BLIP-2

BLIP-2 connects pre-trained, frozen image encoders (CLIP ViT, EVA-CLIP) to frozen LLMs (OPT, Llama) using a learnable Querying Transformer module that acts as a bottleneck. This architecture keeps both the vision and language models frozen during training, requiring only the lightweight Q-Former (~5% of total parameters) to be trained on multimodal data. The Q-Former learns to extract task-relevant visual tokens and project them into the LLM's embedding space through cross-attention mechanisms, enabling efficient knowledge transfer without catastrophic forgetting.

BLIP-2 performs VQA by encoding images through the frozen vision encoder, extracting visual tokens via the Q-Former, and feeding them to a frozen LLM that generates answers in natural language. The architecture supports zero-shot VQA without task-specific fine-tuning by leveraging the LLM's instruction-following capabilities. During inference, the system constructs prompts like 'Question: [Q] Answer:' and uses the LLM's text generation to produce answers, enabling generalization to unseen question types and visual domains without retraining.

BLIP-2 evaluation is standardized through LAVIS's metrics system, which computes task-specific metrics (BLEU, CIDEr, SPICE for captioning; VQA accuracy, F1 for VQA; Recall@K for retrieval) using established implementations (COCO evaluation server, VQA evaluation toolkit). The system provides a unified evaluation interface that works across different tasks and models. Metrics are computed on validation sets during training and logged to tensorboard. The evaluation pipeline supports distributed evaluation across multiple GPUs.

BLIP-2 generates image captions by encoding images through the frozen vision encoder, extracting visual tokens via the Q-Former, and prompting the frozen LLM with instructions like 'A short image description:' or 'Describe the image in detail:'. The LLM's instruction-following capabilities enable controllable caption generation (short, detailed, factual) without task-specific fine-tuning. The system leverages beam search or nucleus sampling during decoding to generate diverse, coherent captions that align with the visual content.

BLIP-2 extracts aligned visual-semantic embeddings by passing images through the frozen vision encoder and Q-Former, then optionally through the LLM's embedding layer. The LAVIS library provides a unified feature extraction interface via `extract_features()` that works across different models (BLIP, BLIP-2, ALBEF, CLIP) with minimal code changes. Features can be extracted at multiple levels: Q-Former output tokens (visual-semantic aligned), LLM embedding space, or intermediate layer activations. These embeddings enable downstream tasks like image-text retrieval, clustering, and similarity search.

BLIP-2 integrates with LAVIS's registry-based architecture that centralizes model and dataset management. The `load_model_and_preprocess()` function uses a hierarchical registry to instantiate models, load pre-trained checkpoints from Hugging Face or Salesforce servers, and initialize data processors (image normalization, text tokenization) in a single call. The registry pattern enables extensibility — new models, datasets, and processors are registered via YAML configs and Python classes without modifying core code. Checkpoints are automatically downloaded and cached locally on first use.

BLIP-2 training is orchestrated through LAVIS's runner system, which abstracts the training loop, loss computation, and evaluation across different tasks (VQA, captioning, retrieval, classification). The runner loads task-specific configs (learning rate, batch size, loss weights), manages distributed training via PyTorch DistributedDataParallel, handles mixed-precision training with automatic mixed precision (AMP), and logs metrics to tensorboard. The pipeline supports multi-task learning by combining losses from different tasks with configurable weights. Training is reproducible via seed management and config-based hyperparameter specification.

BLIP-2 performs image-text retrieval by extracting aligned embeddings from both modalities (images via vision encoder + Q-Former, text via LLM embeddings) and computing similarity scores. The system uses contrastive learning objectives (InfoNCE loss) during training to align visual and textual embeddings in a shared space. At inference, retrieval is performed via cosine similarity between image and text embeddings, enabling both image-to-text and text-to-image search. The Q-Former acts as a bottleneck that forces visual information to be compressed into tokens that align with the LLM's semantic space.

InstructBLIP extends BLIP-2 by adding instruction-tuning on top of the frozen Q-Former and LLM, enabling the model to follow complex visual instructions without task-specific fine-tuning. The system uses a two-stage training approach: first, BLIP-2 is trained with contrastive and generative objectives; second, the frozen BLIP-2 is instruction-tuned on diverse vision-language tasks (VQA, captioning, visual reasoning, object detection) using instruction-following datasets. This enables zero-shot generalization to new visual tasks by phrasing them as natural language instructions.

BLIP-2 inference is optimized through LAVIS's support for mixed-precision inference (FP16), which reduces memory usage and latency by ~40% with minimal accuracy loss. The system supports batched inference to amortize model loading overhead across multiple images. Optional quantization (INT8, dynamic quantization) further reduces memory footprint for deployment. Inference is implemented with torch.no_grad() context to disable gradient computation, and the system supports both GPU and CPU inference (though GPU is strongly recommended for latency).

BLIP-2 integrates with LAVIS's dataset system, which provides unified loading for 20+ multimodal datasets (COCO, Flickr30K, Nocaps, Visual Genome, SBU, etc.) via the `load_dataset()` function. The system automatically downloads and caches datasets, applies dataset-specific preprocessing (image resizing, text tokenization), and supports data augmentation (random crops, color jittering, text masking). Data processors are registered per-dataset and handle modality-specific transformations. The dataset system returns PyTorch DataLoaders with configurable batch sizes and sampling strategies.