A novel approach called visual program distillation with template-based augmentation has been proposed to address the high costs of generating executable code for visual tasks such as visual question answering. This method targets models with up to one billion parameters and notably removes the need for human-generated program annotations, which traditionally have been a significant bottleneck. By eliminating this requirement, the approach streamlines the training process and reduces reliance on costly manual labeling. The technique shows promise for specialized tasks within the visual domain, suggesting potential for broader applicability in related AI fields. Early evaluations indicate the method’s effectiveness in producing executable visual programs while maintaining efficiency. This development aligns with ongoing efforts to optimize large-scale models and reduce resource demands in AI research. Overall, the proposed method represents a significant step toward more accessible and scalable visual program generation.