| Connector | General | Knowledge | OCR& Chart Vision-Centric | |
| Concat. [126] | 67.2 | 48.9 | 50.1 | 52.6 |
| Resampler [58] | 63.1 | 46.5 | 27.1 | 42.6 |
| SVA-no-multi-agg | 68.0 | 49.5 | 55.2 | 52.6 |
| SVA | 68.5 | 49.7 | 55.5 | 53.2 |
| 150k | 53.7 | 68.0 | 51.3 | 45.2 | 50.5 |
| 250k54.3 | 68.1 | 51.5 | 45.3 | 52.2 | |
| 350k | 54.3 | 67.4 | 51.446.0 | 52.3 | |
| 450k | 54.2 | 68.0 | 52.2 | 45.5 | 50.7 |
| LLaVA-665K | 40.7 | 64.7 45.2 | 20.8 | 32.0 |
| Cambrian-10M | 54.8 | 68.7 | 51.647.3 | 51.4 |
| Cambrian-7M | 55.9 | 69.6 52.6 | 47.3 | 54.1 |
| Model | Knowledge | OCR & Chart | Vision-Centric | ||||||||||||||||||||||
| BAA | EEE | COO | BAA | ANNNN | A8 | VOAIXOL | dANN | ||||||||||||||||||
| Method GPT-4V | UNK. | 63.0 | 1409.4 | 75.8 | 69.1 | 65.2 | 75.7 | 56.8 | 49.9 | 78.2 | 64.5 | A | 64.3 | ||||||||||||
| Gemini-1.0 Pro | UNK. | 1496.6 | 73.6 | 70.7 | 36.8 | 79.5 | 47.9 | 45.2 | 77.4 | 78.5 | 65.9 | 78.0 - | 88.4 - | 62.4 | 50.0 | 61.4 | 73.8 | ||||||||
| Gemini-1.5 Pro | UNK. | 58.5 | 52.1 | - 80.3 | - 81.3 | 73.5 | 86.5 | = | 67.5 | = | |||||||||||||||
| Grok-1.5 | 53.6 | 52.8 | 88.3 | 76.1 | 78.1 | 68.7 | |||||||||||||||||||
| MM-1-8B | UNK. 144 | 1529.3 | 72.3 | 69.9 | 72.6 | 37.0 | 35.9 | 85.6 | |||||||||||||||||
| MM-1-30B | 144 | 1637.6 | 75.1 | 72.1 | 81.0 | 44.7 | 39.4 | = | = | = | = | ||||||||||||||
| BaseLLM:Llama-3-Ins-8B | |||||||||||||||||||||||||
| Mini-Gemini-HD-8B | 2880 | 72.7 | 1606.0 | 72.7 | 73.264.5 | 55.7 | 75.1 | 37.3 | 37.073.5 | 62.9 | 59.1 | 47.7 | 70.2 | 74.6 | 51.518.7 | 62.1 | 62.263.0 | ||||||||
| LLaVA-NeXT-8B | 2880 | 72.5 | 1603.7 | 72.1 | 72.7 | 65.2 | 55.6 | 72.8 | 41.7 | 36.3 | 71.6 | 63.9 | 69.5 | 49.0 | 64.6 | 72.6 | 56.6 | 38.7 | 60.1 | 62.2 | 65.3 | ||||
| Cambrian-1-8B | 576 | 73.1 | 1,547.1 | 75.9 | 74.7 | 64.6 | 61.3 | 80.4 | 42.7 | 49.0 | 73.0 | 71.3 | 73.3 | 62.4 | 71.7 | 77.8 | 65.051.3 | 64.2 | 72.3 | 72.0 | |||||
| Base LLM: Vicuna-1.5-13B | |||||||||||||||||||||||||
| Mini-Gemini-HD-13B | 2880 | 70.7 | 1597.0 | 68.6 | 70.663.7 | 54.171.9 | 37.337.070.1 | 60.856.646.6 | 70.269.8 | 49.4 | 19.3 | 57.5 | 53.6 | 67.3 | |||||||||||
| LLaVA-NeXT-13B | 2880 | 69.9 73.7 | 1575.0 | 70.0 | 65.6 74.4 | 65.4 64.3 | 53.7 60.2 | 73.5 79.3 | 36.2 40.0 | 35.1 | 70.0 48.073.6 | 62.9 | 62.2 | 51.4 | 67.1 | 70.9 | 55.9 | 36.0 | 59.1 | 62.765.7 | |||||
| Cambrian-1-13B | 576 | 1,610.4 | 75.7 | 71.3 | 73.8 | 61.9 | 72.876.8 | 62.2 | 41.3 | 63.0 | 72.5 | 71.8 | |||||||||||||
| BaseLLM:Hermes2-Yi-34B | 2880 | 76.2 | 1659.0 | 80.6 | 75.365.8 | 62.477.7 | 48.0 | 43.480.5 | 68.1 | 67.6 | 51.8 | 74.1 | 78.9 | 63.837.367.2 | 71.579.2 | ||||||||||
| Mini-Gemini-HD-34B | 2880 | 76.0 76.8 | 1633.2 | 79.3 | 81.4 75.3 | 75.9 | 67.1 65.8 | 62.581.8 67.0 | 85.6 | 46.7 49.7 | 53.2 | 46.574.9 79.7 | 67.7 71.9 | 68.7 75.6 | 54.5 60.0 | 76.7 | 69.578.1 75.5 | 64.047.3 | 61.073.074.8 | 79.7 | |||||
| LLaVA-NeXT-34B Cambrian-1-34B | 576 |
| Type | Task | Description | Sources | #Samples |
| 2D | Rpeationship | Determinr te eati ostin eftaihiet sart. bottom relationship. | ADE20K | 650 |
| Object Count | Determine the number of instances present in the image. | ADE20K COCO | 788 | |
| 3D | Depter | Determine whichof te twoditinct objets is | Omni3D | 600 |
| Relative Distance | Determine which of the two distinct objects is closer to the anchor object. | 0mni3D | 600 |
| Supervision Type | Method | Architecture | Patch Size | Res. | # Tok. | Hidden Size |
| Language-Supervised | ||||||
| Language | OpenAI CLIP | ViT-L | 14 | 336 | 576 | 768 |
| DFN-CLIP | ViT-L | 14 | 224 | 256 | 1024 | |
| DFN-CLIP | ViT-H | 14 | 378 | 729 | 1280 | |
| EVA-CLIP-02 | ViT-L | 14 | 336 | 576 | 1024 | |
| SigLIP | ViT-L | 16 | 384 | 576 | 1024 | |
| SigLIP | ViT-SO400M | 14 | 384 | 729 | 1152 | |
| OpenCLIP | ConvNeXT-L | 1 | 512 | 1576 | 1536 | |
| OpenCLIP | ConvNeXT-L | = | 1024 | 1576 | 1536 | |
| OpenCLIP | ConvNeXT-XXL | = | 1024 | 1576 | 3072 | |
