I tried it and it gave very similar results to dinov2(even slightly worse)
\ndid any one had a noticeable difference with it compared to dinov2?
", "messages": [ { "id": "3381185", "postDate": "12/24/2025 01:20:56", "content": "I tried it and it gave very similar results to dinov2(even slightly worse)
\ndid any one had a noticeable difference with it compared to dinov2?
", "rawMarkdown": "I tried it and it gave very similar results to dinov2(even slightly worse)\n\ndid any one had a noticeable difference with it compared to dinov2?", "votes": null }, { "id": "3381190", "postDate": "12/24/2025 01:56:41", "content": "I've had a little luck with it, but as you mention it's been pretty similar to DINOv2 for me.
", "rawMarkdown": "I've had a little luck with it, but as you mention it's been pretty similar to DINOv2 for me.", "votes": null }, { "id": "3381376", "postDate": "12/24/2025 11:28:45", "content": "For my part, I got lower results when I replaced Dino v2 with V3. I was rather surprised at first, wondering how a new generation model could be worse than its older version, then by looking into it more I discovered that it didn't work exactly the same way—just the way it divides the image (one is a ViT14 the other is ViT16) and they have subtleties. Both models are pretty good in general but one will always be better than the other depending on your approach, what you give it in addition and the way you do it. However, if I have to bet, I would be sure that the top rated among us are using Dino V3. Besides, I noticed a lot of people climbing to the top since the CSIRO team confirmed that we could use it, but it is absolutely not to be ruled out that certain approaches with Dino V2 can at any moment surpass approaches with Dino V3
", "rawMarkdown": "For my part, I got lower results when I replaced Dino v2 with V3. I was rather surprised at first, wondering how a new generation model could be worse than its older version, then by looking into it more I discovered that it didn't work exactly the same way—just the way it divides the image (one is a ViT14 the other is ViT16) and they have subtleties. Both models are pretty good in general but one will always be better than the other depending on your approach, what you give it in addition and the way you do it. However, if I have to bet, I would be sure that the top rated among us are using Dino V3. Besides, I noticed a lot of people climbing to the top since the CSIRO team confirmed that we could use it, but it is absolutely not to be ruled out that certain approaches with Dino V2 can at any moment surpass approaches with Dino V3", "votes": null }, { "id": "3381884", "postDate": "12/25/2025 20:09:25", "content": "DinoV3 (300M - normal one) sucked for me too, idk why\n0.72 -> 0.70 (this was leaky splitting though)\nWill try one last time…
", "rawMarkdown": "DinoV3 (300M - normal one) sucked for me too, idk why\n0.72 -> 0.70 (this was leaky splitting though)\nWill try one last time...", "votes": null }, { "id": "3381937", "postDate": "12/26/2025 02:04:35", "content": "My understanding is that Dino V2 large was trained on a plant data set - I could not find what the V3 versions were trained on. So if you freezing most of the layers V2 might do better unless you find a V3 that was trained on a plant set.
", "rawMarkdown": "My understanding is that Dino V2 large was trained on a plant data set - I could not find what the V3 versions were trained on. So if you freezing most of the layers V2 might do better unless you find a V3 that was trained on a plant set.", "votes": null }, { "id": "3381950", "postDate": "12/26/2025 02:58:29", "content": "well I never thought about that I assumed both of them are trained on the same datasets so thanks🙂
", "rawMarkdown": "well I never thought about that I assumed both of them are trained on the same datasets so thanks🙂", "votes": null }, { "id": "3381956", "postDate": "12/26/2025 03:49:57", "content": "I asked Gemini Pro if my understanding was correct. I think it said - kind of.
\nThe method for both versions apparently followed similar path - but a ton more images pulled from the web for V3. I need to read the paper - I think I am trying something similar without knowing wtf I am doing :)
", "rawMarkdown": "I asked Gemini Pro if my understanding was correct. I think it said - kind of.\n\nThe method for both versions apparently followed similar path - but a ton more images pulled from the web for V3. I need to read the [paper](https://arxiv.org/abs/2508.10104) - I think I am trying something similar without knowing wtf I am doing :)", "votes": null }, { "id": "3381957", "postDate": "12/26/2025 03:55:20", "content": "DINOv2 is pretrained on LVD-142M, which is a combination of mulit datasets has about 142M images. There are plant datasets in it, but not fully trained on plant datasets. And DINOv3 is pretrained on LVD-1689M, contains 1.7B images collected on Instagram, significantly larger than DINOv2. There's also another version of DINOv3 pretrained on SAT-493M, contains 493M satellite images.
