DeepSeek Tried to Paint–and Hit a Wall (www.stork.ai)

DeepSeek V4.1 Flash recently faced an intriguing challenge: recreating a portrait not through traditional pixel generation, but via a web-based version of Microsoft Paint. This task required the multimodal AI model to convert visual inputs into a series of actions on a digital canvas, including selecting colors and applying brush strokes. The experiment aimed to test whether AI can move beyond simple image generation to perform interactive graphical tasks, revealing its ability—or limitations—in interpreting and executing visual commands in real-time. The outcome of the test was a stylized yet abstract portrait that demonstrated DeepSeek’s capacity for basic tool use but highlighted significant shortcomings in detailed layering and refinement. Although the model successfully captured general shapes and color impressions, it failed to execute finer details such as facial features, which are crucial for realism. The experiment emphasized the challenges of agentic motor control and spatial reasoning necessary for complex tasks, revealing that while DeepSeek can manage basic interactions, it struggles with the intricate planning required for nuanced, iterative artistic composition. This underscores a current hurdle in AI’s evolution, particularly in tasks that demand a higher level of precision and control.