| Self-Supervised | ||||||
| Contrastive | DINOv2 | ViT-L | 14 | 336 | 576 | 1024 |
| DINOv2 | ViT-L | 14 | 518 | 1576 | 1024 | |
| MoCo v3 | ViT-B | 16 | 224 | 196 | 768 | |
| MoCo v3 | ViT-L | 16 | 224 | 196 | 1024 | |
| Masked | MAE | ViT-L | 16 | 224 | 196 | 1024 |
| MAE | ViT-H | 14 | 224 | 256 | 1280 | |
| I-JEPA | ViT-H | 14 | 224 | 256 | 1280 | |
| JEPA Other | ||||||
| Segmentation | SAM | ViT-L | 16 | 1024 | 1576 | 1024 |
| SAM | ViT-L | 16 | 1024 | 1576 | 1280 | |
| Depth | MiDaS 3.0 | ViT-L | 16 | 384 | 576 | 1024 |
| MiDaS 3.1 | ViT-L | 16 | 518 | 1024 | 1024 | |
| Diffusion | Stable Diffusion | VAE+UNet | 16 | 512 | 1024 | 3520 |
| Class Labels | 2.1 SupViT | ViT-L | 16 | 224 | 196 | 1024 |
| SupViT | ViT-H | 14 | 224 | 256 | 1280 |
| Method | Architecture | Patch Size | Resolution | # Tokens | Linear Probing (%) |
| EVA-CLIP-02 | ViT-L | 14 | 336 | 576 | 85.0 |
| DFN-CLIP | ViT-L | 14 | 224 | 256 | 83.6 |
| DINOv2 | ViT-L | 14 | 336 | 576 | 83.1 |
| OpenCLIP | ConvNeXt-L | - | 512 | 576 | 82.9 |
| OpenAI CLIP | ViT-L | 14 | 336 | 576 | 80.3 |
| I-JEPA | ViT-H | 14 | 224 | 256 | 77.0 |
| Supervised | ViT-L | 16 | 224 | 196 | 74.5 |
| MoCo v3 | ViT-B | 16 | 224 | 196 | 71.9 |
| MiDaS | ViT-L | 16 | 384 | 576 | 70.1 |
| MAE | ViT-L | 16 | 224 | 196 | 68.3 |
| Vision Backbone | General | Knowledge M | OCR &Chart | Vision-Centric | |||||||
| Model | JAWN | ||||||||||
| Language Supervised | Architecture | VOS | |||||||||
| OpenAI CLIP ViT-L/14@336 | 48.371,419.43 61.4559.85 62.2669.8734.50 27.80 59.8233.7331.7055.39 28.00 | ||||||||||
| DFN-CLIP | ViT-L/14@224 | 11.33 53.46 56.44 57.40 | |||||||||
| 38.781,172.5049.5349.7452.9467.7434.0027.3056.9916.75 4.8744.8111.19 | 6.67 46.97 40.29 52.08 | ||||||||||
| DFN-CLIP | ViT-H/14@378 | 36.791.091.7641.2844.32 50.4865.54 33.65 26.50 56.7615.56 2.7043.41 10.15 | 4.67 47.06 39.07 52.91 | ||||||||
| EVA-CLIP-02 ViT-L/14@336 | 45.841,325.1758.2162.9962.0368.6735.00 27.50 58.2619.40 22.50 51.08 16.3624.6752.68 52.98 54.83 | ||||||||||
| SigLIP | ViT-L/16@ 384 | 48.801,383.42 61.0263.5661.8568.91 35.2929.7057.8734.9629.6056.7328.3123.3352.68 52.95 54.83 | |||||||||
| SigLIP OpenCLIP | ViT-SO400M/14@384 ConvNeXt-L@512 | 47.571,376.75 58.76 60.59 60.9269.0134.40 26.50 58.3530.72 28.605.1028.3119.3350.71 52.3358.67 | |||||||||
| OpenCLIP | ConvNeXt-L@1024 | 47.381,404.0157.6261.90 60.3469.0633.9029.10 58.3928.0425.205.4528.4124.0054.12 53.46 48.91 39.021,139.6014.6449.5937.9165.7134.3027.3054.1332.9712.0552.61 38.36 | |||||||||
| OpenCLIP | ConvNeXt-XXL@102441.831219.4748.049.8855.09.1435.6927.6056.6716.925.0046.9040.9816.0047.324.405.75 | 9.67 47.45 52.68 38.04 | |||||||||
| Self Supervised | |||||||||||
| DINOv2 | ViT-L/14@336 | 41.181,262.66 49.62 56.80 60.30|65.10 35.00 26.40 56.4116.48 | 3.10 44.04 11.9018.67 50.20 49.43 52.25 | ||||||||
| DINOv2 | ViT-L/14@518 | 40.60|1,242.48 51.00 53.39 60.3864.55 34.50 26.20 57.5315.11 | |||||||||
| MoCo v3 | ViT-B/16@224 | 34.94 | 966.45 36.77 33.0047.3562.96 32.80 26.20 55.0516.04 | 2.60 43.81 10.31 | 2.90 44.28 10.9514.00 48.63 46.13 57.90 | ||||||
| MoCo v3 | ViT-L/16@ 224 | 34.70 | 1010.18 34.64 41.71 47.4664.70 33.70 26.30 55.0516.24 | 2.70 42.60 10.39 | 6.67 45.36 39.03 52.83 | ||||||
| MAE | ViT-L/16@ 224 | 37.691,114.0742.3035.9355.2063.5134.6026.0056.1016.11 | 2.70 43.63 10.8314.00 44.80 45.81 55.75 | 4.00 45.36 44.67 35.16 | |||||||
| MAE | ViT-H/14@224 | 38.581,083.3541.1550.9955.3064.9034.1026.0056.4915.63 | 3.20 43.98 11.00 | 12.00 46.30 47.18 54.90 | |||||||
| I-JEPA | ViT-H/14@ 224 | 38.881,132.074.6851.7455.3766.0434.2026.4056.0915.4 | 3.00 43.66 11.48 | 10.6746.0146.7453.50 | |||||||
| Other | |||||||||||
| SAM | ViT-L/16@1024 | 31.74 | 585.78 20.34 36.34 39.8565.49 34.50 25.10 53.92|16.162.70 42.37 9.25 | 2.00 44.44 35.65 50.50 | |||||||
| SAM | ViT-H/16@ 1024 | 32.37 | 648.96 22.30 36.31 40.5265.20 34.10 26.00 54.4415.56 | 2.40 42.39 8.75 | 2.00 45.36 34.83 55.25 | ||||||