", "rawMarkdown": "DINOv2 is pretrained on LVD-142M, which is a combination of mulit datasets has about 142M images. There are plant datasets in it, but not fully trained on plant datasets. And DINOv3 is pretrained on LVD-1689M, contains 1.7B images collected on Instagram, significantly larger than DINOv2. There's also another version of DINOv3 pretrained on SAT-493M, contains 493M satellite images.", "votes": null }, { "id": "3381960", "postDate": "12/26/2025 04:16:49", "content": "It seems to me that a key difference exists - V2 should be using image sizes like 224, that is divisible by 14. V3 should be using image sizes like 256, that is divisible by 16. I did not change my image size when I looked at V3.
", "rawMarkdown": "It seems to me that a key difference exists - V2 should be using image sizes like 224, that is divisible by 14. V3 should be using image sizes like 256, that is divisible by 16. I did not change my image size when I looked at V3.", "votes": null }, { "id": "3382262", "postDate": "12/27/2025 04:34:25", "content": "dinov3 is allowed ?
", "rawMarkdown": "dinov3 is allowed ?", "votes": null }, { "id": "3382276", "postDate": "12/27/2025 05:24:20", "content": "Got back to trying V3 - have two identical notebooks running on two machines. One with v2 at 448x448 and the other with v3 at 512*512. My initial tune is using 5K worth of pseudo images so sloooooow. Maybe in 2 days I will have a result.
", "rawMarkdown": "Got back to trying V3 - have two identical notebooks running on two machines. One with v2 at 448x448 and the other with v3 at 512*512. My initial tune is using 5K worth of pseudo images so sloooooow. Maybe in 2 days I will have a result.", "votes": null }, { "id": "3382279", "postDate": "12/27/2025 05:39:03", "content": "yes
\nhttps://www.kaggle.com/competitions/csiro-biomass/discussion/617603#3377219
", "rawMarkdown": "yes\n\nhttps://www.kaggle.com/competitions/csiro-biomass/discussion/617603#3377219", "votes": null }, { "id": "3382287", "postDate": "12/27/2025 06:05:26", "content": "regarding the 5k pseudo images I didn't understand the idea are you using self-supervised pre training on non labeled images or pseudo labelling🙂
", "rawMarkdown": "regarding the 5k pseudo images I didn't understand the idea are you using self-supervised pre training on non labeled images or pseudo labelling🙂", "votes": null }, { "id": "3382318", "postDate": "12/27/2025 07:42:18", "content": "It's pseudo labeling - started with 185K of images and selected down to 5K with my current best model folds predictions, along with not using any image with less than 30 grams for total. Still playing with the upsample size for the 357 training images. Not sure its going to work and its pretty slow on a pretty high level PC. I am talking days - not hours.
\nThe 357 training images taking me places I have never been :) I think its a real sucky data set but here to do new and different things and its sure providing that experience.
", "rawMarkdown": "It's pseudo labeling - started with 185K of images and selected down to 5K with my current best model folds predictions, along with not using any image with less than 30 grams for total. Still playing with the upsample size for the 357 training images. Not sure its going to work and its pretty slow on a pretty high level PC. I am talking days - not hours.\n\nThe 357 training images taking me places I have never been :) I think its a real sucky data set but here to do new and different things and its sure providing that experience.", "votes": null }, { "id": "3383136", "postDate": "12/29/2025 13:17:07", "content": "May I ask what your local CV is? Thank you very much
", "rawMarkdown": "May I ask what your local CV is? Thank you very much", "votes": null }, { "id": "3383245", "postDate": "12/29/2025 17:19:45", "content": "Minor differences during this experiment - moving on to the next idea on my list - v3 not a big help for my pipline.
", "rawMarkdown": "Minor differences during this experiment - moving on to the next idea on my list - v3 not a big help for my pipline.", "votes": null }, { "id": "3384510", "postDate": "01/01/2026 12:15:43", "content": "I don't quite remember, but it sucked both on CV and LB (more)… (0.72 -> 0.70 LB)
", "rawMarkdown": "I don't quite remember, but it sucked both on CV and LB (more)... (0.72 -> 0.70 LB)", "votes": null } ], "comments": [ { "id": 3381190, "author_name": "chemdatafarmer", "author_url": "", "post_date": "12/24/2025 01:56:41", "content": "I've had a little luck with it, but as you mention it's been pretty similar to DINOv2 for me.
