This purely visual large-model framework learns visual representations through sequence modeling, without requiring any language data. The method first uses a VQGAN tokenizer to independently encode each image into 256 discrete tokens, converting images into serialized symbolic representations. The tokens from all images are then concatenated into a unified visual sentence sequence, and an autoregressive Transformer based on the LLaMA architecture is trained by minimizing the cross-entropy loss of next-token prediction, enabling the model to infer task relationships from the context of the image sequence. Finally, during inference, visual prompts (e.g., partial image sequences) are designed to guide the model in generating subsequent content, enabling zero-shot solutions for diverse visual tasks such as image completion, segmentation, and prediction. The entire pipeline achieves unified representation of visual data and flexible task generalization.