| MiDaS 3.0 | ViT-L/16@384 | 35.65 | 981.36 38.57 40.93 49.0463.41 31.80 25.70 54.7216.362.60 43.19 11.24 | 6.67 44.97 38.78 53.40 | |||||||
| MiDaS 3.1 | ViT-L/16@518 | 35.44 | 983.3434.79 40.20 48.5364.60 33.90 25.00 55.1815.64 | 2.60 42.76 12.08 | 6.66 43.66 39.63 52.58 | ||||||
| Diffusion | SD2.1/16@512 | 36.591,044.2837.7142.00 48.3864.55 33.40 25.70 56.9915.56 | 3.10 43.14 10.40 | 9.33 45.88 44.68 52.40 | |||||||
| SupViT | ViT-L/16@ 224 | 40.131,197.39 46.5554.72 57.2765.9434.00 28.00 56.2216.44 | 3.10 43.52 11.82|16.67 46.67 48.49 52.75 | ||||||||
| SupViT | ViT-H/14@224 | 37.451,082.43 42.61 48.45 52.9863.51 35.29 26.50 55.7815.16 | 3.30 44.16 11.49 | 4.66 43.79 44.55 52.91 | |||||||
| Model | Architecture | AllGKOV | Model | Architecture | All G KO V | ||||||||||
| SigLIP | ViT-SO400M/14@384 | 1 | 1 | 1 | 21 | DINOv2 | ViT-L/14@518 | 1 | 1 | 1 | 1 | ||||
| OpenCLIP | ConvNeXt-XXL@ 1024 | 2 | 6 | 8 | 13 | DINOv2 | ViT-L/14@336 | 2 | 2 | 33 | 2 | ||||
| DFN-CLIP | ViT-H/14@378 | 3 | 4 | 2 | 5 | 4 MAE | ViT-L/16@224 | 3 5 | 22 | 4 | |||||
| OpenCLIP | ConvNeXt-L@1024 | 4 | 8 | 7 | 3 | 8 | I-JEPA | ViT-H/14@224 | 4 3 | 6 | 8 | 3 | |||
| SigLIP | ViT-L/16@384 | 5 | 5 | 4 | 4 | 6 | SD2.1 | VAE+UNet/16@512 | 5 7 | 9 | 9 | 5 | |||
| OpenAI CLIP ViT-L/14 @336 | 6 | 3 | 6 | 6 | 7 | MiDaS 3.0 ViT-L/16@384 | 6 | 6 | 8 | 5 | 6 | ||||
| EVA-CLIP-02 ViT-L/14@336 | 7 | 2 | 5 | 82SupViT | ViT-L/16@224 | 7 | 4 | 9 | 4 | 8 | |||||
| OpenCLIP | ConvNeXt-L@512 | 8 | 7 | 3 | 7 | 9 | MoCo v3 | ViT-B/16@224 | 8 8 | 4 | 7 | 7 | |||
| DFN-CLIP | ViT-L/14@224 | 9 | 9 | 9 | 9 | 10 | MoCo v3 | ViT-L/16@224 | 9 9 | 5 | 6 | 9 | |||
| DINOv2* | ViT-L/14@518 | 10 | 10 10 10 | 5 SAM | ViT-H/16@1024 | 10 | 10 10 1010 |
| Vision Backbone | General | Knowledge | OCR & Chart | Vision-Centric | |||||||
| 1VOS ANWNW | dAWN | ||||||||||
| Model Architecture Language-Supervised | |||||||||||
| OpenAI CLIP ViT-L/14@336 | 50.49 | 1,476.6561.96 65.45 62.78 | 69.06 35.00 29.50 58.9437.84 30.90 58.21 32.1128.66 54.90 54.14 54.60 | ||||||||
| DFN-CLIP | ViT-L/14@224 | 46.01 | 1,341.14 56.68 63.74 60.7566.96 33.80 28.6557.0423.32 23.20 52.85 18.9726.67 51.44 52.91 52.08 | ||||||||
| DFN-CLIP | ViT-H/14@378 | 51.17 | 1,426.3262.38 67.29 62.8969.0135.89 30.0060.0141.08 30.60 57.53 31.6932.67 55.95 55.46 55.00 | ||||||||
| EVA-CLIP-02 ViT-L/14@336 | 49.71 | 1,449.7864.0067.53 63.6069.9135.49 28.40 59.1624.7627.10 55.39 21.6334.67 55.69 57.83 57.75 | |||||||||
| SigLIP | ViT-L/16@384 | 50.87 | 1,424.20 59.40 65.48 62.5668.6735.99 29.70 59.2940.52 33.50 59.59 35.2028.00 53.3355.42 56.08 | ||||||||
| SigLIP | ViT-SO400M/14@384 | 53.91 | 1,455.6463.6667.62 63.7072.10 36.0929.3061.5943.7637.2061.82 40.1936.60 56.99 59.61 59.58 | ||||||||
| OpenCLIP | ConvNeXt-L@512 | 49.16 | 1,416.8760.60 63.87 61.8769.92 35.79 29.50 59.36|34.40 28.00 58.36 28.4127.33 51.90 54.64 51.80 | ||||||||
| OpenCLIP | ConvNeXt-L@1024 | 51.001,392.9258.2165.4762.8967.4334.9029.9059.1346.0825.5062.1444.1326.6755.2953.5755.08 | |||||||||
| OpenCLIP | ConvNeXt-XXL@102452.18 | 1,402.9459.4065.2162.7368.2733.1029.3059.8448.0028.0063.2748.1134.6755.9553.8355.08 | |||||||||
| Self-Supervsied | |||||||||||
| DINOv2 | ViT-L/14@336 | 41.85|1,190.81 51.83 56.90 60.3866.04 34.20 27.40 56.4116.44 | 3.30 45.12 11.79|21.33 49.67 53.91 55.33 | ||||||||
| MoCo v3 | ViT-B/16@224 | 38.881,129.3241.62 52.19 55.0365.89 33.30 28.30 56.4416.48 | 3.00 44.0911.4712.0047.5845.17 53.00 | ||||||||
| MoCo v3 | ViT-L/16@224 | 37.07 | 1015.20 37.28 48.31 52.6365.49 34.50 27.60 55.41 | 16.92 | 3.10 43.57 11.5014.67 45.49 45.17 40.75 | ||||||
| MAE | ViT-L/16@224 | 40.39 | 1,132.8043.40 55.67 57.4266.04 35.59 27.80 56.48 | 17.36 | 3.3044.5312.3016.0047.7149.24 56.75 | ||||||
| I-JEPA | ViT-H/14@224 | 40.27 | 1,207.88 45.79 54.51 56.1565.29 34.40 27.10 56.1916.20 | 3.20 43.4511.5818.00 45.88 49.57 56.58 | |||||||
| Other | 32.54 | 682.81 23.32 36.16 40.3265.20 33.20 26.50 54.21 | 15.682.50 41.76 8.98 | ||||||||
