GPT Image 2 impressed me with its realism and ability to create useful graphics. In the examples I was comparing, facial contours, anatomy, and overall consistency were noticeably better than earlier image generators. That is my impression, not proof that it has eliminated every visual error.
The model is also useful for posters and advertising layouts. A short prompt can produce a recognizable design, such as the Python-themed graphic below.
Other uses include changing a photo background, creating an anime-style version of yourself, or making an avatar. The quality is exciting, but it also makes misleading images more convincing. A realistic receipt, document, or screenshot is not reliable evidence by itself.
My original post argued that restrictions were too weak. OpenAI does describe safeguards, so it would be inaccurate to say there are none. The concern is whether those safeguards and ordinary verification practices keep pace with increasing realism.
Read OpenAI's Images 2.0 announcement and the system card for the documented capabilities and safeguards.
Adapted from the original Chinese article, published on May 4, 2026.
Comments NOTHING