", "votes": null, "replies": [] }, { "id": 3381376, "author_name": "zera12", "author_url": "", "post_date": "12/24/2025 11:28:45", "content": "For my part, I got lower results when I replaced Dino v2 with V3. I was rather surprised at first, wondering how a new generation model could be worse than its older version, then by looking into it more I discovered that it didn't work exactly the same way—just the way it divides the image (one is a ViT14 the other is ViT16) and they have subtleties. Both models are pretty good in general but one will always be better than the other depending on your approach, what you give it in addition and the way you do it. However, if I have to bet, I would be sure that the top rated among us are using Dino V3. Besides, I noticed a lot of people climbing to the top since the CSIRO team confirmed that we could use it, but it is absolutely not to be ruled out that certain approaches with Dino V2 can at any moment surpass approaches with Dino V3
", "votes": null, "replies": [] }, { "id": 3381884, "author_name": "omashosleeps", "author_url": "", "post_date": "12/25/2025 20:09:25", "content": "DinoV3 (300M - normal one) sucked for me too, idk why\n0.72 -> 0.70 (this was leaky splitting though)\nWill try one last time…
", "votes": null, "replies": [ { "id": 3383136, "author_name": "kagger3", "author_url": "", "post_date": "12/29/2025 13:17:07", "content": "May I ask what your local CV is? Thank you very much
", "votes": null, "replies": [ { "id": 3384510, "author_name": "omashosleeps", "author_url": "", "post_date": "01/01/2026 12:15:43", "content": "I don't quite remember, but it sucked both on CV and LB (more)… (0.72 -> 0.70 LB)
", "votes": null, "replies": [] } ] } ] }, { "id": 3381937, "author_name": "pcjimmmy", "author_url": "", "post_date": "12/26/2025 02:04:35", "content": "My understanding is that Dino V2 large was trained on a plant data set - I could not find what the V3 versions were trained on. So if you freezing most of the layers V2 might do better unless you find a V3 that was trained on a plant set.
", "votes": null, "replies": [ { "id": 3381950, "author_name": "tahaalshatiri", "author_url": "", "post_date": "12/26/2025 02:58:29", "content": "well I never thought about that I assumed both of them are trained on the same datasets so thanks🙂
", "votes": null, "replies": [ { "id": 3381956, "author_name": "pcjimmmy", "author_url": "", "post_date": "12/26/2025 03:49:57", "content": "I asked Gemini Pro if my understanding was correct. I think it said - kind of.
\nThe method for both versions apparently followed similar path - but a ton more images pulled from the web for V3. I need to read the paper - I think I am trying something similar without knowing wtf I am doing :)
", "votes": null, "replies": [ { "id": 3381960, "author_name": "pcjimmmy", "author_url": "", "post_date": "12/26/2025 04:16:49", "content": "It seems to me that a key difference exists - V2 should be using image sizes like 224, that is divisible by 14. V3 should be using image sizes like 256, that is divisible by 16. I did not change my image size when I looked at V3.
", "votes": null, "replies": [ { "id": 3382276, "author_name": "pcjimmmy", "author_url": "", "post_date": "12/27/2025 05:24:20", "content": "Got back to trying V3 - have two identical notebooks running on two machines. One with v2 at 448x448 and the other with v3 at 512*512. My initial tune is using 5K worth of pseudo images so sloooooow. Maybe in 2 days I will have a result.
", "votes": null, "replies": [ { "id": 3382287, "author_name": "tahaalshatiri", "author_url": "", "post_date": "12/27/2025 06:05:26", "content": "regarding the 5k pseudo images I didn't understand the idea are you using self-supervised pre training on non labeled images or pseudo labelling🙂
", "votes": null, "replies": [ { "id": 3382318, "author_name": "pcjimmmy", "author_url": "", "post_date": "12/27/2025 07:42:18", "content": "It's pseudo labeling - started with 185K of images and selected down to 5K with my current best model folds predictions, along with not using any image with less than 30 grams for total. Still playing with the upsample size for the 357 training images. Not sure its going to work and its pretty slow on a pretty high level PC. I am talking days - not hours.
\nThe 357 training images taking me places I have never been :) I think its a real sucky data set but here to do new and different things and its sure providing that experience.
", "votes": null, "replies": [] } ] }, { "id": 3383245, "author_name": "pcjimmmy", "author_url": "", "post_date": "12/29/2025 17:19:45", "content": "Minor differences during this experiment - moving on to the next idea on my list - v3 not a big help for my pipline.
