| Image Encoder | |
| 1. A ViT-H/14 variant of the vision transformer with 630M parameters, trained on 2.5B image-text pairs | |
| 2. Processes images of resolution 224x224, divided into 16x16 patches of 14x14 pixels each | |
| 3. Leverages multi-layer feature extraction utilizing features from the 4th, 8th, 16th, 24th, and 31st layers, in addition to the final layer features | |
| 4. 8 gated self-attention layers, resulting in a total of 40 transformer blocks, and with 850M total parameters | |
| 5. Encoder generates a 7680-dimensional representation for each patch, producing a total of 256 patches | |
| Image Adapter | |
| 1. Cross-attention layers b/w visual and language model token representations, applied after every fourth self-attention layer in the language model, utilizing GQA | |
| 2. Initial pre-training: Trained on ≈6B image-text pairs, with images resized to fit within four tiles of 336x336 pixels, arranged to accommodate various aspect ratios. | |
| 3. Annealing: Training continues on ≈500M images, increasing per-tile resolution to enhance performance on downstream tasks | |
| Video Adapter | |
| 1. Input: Up to 64 video frames, each encoded | |
| 2. Temporal modeling: Combines 32 consecutive frames and adds video cross-attention layers | |
| 3. Aggregator: Implemented as a perceiver resampler | |
| 4. Parameters: 0.6B and 4.6B for Llama 3 7B and 70B, respectively |