{
  "id": 667753,
  "title": "Does Dinov3-huge underperform Dinov3-large? ",
  "url": "/competitions/csiro-biomass/discussion/667753",
  "author_name": "aban",
  "post_date": "2026-01-14T05:47:40.885000",
  "votes": 1,
  "comment_count": 6,
  "views": 0,
  "content": "<p>I conducted tests on both the huge and large variants of Dinov3 and observed that the huge version exhibited worse performance than the large one. Could you please explain the potential reasons behind this phenomenon?</p>",
  "messages": [
    {
      "id": 3390892,
      "postDate": "2026-01-14T05:47:40.887Z",
      "content": "<p>I conducted tests on both the huge and large variants of Dinov3 and observed that the huge version exhibited worse performance than the large one. Could you please explain the potential reasons behind this phenomenon?</p>",
      "rawMarkdown": "I conducted tests on both the huge and large variants of Dinov3 and observed that the huge version exhibited worse performance than the large one. Could you please explain the potential reasons behind this phenomenon?",
      "votes": 1
    },
    {
      "id": 3391009,
      "postDate": "2026-01-14T09:35:09.690Z",
      "content": "<p>For me best large: .67, best huge: .69. I'm doing most experiments on large for compute reasons.</p>",
      "rawMarkdown": "For me best large: .67, best huge: .69. I'm doing most experiments on large for compute reasons.",
      "votes": 2
    },
    {
      "id": 3392231,
      "postDate": "2026-01-16T14:02:27.860Z",
      "content": "<p>it's in the official paper, page 30 (DINOv3 by Meta)\n<a href=\"https://ai.meta.com/blog/dinov3-self-supervised-vision-model/\" target=\"_blank\">https://ai.meta.com/blog/dinov3-self-supervised-vision-model/</a></p>",
      "rawMarkdown": "it's in the official paper, page 30 (DINOv3 by Meta)\nhttps://ai.meta.com/blog/dinov3-self-supervised-vision-model/"
    },
    {
      "id": 3390951,
      "postDate": "2026-01-14T07:44:59.763Z",
      "rawMarkdown": "",
      "isDeleted": true,
      "replies": [
        {
          "id": 3390956,
          "postDate": "2026-01-14T07:55:37.857Z",
          "content": "<p>Writing in the config, Dino v3 Huge's embedding size is 1280. Can I ask if you're using the large or huge version?</p>",
          "rawMarkdown": "Writing in the config, Dino v3 Huge's embedding size is 1280. Can I ask if you're using the large or huge version?",
          "replies": [
            {
              "id": 3391005,
              "postDate": "2026-01-14T09:24:58.397Z",
              "content": "<p><a href=\"https://www.kaggle.com/code/llkh0a/cisro-baseline-train-infer-21-12-dinov3-siglip/comments#3387066\" target=\"_blank\">https://www.kaggle.com/code/llkh0a/cisro-baseline-train-infer-21-12-dinov3-siglip/comments#3387066</a></p>",
              "rawMarkdown": "https://www.kaggle.com/code/llkh0a/cisro-baseline-train-infer-21-12-dinov3-siglip/comments#3387066"
            },
            {
              "id": 3391014,
              "postDate": "2026-01-14T09:39:28.463Z",
              "content": "<p>Thank you！！</p>",
              "rawMarkdown": "Thank you！！\n"
            }
          ]
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 3391009,
      "author_name": "tennogh",
      "author_url": "",
      "post_date": "2026-01-14T09:35:09.690000",
      "content": "<p>For me best large: .67, best huge: .69. I'm doing most experiments on large for compute reasons.</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 3392231,
      "author_name": "Zack M123",
      "author_url": "",
      "post_date": "2026-01-16T14:02:27.860000",
      "content": "<p>it's in the official paper, page 30 (DINOv3 by Meta)\n<a href=\"https://ai.meta.com/blog/dinov3-self-supervised-vision-model/\" target=\"_blank\">https://ai.meta.com/blog/dinov3-self-supervised-vision-model/</a></p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3390951,
      "author_name": "",
      "author_url": "",
      "post_date": "2026-01-14T07:44:59.763000",
      "content": "",
      "votes": 0,
      "replies": [
        {
          "id": 3390956,
          "author_name": "aban",
          "author_url": "",
          "post_date": "2026-01-14T07:55:37.857000",
          "content": "<p>Writing in the config, Dino v3 Huge's embedding size is 1280. Can I ask if you're using the large or huge version?</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3391005,
              "author_name": "Paritosh Kumar Tripathi",
              "author_url": "",
              "post_date": "2026-01-14T09:24:58.397000",
              "content": "<p><a href=\"https://www.kaggle.com/code/llkh0a/cisro-baseline-train-infer-21-12-dinov3-siglip/comments#3387066\" target=\"_blank\">https://www.kaggle.com/code/llkh0a/cisro-baseline-train-infer-21-12-dinov3-siglip/comments#3387066</a></p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3391014,
              "author_name": "aban",
              "author_url": "",
              "post_date": "2026-01-14T09:39:28.463000",
              "content": "<p>Thank you！！</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3390892": "I conducted tests on both the huge and large variants of Dinov3 and observed that the huge version exhibited worse performance than the large one. Could you please explain the potential reasons behind this phenomenon?",
    "3391009": "For me best large: .67, best huge: .69. I'm doing most experiments on large for compute reasons.",
    "3392231": "it's in the official paper, page 30 (DINOv3 by Meta)\nhttps://ai.meta.com/blog/dinov3-self-supervised-vision-model/",
    "3390951": ""
  }
}