| SAM | ViT-H/16@ 1024 ViT-L/16@384 | 39.151,132.186.2151.7555.576303.7026.7056.0617.083.1043.6511.6615.3045.7544.445. | 1.33 46.80 37.66 52.90 | ||||||||
| MiDaS 3.0 | SD2.1/16@512 | 39.51 | 1,168.5240.00 53.80 55.3364.6035.00 26.10 57.1615.363.10 44.2311.0618.6747.32 48.04 53.90 | ||||||||
| Diffusion | ViT-L/16@224 | 39.12 | 1,216.1145.28 51.46 55.8864.15 34.70 26.80 55.7616.802.80 44.42 11.6111.3347.9744.62 51.60 | ||||||||
| Supervised |
| Vision Backbone | General | Knowledge | OCR& Chart | Vision-Centric | ||||||
| Model | COO | vOS | ||||||||
| Architecture Language-Supervised | ||||||||||
| OpenAI CLIP ViT-L/14@336 | 49.031,413.51 60.34 62.17 60.8169.76 36.49 29.90 58.48 | 36.80 30.20 57.63 30.9821.33 51.63 52.57 54.75 | ||||||||
| DFN-CLIP | ViT-L/14@ 224 | 45.59|1,382.75 57.36 63.26 60.5466.81 35.09 29.20 57.71 | 22.88 23.45 52.27 18.8221.33 51.31 49.59 50.63 | |||||||
| DFN-CLIP | ViT-H/14@378 | 50.621,500.45 62.64 66.44 62.5370.7535.6930.3058.78 | 39.20 29.80 56.98 31.3929.3353.59 54.96 52.58 | |||||||
| EVA-CLIP-02 ViT-L/14@336 | 47.13 | 1,362.07 62.64 63.96 61.6669.46 35.89 27.90 56.96 | 20.96 26.10 53.93 19.0720.00 53.07 56.28 58.16 | |||||||
| SigLIP | ViT-L/16@384 | 48.11 | 1,381.48 61.79 61.87 59.4570.2535.99 28.80 57.5828.76 28.20 54.90 25.60|26.00 52.29 52.89 56.33 | |||||||
| SigLIP | ViT-SO400M/14@384 | 50.411,327.79 62.13 63.92 61.3170.3836.9930.00 59.52 | 40.08 33.2060.37 36.5822.00 53.99 55.59 54.08 | |||||||
| OpenCLIP | ConvNeXt-L@512 | 48.011,366.85 59.66 62.89 61.3168.77 36.99 28.50 59.29 | 27.88 28.50 57.57 29.4816.00 53.20 53.81 55.91 | |||||||
| OpenCLIP | ConvNeXt-L@1024 | 40.29|1,084.62 12.94 51.02 49.7865.47 34.20 27.60 56.3629.9213.25 50.3743.6713.3349.08 51.85 41.58 | ||||||||
| OpenCLIP | ConveXt-XXL@102450.451405.6557.9663.5862.4168.0234.3029.4059.6242.9626.2061.8242.672.675.1649.9254.16 | |||||||||
| Self-Supervised | ||||||||||
| DINOv2 | ViT-L/14@336 | 42.64|1,283.95 54.64 59.03 60.1966.39 35.29 25.70 58.03 | 16.00 | 3.20 45.39 11.7920.00 50.59 53.5158.33 | ||||||
| MoCo v3 | ViT-B/16@ 224 | 38.50|1,159.10 40.00 51.37 54.97 | 65.25 33.7027.20 55.51 | 16.36 | 3.30 44.42 11.42|10.67 46.14 45.80 52.00 | |||||
| MoCo v3 | ViT-L/16@224 | 37.711,074.1341.1949.46 53.6163.6633.7027.40 55.83 | 17.04 | 3.40 43.84 11.98 | 8.00 46.27 48.69 45.59 | |||||
| MAE | ViT-L/16@ 224 | 39.99|1,138.3544.60 54.91 56.6965.6436.19 27.90 56.48 | 17.203.20 44.45 12.4214.00 47.32 48.83 53.08 | |||||||
| I-JEPA | ViT-L/14@224 | 39.911,180.12 44.26 52.86 55.3265.9434.40 27.00 57.16 | 15.88 | 3.20 44.3611.61 | 13.3346.27 52.19 55.83 | |||||
| Other | ||||||||||
| SAM | ViT-H/16@ 1024 | 32.18 | 40.071,183.95 47.40 53.00 56.1566.19 32.90 27.60 56.61 | 649.99 22.47 36.37 40.4664.60 32.50 25.80 54.66 | 15.802.70 42.40 8.89 | 0.00 45.62 37.02 53.08 | ||||
| MiDaS 3.0 | ViT-L/16@384 | 38.261,123.4642.04 50.66 53.6365.7433.30 24.5057.48 | 17.00 | 3.00 44.34 11.55 | 19.33 47.32 45.01 54.58 | |||||
| Diffusion | SD2.1/16@512 | 39.661,186.88 48.43 54.28 56.3565.49 33.00 28.1057.1617.562.8044.9212.2312.6747.06 43.59 51.67 | 14.52 | 3.30 43.95 10.62|10.00 43.53 48.24 54.50 | ||||||
| SupViT | ViT-L/16@224 | |||||||||
| Vision Backbone | General | Knowledge | OCR& Chart | Vision-Centric | ||||||||
| vOS | ||||||||||||
| Model Other | Architecture | |||||||||||
| OpenAI CLIP ViT-L/14@336 | ||||||||||||
| 52.901,477.15 63.15 68.49 64.2469.2134.90 28.70 61.2447.68 40.20 58.92 35.2226.00 58.82 59.65 56.16 | ||||||||||||
| DFN-CLIP | ViT-L/14@224 EVA-CLIP-02ViT-L/14@336 | 51.301,492.3565.5369.7565.1868.62 35.0029.50607829.329.9056.8721.5944.6758.9554.6655.91 | 45.801,364.9155.1563.0659.9467.2835.5929.8058.6520.4823.4052.2318.2922.6750.85 53.59 53.50 | |||||||||
| SigLIP | ViT-L/16@384 | 52.471,429.1163.5767.3463.4468.0236.0929.7061.5646.68 35.7059.8635.9332.6755.2955.24 57.00 | ||||||||||
