Abstract: Due to the small size, low signal-to-noise ratio, weak contrast, and susceptibility to being covered by background clutter of infrared small targets, their detection becomes exceptionally challenging. Traditional methods rely on complex handcrafted feature design and hyperparameter tuning, resulting in poor adaptability and robustness. In recent years, deep learning techniques, particularly Convolutional Neural Networks (CNNs), have been applied to infrared small target detection due to their powerful feature learning and representation capabilities. However, they require a large amount of precisely annotated data for training. This paper investigates a single-point supervised infrared small target detection method, optimizing the network structure using a label evolution framework and a dynamic upsampler. By gradually expanding point labels through intermediate predictions of the CNN, pixel-level target masks are obtained, reducing annotation costs while maintaining high detection performance. The research contributions include combining a dynamic upsampler with a weakly supervised infrared small target detection framework, leveraging the "mapping degradation" phenomenon to generate pseudo-labels through auto-regression to improve detection accuracy, and incorporating multiple datasets for model training and testing to enhance adaptability and robustness.