neither gpt nor gemini can make the wizard turned to his *right* – looking at the female on his right. no matter the prompt, the wizard will ALWAYS look to the center of the image.

Unlocking the Limitations of AI Image Generators: The Case of the Wizard’s Gaze

In the rapidly evolving field of AI-generated imagery, even the most advanced models encounter intriguing limitations. A recent exploration highlights one such challenge: regardless of prompting efforts, certain subjects within generated images exhibit persistent behavioral patterns that defy user expectations.

The specific scenario involves a wizard character turned to his right, ostensibly gazing at a female figure positioned on his right side. While intuitive prompting might suggest that the AI should produce an image reflecting this orientation and gaze direction, this is not always the case. Both GPT-based image generation and Gemini models have demonstrated a tendency for the wizard to fix his gaze toward the center of the image, ignoring explicit instructions to look to his right.

Efforts to circumvent this behavior often involve crafting carefully designed prompts. However, even with refined prompts—sometimes touted as “guaranteed” solutions—the generated images consistently fall back to the default gaze toward the center. This reveals an inherent limitation within these models: they tend to default to symmetrical or centered compositions regardless of specific directional cues.

Notably, while Gemini has shown some success in producing images aligned with the intended prompt (as showcased in an example image linked in the original discussion), GPT-based models seem to reliably disregard such directives. This disparity underscores a broader challenge within AI image synthesis: the models’ understanding of directional context and gaze is still developing.

For practitioners and enthusiasts eager to explore these boundaries, the takeaway is clear. Achieving precise gaze direction and orientation in AI-generated images remains imperfect and often unpredictable. Continued experimentation with prompt phrasing can yield some success, but a perfect, consistent portrayal may still be out of reach with current technologies.

Authors and developers working on future iterations of these models should consider refining their understanding of spatial and contextual cues, enhancing the ability to translate nuanced prompts into accurate visual representations.

Conclusion

While AI image generators like GPT and Gemini continue to improve, certain limitations—such as the inability to reliably depict directional gaze—persist. Recognizing these boundaries is essential for users aiming to harness these tools effectively. As technology advances, overcoming these challenges will likely become feasible, opening new horizons for creative and professional applications.

Have you tested similar prompts? Share your experiences and successful techniques in the comments below!