| SigLIP | ViT-SO400M/14@38455.271489.0566.5569.5964.5870.4535.6929.2062.3451.2840.8063.2843.0238.0059.6161585.91 | |||||||||||
| OpenCLIP | ConvNeXt-L@512 | 52.761,467.3863.4066.9263.1769.1634.9029.70 58.4552.0435.4061.8738.7930.6756.2154.24 55.91 | ||||||||||
| Other | ||||||||||||
| DINOv2 | ViT-L/14@336 | 43.261,261.43 53.96 63.22 62.6165.49 34.50 27.70 56.90|15.40 | 3.40 44.8711.2226.00 54.38 53.06 56.40 | |||||||||
| MoCo v3 | ViT-B/16@224 | 39.511,175.3441.70 53.31 56.1565.25 33.40 28.30 55.7615.48 | 3.2044.4211.10|18.00 46.67 45.38 55.25 | |||||||||
| MoCo v3 | ViT-L/16@224 | 37.591075.3939.15 50.14 53.6565.49 34.60 27.30 55.4417.28 | 3.0044.2111.7014.6744.5845.3841.12 | |||||||||
| MAE | ViT-L/16@224 | 41.43|1,181.51 45.53 58.92 58.7564.65 35.00 29.20 57.1216.88 41.901,175.70 48.00 59.60 59.3564.4535.0927.6057.3216.20 | 3.10 44.6711.7418.67 49.67 53.04 56.83 | |||||||||
| I-JEPA | ViT-H/14@ 224 | 3.00 45.5011.4022.6749.9352.3859.08 | ||||||||||
| Other | ViT-L/16@384 | 38.281,065.2642.64 50.9556.1065.3935.0027.20 52.9815.962.8043.491.2312.0046.1443.41 53.90 | ||||||||||
| MiDaS 3.0 | ViT-L/16@224 | 40.011,222.4147.40 54.15 57.2664.3534.4026.40 56.0916.203.204.73117414.0046.4149.33 53.40 | ||||||||||
| Supervised |
| General | Knowledge | OCR&Chart | Vision-Centric | ||||||||||||||
| OpenAI CLIP ViT-L/14@ 336 | Backbone | Unfreeze | COO VOS | dAWI | |||||||||||||
| Method | × | 55.85 | 1,577.3369.7070.22 63.3373.6736.19 36.60 64.8049.1236.90 60.3339.7932.67 55.5666.11 59.75 | ||||||||||||||
| DINOv2 | ViT-L/14@336 | × | 45.36 | 1,373.14 57.02 64.58 61.6767.1336.19 30.70 60.62 | 19.043.4046.3913.2726.67 52.68 59.81 57.91 | ||||||||||||
| DINOv2 | OpenAI CLIP ViT-L/14@336 | ViT-L/14@336 | √ √ | 47.40 | 57.441,585.3468.6871.4763.9677.3936.0937.30 65.1259.36 48.0062.3945.2431.3356.2161.09 56.16 1,366.6561.62 69.72 63.6868.7236.2935.50 60.88 | 18.64 | 4.40 47.9214.6634.6754.64 60.98 57.83 |
| Vision Backbone | General | Knowledge | OCR&Chart | Vision-Centric | |||||
| EEEE | dAN | ||||||||
| Method SigLIP+DINOv2 | 51.61 | 1,432.0261.2865.99 63.30 | 68.82 35.69 29.40 60.0143.00 35.70 60.40 37.5430.00 53.99 55.52 53.58 | ||||||
| SigLIP+DINOv2+ConvNext | 54.521,50.51 63.83 67.97 695 | 70.40 35.9929.30 60.6948.20 36.9064.97 45.5334.6758.69 55.74 60.33 | |||||||
| SigLIP+DINOv2+ConvNext+CLIP54.74 | 1,479.4663.3267.6364.04 | 71.3935.49 29.10 59.8850.2439.6064.55 46.1232.67 58.9558.54 60.42 | |||||||
| SigLIP+ConvNext | 54.531,494.9764.6067.98 63.5871.05 34.9029.80 60.8550.64 38.00 64.5346.5232.00 57.9158.8356.58 | ||||||||
| CLIP+ConvNext | 54.451.511.0863.8367.4163.6370.8035.0930.4059.9151.3235.0064.4547.8833.3357.2556.3259.08 | ||||||||
| SigLIP+DINOv2+ConvNext-L | 53.78 | 1,450.6463.5767.7963.63 | 71.34 34.80 30.20 61.0449.32 37.70 64.05 45.8330.00 56.21 58.08 54.33 | ||||||
| SigLIP+CLIP+ConvNext-L | 54.531,507.28 63.2368.6463.6371.1035.8930.9059.9752.3638.5065.4047.9228.67 57.25 57.66 55.92 |
| Category | Datasets |
| General Conversation & VQA Data | LVIS-Instruct4V [131], SketchyVQA [129], OODVQA [129], VizWiz [51], ALLaVA [25], IDK [22], Q-Instruct [135], LAION GPT-4V [70], Hateful- Memes [66], Visual7W [159], Visualmrc [123], AlfWorld [117],LNQA [108], LLaVA150K [82], ShareGPT [27], VQAv2 [50], GQA [57], OKVQA [95], A- OKVQA [116], RefCOCO [140], VisualGenome [69], GPT-4V recorded chat |
| OCR Related Data | LLAVAR [149], ChartQA [96], DocVQA [97], DVQA [61], ArxivQA [76], AI2D [65], ScreenQA [56], SynthDog [67], IconQA [89],WTQ[106], WikiSQL [153], FinQA [31], HiTab [32], TAT-QA [158],TabMWP [88], Chart2Text _[62], VisText_[124], InfoVQA [16], ST-VQA_[17], Rendered- Text [134], OCRVQA [100], TextCaps [119], ShareGPTOCRData [27] |
| Counting Data | TallyQA [1], CLEVR [59] |
| Knowledge-Based Data | Code: Design2Code [118], WebSight [71], Datikz [14] Math:MathVision_[132],Geo170K[42],TQA [5],Inter-GPS[90], RAVEN [148], GeomVerse [64] Science: ScienceQA [91], PathVQA [53] |
