Image To Text Captioning Dataset With Egocentric Context And Temporal Grounding

via “vision-language image captioning with conditional generation”

image-to-text model by undefined. 14,17,263 downloads.

Unique: Uses a lightweight query-based attention mechanism (BLIP architecture) that decouples image understanding from text generation, enabling efficient fine-tuning and inference compared to end-to-end vision-language models like CLIP+GPT. The 'large' variant (350M parameters) balances quality and computational efficiency through knowledge distillation from larger models.

vs others: Faster and more memory-efficient than ViLBERT or LXMERT for caption generation while maintaining competitive quality; outperforms CLIP-based caption generation in semantic coherence due to explicit decoder training on caption datasets.