</a></p>\n\n<p>If anyone has enough computing power, you can try what kind of results this paper can get in this competition.</p>",
      "rawMarkdown": "A CVPR2020 paper: 'CNN-generated images are surprisingly easy to spot...for now'\n\n[paper here](https://arxiv.org/abs/1912.11035)\n\n[project page here](https://peterwang512.github.io/CNNDetection/)\n\n[github here (and pretrained model)](https://github.com/peterwang512/CNNDetection)\n\nIf anyone has enough computing power, you can try what kind of results this paper can get in this competition.",
      "votes": null
    },
    {
      "id": "766094",
      "postDate": "03/07/2020 16:53:55",
      "content": "<p>It seems to work good for synthetic images that are very easy to spot (i.e. many pixels corruptions, blurs, etc.).\nI've just tested their pretrained model on some faces from dfdc dataset, and the prob of being fake was always &lt; 5% for any face. Not sure if it worth spending time on it for training on dfdc data (they don't provide training script as well). But I guess it won't work good. Anyway their model is just another classifier.</p>",
      "rawMarkdown": "It seems to work good for synthetic images that are very easy to spot (i.e. many pixels corruptions, blurs, etc.).\nI've just tested their pretrained model on some faces from dfdc dataset, and the prob of being fake was always &lt; 5% for any face. Not sure if it worth spending time on it for training on dfdc data (they don't provide training script as well). But I guess it won't work good. Anyway their model is just another classifier.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 766094,
      "author_name": "vostankovich",
      "author_url": "",
      "post_date": "03/07/2020 16:53:55",
      "content": "<p>It seems to work good for synthetic images that are very easy to spot (i.e. many pixels corruptions, blurs, etc.).\nI've just tested their pretrained model on some faces from dfdc dataset, and the prob of being fake was always &lt; 5% for any face. Not sure if it worth spending time on it for training on dfdc data (they don't provide training script as well). But I guess it won't work good. Anyway their model is just another classifier.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "765833": "A CVPR2020 paper: 'CNN-generated images are surprisingly easy to spot...for now'\n\n[paper here](https://arxiv.org/abs/1912.11035)\n\n[project page here](https://peterwang512.github.io/CNNDetection/)\n\n[github here (and pretrained model)](https://github.com/peterwang512/CNNDetection)\n\nIf anyone has enough computing power, you can try what kind of results this paper can get in this competition.",
    "766094": "It seems to work good for synthetic images that are very easy to spot (i.e. many pixels corruptions, blurs, etc.).\nI've just tested their pretrained model on some faces from dfdc dataset, and the prob of being fake was always &lt; 5% for any face. Not sure if it worth spending time on it for training on dfdc data (they don't provide training script as well). But I guess it won't work good. Anyway their model is just another classifier."
  },
  "source": "meta"
}