| Language Only Data | Dolly [36], MathInstruct [142], WizardCoder [93], OrcaMath [99], OpenCodeIn- terpreter [152], OpenOrca [78] |
| Index | Response formatting prompts |
| 1 | Answer the question using a single word or phrase. |
| 2 | Answer the question using a single number or phrase. |
| 3 | Answer with the option's letter from the given choices directly. |
| 4 | Give the short answer directly. |
| 5 | Answer the question using a single word or phrase. |
| 6 | When the provided information is insufficient, respond with <no answer>. |
| 7 8 | Directly provide the HTML code. |
| 9 | First show your reasoning process and then give the final answer. |
| When the provided information is insufficient, respond with 'Unanswerable'.Answer the question using a single word or phrase. | |
| 10 | Answer with the letter. |
| Dataset SketchyVQA | Prompts added |
| 0ODVQA | 1 |
| 1 | |
| VizWiz | 9 |
| Q-Instruct | 1,3 |
| ChartQA | 2 |
| DocVQA | 4 |
| DVQA | 1 |
| AI2D | 1 |
| ScreenQA | 1,6 |
| CLEVR | 1 |
| TallyQA | 1 |
| PathVQA | 1 |
| MathInstruct | 8 |
| Design2Code | 7 |
| IconQA | 1,10 |
| HiTab | 1 |
| WTQ | 1 |
| WikiSQL | |
| 1 | |
| Inter-GPS | 10 |
| Visual7W | 3 |
| TQA RAVEN | 10 1 |
| General | Knowledge | OCR&Chart | Vision-Centric JAWN | |||||||||||||
| 1VOS | Wesige | |||||||||||||||
| # of Data t=150k 4015k | 53.74 | 1,512.3 67.0 68.3 61.2 | 73.435.1 | 34.362.4 | 44.639.158.5 38.530.0 55.2 61.98 54.7 | |||||||||||
| t= 250k 5218k | 54.311,475.9 67.3 69.2 61.673.4 35.9 34.5 62.446.5 36.9 59.2 38.632.0 56.663.68 57.5 | |||||||||||||||
| t= 350k 5883k | 54.271,461.9 66.2 68.9 61.673.8 36.432.8 62.546.838.3 59.339.331.3 54.9 62.68 60.4 | 36.9 63.845.8 38.6 58.439.428.0 53.6 64.60 56.8 | ||||||||||||||
| t= 450k 6383k | 54.15 | 1,534.1 67.666.361.972.9 35.1 |
| General | Knowledge | OCR&Chart | Vision-Centric | |||
| VOS ANNNN | ||||||
| exp147.49 | 1,309.10 58.00 60.10 54.0072.40 34.80 31.20 59.1034.2034.50 54.2033.0013.30 47.60 57.40 50.58 | |||||
| exp247.781,351.70 60.30 61.20 55.4072.8035.20 29.50 59.1031.20 33.40 54.20 30.5015.30 48.20 58.40 52.25 | exp348.281,299.5360.5661.79 55.7472.04 34.9032.10 59.4033.20 33.90 54.1531.9021.3048.60 58.52 49.41 | |||||
| exp447.471,288.98 58.16 61.47 55.0071.0537.1028.20 58.5033.72 34.50 55.0731.6920.66 47.06 56.30 46.58 | ||||||
| exp548.961,363.2660.48 63.18 55.9270.35 35.70 31.40 57.1932.88 34.60 54.7432.1022.7047.30 58.83 57.75 |
| Category | Test Set | #Images | Data Eng. | Cambrian10M | LLaVA-665k |
| General | MMEP | 2.374 | 0 (0.00%) | 332 (13.98%) | 82 (3.45%) |
| MMB | 4,377 | 7 (0.16%) | 1,122 (25.63%) | 533 (12.18%) | |
| SEEDI | 17,990 | 6 (0.03%) | 26 (0.14%) | 0 (0.00%) | |
| GQA | 398 | 0 (0.00%) | 1 (0.25%) | 0 (0.00%) | |
| Knowledge | SQA1 | 2,017 | 0 (0.00%) | 1,263 (62.62%) | 0 (0.00%) |
| MMMUV | 900 | 1 (0.11%) | 3 (0.33%) | 0 (0.00%) | |
| MathVistaM | 1,000 | 2 (0.20%) | 259 (25.90%) | 15 (1.50%) | |
| AI2D | 3,088 | 0 (0.00%) | 1,458 (47.22%) | 0 (0.00%) | |
| OCR & Chart | ChartQA | 2.,500 | 14 (0.56%) | 670 (26.80%) | 0 (0.00%) |
| OCRBench | 1,000 0 (0.00%) | 177 (17.70%) | 59 | (5.90%) | |
| TextVQA | 5,000 | 2 (0.04%) | 1,122 (22.44%) | 9 (0.18%) | |
| DocVQA | 5,188 | 0 (0.00%) 53 | (1.02%) | 0 (0.00%) | |
| Vision-Centric | MMVP | 300 | 0 (0.00%) | (0.00%) | 0 (0.00%) |
| RealWorldQA | 765 0 (0.00%) | 0 0 (0.00%) | 0 | (0.00%) | |
| CV-Bench | 2,638 | 0 (0.00%) | 758 (28.73%) | 336 (12.74%) | |
| Total | 49,535 | 32 (0.06%) | 7,244 (14.62%) | 1,034 (2.07%) |
| Method | General | Knowledge | OCR&Chart | Vision-Centric |
| Interpolate + MLP | 63.4 | 43.8 | 28.1 | 43.7 |
| C-Abstractor [21] | 64.4 | 42.8 | 26.1 | 44.3 |
| LDPv2 [35] | 62.5 | 43.9 | 28.7 | 43.9 |
| SVA | 65.5 | 44.5 | 31.4 | 46.9 |
| D OCR& Chart | G | OCR& Chart | Multi-agg | OCR&Chart |
| 2 52.1 | 1 | 52.4 | No | 52.4 |
| 3 52.4 | 2 52.6 | Yes | 53.3 | |
| 4 52.8 | 3 53.1 | |||
| (a) #layers | (b)# groups | (c) Multi-layer aggregation |
