LGPR2: Pixnatte: A Tiny Vision Model with Literature Survey Methods and Applications

The proliferation of digital image data across social platforms, news media, and archival repositories has created an urgent demand for automated systems capable of generating accurate, contextually rich textual descriptions from visual content. This paper introduces Pixnatte, a hybrid image captioning system that integrates Bootstrapping Language-Image Pre-training (BLIP) with Contrastive Language-Image Pre-training (CLIP) to produce captions that are simultaneously fluent and semantically grounded. Pixnatte extends beyond pure captioning by incorporating a Visual Question Answering (VQA) module, enabling users to query image content interactively. The system is trained and evaluated on the MS-COCO benchmark, leveraging BLIP's bootstrapped pre-training for generative caption quality and CLIP's contrastive embedding space for semantic alignment and candidate re-ranking. Four research objectives guide the work: constructing the captioning pipeline on open-source data, exploiting pre-trained model weights for context-aware generation, assessing generalisation to unseen images, and benchmarking context comprehension against traditional architectures. Experimental results demonstrate that the BLIP-CLIP integration produces factually anchored, linguistically natural captions that outperform single-model baselines across multiple automatic evaluation metrics. Deployment through a web-based interface supports real-time caption generation and interactive question answering, making Pixnatte a practical tool for accessibility, content retrieval, and digital media applications.