{
  "id": 665866,
  "title": "why not dinov2 large?",
  "url": "/competitions/recodai-luc-scientific-image-forgery-detection/discussion/665866",
  "author_name": "Praphool Yadav",
  "post_date": "2026-01-04T06:46:30.091000",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Does any one tried dinov2 large/small etc. Now we have dinov3, how to run that?</p>",
  "messages": [
    {
      "id": 3385880,
      "postDate": "2026-01-04T06:46:30.093Z",
      "content": "<p>Does any one tried dinov2 large/small etc. Now we have dinov3, how to run that?</p>",
      "rawMarkdown": "Does any one tried dinov2 large/small etc. Now we have dinov3, how to run that?",
      "votes": 1
    },
    {
      "id": 3386139,
      "postDate": "2026-01-04T17:09:55.617Z",
      "content": "<p>Even after applying Dinov3, I’m still unable to break out of the ~0.303 score range (the baseline score obtained when predicting all images as authentic). Without addressing the competition-specific data scarcity and class imbalance, increasing the model size alone is unlikely to lead to any performance improvement.</p>",
      "rawMarkdown": "Even after applying Dinov3, I’m still unable to break out of the ~0.303 score range (the baseline score obtained when predicting all images as authentic). Without addressing the competition-specific data scarcity and class imbalance, increasing the model size alone is unlikely to lead to any performance improvement.",
      "votes": 2,
      "replies": [
        {
          "id": 3386385,
          "postDate": "2026-01-05T07:33:19.337Z",
          "content": "<p>Yes, imbalances need to handle in all model. </p>",
          "rawMarkdown": "Yes, imbalances need to handle in all model. "
        },
        {
          "id": 3386391,
          "postDate": "2026-01-05T08:00:49.770Z",
          "content": "<p>Is there a kaggle model for dinov3 or you download from huggingface? can you give me the link?</p>",
          "rawMarkdown": "Is there a kaggle model for dinov3 or you download from huggingface? can you give me the link?",
          "votes": 1,
          "replies": [
            {
              "id": 3386400,
              "postDate": "2026-01-05T08:30:30.270Z",
              "rawMarkdown": "",
              "votes": 2,
              "isDeleted": true
            }
          ]
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 3386139,
      "author_name": "littlelittleman",
      "author_url": "",
      "post_date": "2026-01-04T17:09:55.617000",
      "content": "<p>Even after applying Dinov3, I’m still unable to break out of the ~0.303 score range (the baseline score obtained when predicting all images as authentic). Without addressing the competition-specific data scarcity and class imbalance, increasing the model size alone is unlikely to lead to any performance improvement.</p>",
      "votes": 2,
      "replies": [
        {
          "id": 3386385,
          "author_name": "Praphool Yadav",
          "author_url": "",
          "post_date": "2026-01-05T07:33:19.337000",
          "content": "<p>Yes, imbalances need to handle in all model. </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 3386391,
          "author_name": "MenghaoZhao",
          "author_url": "",
          "post_date": "2026-01-05T08:00:49.770000",
          "content": "<p>Is there a kaggle model for dinov3 or you download from huggingface? can you give me the link?</p>",
          "votes": 1,
          "replies": [
            {
              "id": 3386400,
              "author_name": "",
              "author_url": "",
              "post_date": "2026-01-05T08:30:30.270000",
              "content": "",
              "votes": 2,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3385880": "Does any one tried dinov2 large/small etc. Now we have dinov3, how to run that?",
    "3386139": "Even after applying Dinov3, I’m still unable to break out of the ~0.303 score range (the baseline score obtained when predicting all images as authentic). Without addressing the competition-specific data scarcity and class imbalance, increasing the model size alone is unlikely to lead to any performance improvement."
  }
}