Flamingo: a Visual Language Model for Few-Shot Learning (Flamingo)

Flamingo processes interleaved sequences of images and text tokens through a unified transformer architecture, enabling the model to learn visual-linguistic patterns from few-shot examples without fine-tuning. The architecture uses gated cross-attention mechanisms to fuse visual features (from a pre-trained vision encoder) with language model embeddings, allowing the model to dynamically attend to relevant image regions when generating text. This enables rapid adaptation to new vision-language tasks by simply conditioning on example image-text pairs in the input context.

Flamingo implements gated cross-attention layers that selectively combine visual features from a frozen vision encoder with the language model's token embeddings. The gating mechanism learns to weight the contribution of visual information at each layer, allowing the model to decide when and how much to incorporate visual context. This is implemented as a learned linear transformation that gates the cross-attention output before residual addition, enabling fine-grained control over vision-language fusion without modifying the underlying language model weights.

Flamingo keeps the vision encoder (e.g., CLIP) frozen during training and only trains the gated cross-attention layers and language model components. This approach leverages pre-trained visual representations without catastrophic forgetting while minimizing training compute. The frozen encoder acts as a fixed feature extractor, with spatial visual features (e.g., 256 patches from a ViT) passed to the cross-attention mechanism. This design enables training on large-scale vision-language datasets without the memory and compute overhead of fine-tuning a billion-parameter vision model.

Flamingo enables few-shot learning by including example image-text pairs directly in the input context, allowing the model to infer task structure from examples without gradient updates. The model processes interleaved sequences like [image₁, text₁, image₂, text₂, ..., image_query, ?] and generates appropriate responses based on learned patterns from the examples. This is implemented through the standard transformer attention mechanism, where the model learns to recognize task patterns (e.g., visual question answering, image captioning, visual reasoning) from the example structure and apply them to new queries. No fine-tuning or task-specific training is required; the model adapts purely through context.

Flamingo generates free-form natural language responses to visual queries by leveraging the language model's text generation capabilities conditioned on visual context. The model can answer questions about images, describe visual scenes, perform visual reasoning, and engage in multimodal dialogue without task-specific output constraints. This is implemented through standard autoregressive text generation (sampling or beam search) where each token is predicted based on previous tokens and the visual context via cross-attention. The model learns to ground language generation in visual features, enabling reasoning about spatial relationships, object properties, and scene understanding.

Flamingo can follow natural language instructions that reference visual content, enabling tasks like 'describe the object in the top-left corner' or 'compare the two images'. The model grounds instructions in visual features by attending to relevant image regions via cross-attention, then generates appropriate responses. This capability emerges from training on diverse vision-language tasks and is enabled by the interleaved image-text input format, which allows instructions and visual references to be processed jointly. The model learns to map natural language spatial and semantic references to visual features without explicit supervision for instruction following.

Flamingo is trained on large-scale interleaved image-text data (e.g., web-crawled multimodal datasets) using efficient distributed training. The architecture is designed to scale to billions of image-text pairs by keeping the vision encoder frozen and only training fusion and language components. Training uses standard transformer optimization (AdamW, gradient accumulation, mixed precision) with careful data loading and batching strategies for multimodal data. The model learns from diverse vision-language tasks present in the training data without explicit task labels, enabling emergent few-shot learning capabilities.

Flamingo demonstrates cross-lingual capabilities by understanding images and generating responses in multiple languages, enabled by the language model component's multilingual training. The model can process images with text in different languages and generate responses in the same or different languages. This capability emerges from the language model's multilingual pre-training combined with vision-language alignment learned during training. The cross-attention mechanism is language-agnostic, treating all text tokens uniformly regardless of language, enabling seamless multilingual vision-language understanding.