| Model | GQA | DocVQA | ScienceQA |
| SigLIP | 29.7% | 31.1% | 35.2% |
| CLIP | 18.5% | 13.4% | 16.3% |
| DINOV2 | 24.1% | 11.0% | 17.6% |
| ConvNext | 27.7% | 44.5% | 30.9% |
| Experiment | Backbone | Data | Adapter lrwd bs | Instruction Tuning lrwdbs | ||||||
| LLM | Vision | Adapter | Instruction Tuning | vision lr | ||||||
| 0M Adapter+737K IT | Vicuna-1.5-7B | OpenAI CLIPViT-L/14@336 | 0 | 737k | 2e-50 | 512 | = | |||
| 0.5M Adapter+737K IT | Vicuna-1.5-7B | OpenAI CLIP ViT-L/14@ 336 | 0.5M | 737k | 1e-3 05122e-5 | 0 | 512 | |||
| 1.2M Adapter+737K IT | Vicuna-1.5-7B | OpenAI CLIP ViT-L/14@ 336 | 1.2M | 737k | le-3 0 5122e-5 | 0 | 512 | = | ||
| Unfreeze Vision | Vicuna-1.5-7B | OpenAI CLIP ViT-L/14@ 336 | 1.2M | 737k | 1e-3 0512 | 2e-5 | 0 | 512 | 1e-5 | |
| Model Ensemble | Vicuna-1.5-7B | Chosen Combination | 1.2M | 737k | 1e-30 | 512 | 2e-5 | 0 | 512 | |
| Data Balance | Vicuna-1.5-7B | OpenAI CLIP ViT-L/14@336 | 0 | Mix Based on three t | = | 2e-5 | 0 | 512 | = | |
| Data Ratio | Vicuna-1.5-7B | OpenAI CLIP ViT-L/14@336 | 0 | 1350k Based on Ratio | 2e-5 | 0 | 512 | |||
| LLaVA 665K | Vicuna-1.5-7B | OpenAI CLIP ViT-L/14@ 336 | 0 | LLaVA 665K | 2e-5 | 0 | 512 | = | ||
| Cambrian-10M (Data) | Vicuna-1.5-7B | OpenAI CLIP ViT-L/14@336 | 0 | Cambrian-10M | 2e-5 | 0 | 512 | |||
| Cambrian-7M (Data) | Vicuna-1.5-7B | OpenAI CLIP ViT-L/14@ 336 | 0 | Cambrian-7M | = | 2e-5 | 0 | 512 | ||
| Cambrian-1-8B | Llama-3-Ins-8B | SVA with 4 encoders* | 2.5M | Cambrian-7M | le-4 0 512 | 2e-5 | 0 | 512 | ||
| Cambrian-1-13B Cambrian-1-34B | Vicuna-1.5-13B | SVA with 4 encoders* | 2.5M | Cambrian-7M | 1e-4 0 5122e-5 | 0 | 512 | |||
| Hermes-2-Yi-34B SVA with 4 encoders* | 2.5M | Cambrian-7M | 1e-4 0 512|2e-5 | 0 | 1024 |
| You will be given an 'answer’ and a 'gt_answer’ (ground truth answer) ,and you must reply with either CORRECT or INCORRECT based on the |
| response. Tolerate a O.O5 relative error for numerical answers. |
| answer: 25 |
| gt_answer: 29 |
| evaluation: INCORRECT |
| answer: Yes gt_answer: Yes |
| evaluation: CORRECT |
| answer: 80 |
| gt_answer: 80 |
| evaluation: CORRECT |
| answer: Ireland |
| gt_answer: Italy evaluation: INCORRECT |
| answer: UK |
| gt_answer: UK |
| evaluation: CORRECT |
| answer: 2019 |
| gt_answer: 2011 |
| evaluation: INCORRECT |
| answer: {answer} |
| gt_answer: {gt_answer} |
| evaluation: |
| LLM Backbone | System Prompt |
| Vicuna 1.5 7B | A chat between a curious user and an artificial intelligence assistant. The assstant gives helpful, detailed,and polite answers to the user's questions. |
| LLAMA-38B | You are Cambrian, a highly intelligent multimodal AI trained by NYU Vision X. As a multimodal AI,you have the ability to process and analyze images.Whenever an image is present in the conversation, very carefully examine it and consider its content when formulating your response.You should give concise responses to very simple questions,but provide thorough responses to more complex and open-ended questions. |
| Nous-Yi34B | You are Cambrian,a highly intelligent multimodal AI trained by NYU Vision X.As a multimodal AI, you have the ability to process and analyze images. Whenever an image is present in the conversation,very carefull examine it and consider its content when formulating your response. You should give concise responses to very simple questions, but provide thorough responses to more complex and open-ended questions. |
| Method Cambrian-1 8B Cambrian-18B GPT3.5 Grading78.4 65.8 82.0 38.9 78.1 71.2 67.0 69.2 49.3 63.5 â–³ | Eval Fuzzy Matching |
| Benchmark | Prompt | Example USER: <image>\\nwhich of these define dairy item\\n(A) c\\n(B) |
| AI2D | \\nAnswer with the option's letter from the given choices di- rectly. | D\\n(C) b\\n(D) a\\nAnswer with the option's letter from the given choices directly. ASSISTANT: |
| ChartQA | \\nAnswer the ques- tion using a single number or phrase. | USER: <image>\\nHow many food item is shown in the bar graph?\\nAnswer the question using a single number or phrase. ASSISTANT: |
