GPT Image 2 preview

A Reddit user's viral post has provided a potential sneak peek at the next generation of OpenAI's image generation technology. The user, Groundbreaking_Tap85, reported having temporary access to a model they identified as 'GPT Image 2' for approximately two days. During this window, the model produced outputs that were consistently more realistic, detailed, and coherent than the standard DALL-E 3 model, suggesting a significant underlying upgrade. The key evidence was a side-by-side comparison using the exact same prompt: 'Amateur photograph of an elderly couple sat inside of a Yorkshire pub, amateur composition, candid.' The image from the preview model displayed a natural, photographic quality with nuanced lighting and authentic details, while the output from the reverted standard model returned to a more stylized, less realistic aesthetic.

This incident suggests OpenAI is actively testing a substantially improved image model, possibly internally referred to as GPT Image 2, in a limited, rolling fashion. The brief exposure and subsequent rollback indicate it was likely an A/B test or an accidental early deployment to a subset of users. The dramatic improvement in photorealism and prompt adherence, as evidenced by the 'amateur photograph' descriptor being accurately rendered, points to advancements in the model's understanding of physical scenes, lighting, and compositional nuance. For the AI community, this leak acts as a strong signal of the rapid progress being made behind the scenes, setting new expectations for the quality and fidelity of consumer-facing AI image tools in the very near future.