Modern research does not train Vision Language Models from absolute scratch due to the immense computational cost and data requirements. Instead, the standard approach involves taking a pretrained text-only language model and finetuning it to acquire vision capabilities. This method is significantly more efficient and avoids the performance degradation often seen in training both modalities simultaneously. The architecture typically consists of three main components: an image backbone, an adapter layer, and a language model. The image backbone, often a frozen Vision Transformer, converts raw pixel inputs into a sequence of vector embeddings. Keeping this component static prevents overfitting, as the available image-text datasets are usually much smaller than the massive corpora used to pretrain the vision model. The most critical and complex module is the adapter layer, specifically the Query Transformer or Q-Former. While the image backbone produces embeddings unaware of language, the Q-Former bridges this gap by grounding pixel data into text-compatible representations. The process begins by passing image embeddings from the frozen backbone and learnable query vectors into a BERT-based model. Through a combination of self-attention and cross-attention layers, the model forces the query vectors to attend to specific features within the image embeddings. This effectively condenses a long sequence of visual tokens into a shorter, semantically rich representation that aligns with textual concepts. Training involves loss functions such as Image-Text Contrastive Loss, which aligns global representations, or Image-Text Matching Loss, which enables fine-grained verification between image and text details. Once the visual information is condensed by the Q-Former, it is passed through a Multi-Layer Perceptron (MLP) to match the dimensionality of the target language model. This sequence is then inserted between the system prompt and user query tokens. To ensure the model can generate text based on the visual input, the architecture relies on autoregressive generation where the language model treats the visual embeddings as a prefix. To make this process feasible on consumer hardware, the training strategy freezes the original language model parameters. Instead, only Low-Rank Adaptation (LoRA) matrices are trained alongside the Q-Former and the MLP adapter. This allows the model to learn how to interpret visual tokens and integrate them with existing world knowledge without requiring massive compute resources. The final model successfully learns to process image inputs and generate coherent textual descriptions or answers, demonstrating that transferring vision capabilities to a small text model is a viable and efficient path for developing multimodal AI.