| DocVQA | InGive the short an- swer directly. \\nAnswer the ques- | USER: <image>\\nWhat is the dividend payout in 2012?nGive the short answer directly. ASSISTANT: USER: <image>\\nIs it overcast?\\nAnswer the question using |
| GQA | tionusingsingle word or phrase. | single word or phrase. ASSISTANT: |
| MathVista | \\nFirst show your rea- soning process and then give the final an- swer. | USER: <image>\\nwhat is the total volume of the measuring cup?nFirst show your reasoning process and then give the final answer. ASSISTANT: |
| MM-Bench EN | InAnswer with the option's letter from the given choices di- rectly. | USER: <image>nWhich of the following was a dependent variable in this experiment?\\n(A) cocoon\\n(B) chrysalis\\n(C) nan\\n(D) nan\\nAnswer with the option's letter from the given choices directly. ASSISTANT: |
| MME | \\nPlease answer the question using a sin- gle word or phrase. | USER: <image>\\nIs a python code shown in the picture? Please answer yes or no.\\nAnswer the question using a single word or phrase.ASSISTANT: |
| MMMU | \\nAnswer with the option's letter from the given choices di- rectly. | USER: <image>\\nWhat causes these unusual formations on Mountain papaya? Options:\\nA. Abiotic\\nB. Confused\\nC. Bi- otic\\nD. NormainAnswer with the option's letter from the given choices directly. ASSISTANT: |
| MMVP | \\nAnswer with the option's letter from the given choices di- rectly. | USER: <image>\\nAre the buterfly's wings closer to being open or closed? Options:\\n(a) Open\\n(b) Closed\\nAnswer with the option's lettr from the given choices directly. ASSISTANT: |
| OCR Bench | \\nGive the short an- swer directly. | USER: <image>\\nwhat is writen in the image?\\nGive the short answer directly. ASSISTANT: |
| RealWorld QA | InAnswer the ques- tion using a single word or phrase. | USER: <image>\\nIn which direction is the front wheel of the car on the right side facing?\\n\\nA.Left\\nB. Straight\\nC. Right\\nAnswer the question using a single word or phrase. AS- SISTANT: |
| SQA-I | \\nAnswer with the option's letter from the given choices di- | USER: <image>\\nWhat is the name of the colony shown?nA. Maryland\\nB.New Hampshire\\nC.Rhode Island\\nD.Ver- mont\\nAnswer with the option's letter from the given choices directly. ASSISTANT: |
| SEED-I | rectly. InAnswer with the option's letter from the given choices di- | USER: <image>\\nHow many towels are in the image? Op- tions:\\nA. One\\nB. Two\\nC. Three\\nD.Four\\nAnswer with the option's letter from the given choices directly. ASSISTANT: |
| Text-VQA | rectly. \\nAnswer the ques- tion using a single word or phrase. | USER: <image>\\nwhat is the time?\\nReference OCR tokens: N, u, gO\\nAnswer the question using a single word or phrase. ASSISTANT: |
| ADE | \\nAnswer with the option's letter from the given choices di- rectly. | USER: <image>\\nConsidering the relative positions of the cush- ion and the sofa in the image provided, where is the cushion lo- cated with respect to the sofa? Select from the following choices. \\n(A) right\\n(B) left\\nAnswer with the option's letter from the |
| Benchmark | Prompt | Example |
| COCO | \\nAnswer with the op- tion's letter from the given choices directly. | USER: <image>\\nHow many trains are in the image? Select from the following choices.\\n(A) 3\\n(B) 0\\n(C)1\\n(D) 2 \\n(E)4\\nAnswer with the option's letter from the given choices directly. ASSISTANT: |
| Omni3D | \\nAnswer withthe option's letter from the given choices di- rectly. | USER: <image>\\nEstimate the real-world distances between ob- jects in this image. Which object is closer to the traffic cone (highlighted by a red box),the motorcycle (highlighted by a blue box) or the bus (highlighted by a green box)?\\n(A) motor- cycle\\n(B) bus\\nAnswer with the option's letter from the given choices directly. ASSISTANT: |