{
  "id": 668017,
  "title": "Appreciation Post",
  "url": "/competitions/recodai-luc-scientific-image-forgery-detection/discussion/668017",
  "author_name": "Maapu",
  "post_date": "2026-01-14T21:58:29.914000",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Just a huge appreciation post to the top 50 contestants. This was tougher than I thought. I'm not sure how everyone ran experiments to crack the top. Would love to read the experimentation process and methods adopted once the competition is closed. </p>\n<p>Apart from BackBone of the architecture, DinoV2 or BiomedClip.. none of the experiments guided mine to improve accuracy. But just a push-pull effect. \nGlobal gate, self-similarity branch, forensic head with BayarConv, prompt pairing with difference in regions -none helped mine in finetuning. Dang. </p>",
  "messages": [
    {
      "id": 3391399,
      "postDate": "2026-01-14T21:58:29.913Z",
      "content": "<p>Just a huge appreciation post to the top 50 contestants. This was tougher than I thought. I'm not sure how everyone ran experiments to crack the top. Would love to read the experimentation process and methods adopted once the competition is closed. </p>\n<p>Apart from BackBone of the architecture, DinoV2 or BiomedClip.. none of the experiments guided mine to improve accuracy. But just a push-pull effect. \nGlobal gate, self-similarity branch, forensic head with BayarConv, prompt pairing with difference in regions -none helped mine in finetuning. Dang. </p>",
      "rawMarkdown": "Just a huge appreciation post to the top 50 contestants. This was tougher than I thought. I'm not sure how everyone ran experiments to crack the top. Would love to read the experimentation process and methods adopted once the competition is closed. \n\nApart from BackBone of the architecture, DinoV2 or BiomedClip.. none of the experiments guided mine to improve accuracy. But just a push-pull effect. \nGlobal gate, self-similarity branch, forensic head with BayarConv, prompt pairing with difference in regions -none helped mine in finetuning. Dang. ",
      "votes": 3
    },
    {
      "id": 3392034,
      "postDate": "2026-01-16T07:05:26.707Z",
      "content": "<p>I’ve tried a bunch of block scanning methods, and they all turned out pretty terrible.”</p>",
      "rawMarkdown": "I’ve tried a bunch of block scanning methods, and they all turned out pretty terrible.”",
      "votes": 1
    },
    {
      "id": 3391502,
      "postDate": "2026-01-15T04:39:44.923Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 3392034,
      "author_name": "meepolimi1",
      "author_url": "",
      "post_date": "2026-01-16T07:05:26.707000",
      "content": "<p>I’ve tried a bunch of block scanning methods, and they all turned out pretty terrible.”</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 3391502,
      "author_name": "",
      "author_url": "",
      "post_date": "2026-01-15T04:39:44.923000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3391399": "Just a huge appreciation post to the top 50 contestants. This was tougher than I thought. I'm not sure how everyone ran experiments to crack the top. Would love to read the experimentation process and methods adopted once the competition is closed. \n\nApart from BackBone of the architecture, DinoV2 or BiomedClip.. none of the experiments guided mine to improve accuracy. But just a push-pull effect. \nGlobal gate, self-similarity branch, forensic head with BayarConv, prompt pairing with difference in regions -none helped mine in finetuning. Dang. ",
    "3392034": "I’ve tried a bunch of block scanning methods, and they all turned out pretty terrible.”",
    "3391502": ""
  }
}