", "votes": null, "replies": [] } ] } ] } ] } ] }, { "id": 3381957, "author_name": "linkinpony", "author_url": "", "post_date": "12/26/2025 03:55:20", "content": "DINOv2 is pretrained on LVD-142M, which is a combination of mulit datasets has about 142M images. There are plant datasets in it, but not fully trained on plant datasets. And DINOv3 is pretrained on LVD-1689M, contains 1.7B images collected on Instagram, significantly larger than DINOv2. There's also another version of DINOv3 pretrained on SAT-493M, contains 493M satellite images.
", "votes": null, "replies": [] } ] }, { "id": 3382262, "author_name": "gokulakrishnanp", "author_url": "", "post_date": "12/27/2025 04:34:25", "content": "dinov3 is allowed ?
", "votes": null, "replies": [ { "id": 3382279, "author_name": "tahaalshatiri", "author_url": "", "post_date": "12/27/2025 05:39:03", "content": "yes
\nhttps://www.kaggle.com/competitions/csiro-biomass/discussion/617603#3377219
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3381185": "I tried it and it gave very similar results to dinov2(even slightly worse)\n\ndid any one had a noticeable difference with it compared to dinov2?", "3381190": "I've had a little luck with it, but as you mention it's been pretty similar to DINOv2 for me.", "3381376": "For my part, I got lower results when I replaced Dino v2 with V3. I was rather surprised at first, wondering how a new generation model could be worse than its older version, then by looking into it more I discovered that it didn't work exactly the same way—just the way it divides the image (one is a ViT14 the other is ViT16) and they have subtleties. Both models are pretty good in general but one will always be better than the other depending on your approach, what you give it in addition and the way you do it. However, if I have to bet, I would be sure that the top rated among us are using Dino V3. Besides, I noticed a lot of people climbing to the top since the CSIRO team confirmed that we could use it, but it is absolutely not to be ruled out that certain approaches with Dino V2 can at any moment surpass approaches with Dino V3", "3381884": "DinoV3 (300M - normal one) sucked for me too, idk why\n0.72 -> 0.70 (this was leaky splitting though)\nWill try one last time...", "3381937": "My understanding is that Dino V2 large was trained on a plant data set - I could not find what the V3 versions were trained on. So if you freezing most of the layers V2 might do better unless you find a V3 that was trained on a plant set.", "3381950": "well I never thought about that I assumed both of them are trained on the same datasets so thanks🙂", "3381956": "I asked Gemini Pro if my understanding was correct. I think it said - kind of.\n\nThe method for both versions apparently followed similar path - but a ton more images pulled from the web for V3. I need to read the [paper](https://arxiv.org/abs/2508.10104) - I think I am trying something similar without knowing wtf I am doing :)", "3381957": "DINOv2 is pretrained on LVD-142M, which is a combination of mulit datasets has about 142M images. There are plant datasets in it, but not fully trained on plant datasets. And DINOv3 is pretrained on LVD-1689M, contains 1.7B images collected on Instagram, significantly larger than DINOv2. There's also another version of DINOv3 pretrained on SAT-493M, contains 493M satellite images.", "3381960": "It seems to me that a key difference exists - V2 should be using image sizes like 224, that is divisible by 14. V3 should be using image sizes like 256, that is divisible by 16. I did not change my image size when I looked at V3.", "3382262": "dinov3 is allowed ?", "3382276": "Got back to trying V3 - have two identical notebooks running on two machines. One with v2 at 448x448 and the other with v3 at 512*512. My initial tune is using 5K worth of pseudo images so sloooooow. Maybe in 2 days I will have a result.", "3382279": "yes\n\nhttps://www.kaggle.com/competitions/csiro-biomass/discussion/617603#3377219", "3382287": "regarding the 5k pseudo images I didn't understand the idea are you using self-supervised pre training on non labeled images or pseudo labelling🙂", "3382318": "It's pseudo labeling - started with 185K of images and selected down to 5K with my current best model folds predictions, along with not using any image with less than 30 grams for total. Still playing with the upsample size for the 357 training images. Not sure its going to work and its pretty slow on a pretty high level PC. I am talking days - not hours.\n\nThe 357 training images taking me places I have never been :) I think its a real sucky data set but here to do new and different things and its sure providing that experience.", "3383136": "May I ask what your local CV is? Thank you very much", "3383245": "Minor differences during this experiment - moving on to the next idea on my list - v3 not a big help for my pipline.", "3384510": "I don't quite remember, but it sucked both on CV and LB (more)... (0.72 -> 0.70 LB)" }, "source": "meta" }