ChatGPT and Gemini: When Image Restoration Turns into a Nightmare

ChatGPT and Gemini: When Image Restoration Turns into a Nightmare

On the platform X, a strange trend has emerged: simply asking ChatGPT to restore an image without attaching a file triggers a series of unexpected results. ChatGPT Images 2.0, an enhanced version of OpenAI's tool launched in April 2026, is at the heart of this experiment. Instead of restoring a non-existent image, the model generates absurd and grotesque scenes, often in a retro style.

When deprived of an input image, the model switches to pure generation mode. It relies on the textual cues provided, often marked by implicit oddity, to create visuals that can be disturbing. The problem primarily lies in the user interface (UX): ineffective detection of the absence of an image and a lack of clear feedback to the user.

A particular prompt, spotted in May 2026, caught attention. It invites ChatGPT to restore an image without an attached file, producing results that are as strange as they are varied. Although ChatGPT Images 2.0 has been praised for its flexibility compared to Google's Gemini, it reveals an unexpected facet of its capabilities here.

Results as Strange as They Are Unexpected

Users who tested this prompt were surprised by the diversity of the generated images. Among the most notable creations are a meme of a stuffed fox, a half-man half-deer figure sitting on a toilet, and a rabbit mascot in an 80s scene. These results illustrate the AI's ability to produce images from textual prompts without visual support.

For those looking to experiment with a more moderate version, another prompt is circulating on social media. This one requests a restored image that is "family-friendly" and adheres to guidelines. The results include scenes like a shirtless man with a cat on a bed, showcasing another facet of the AI's creativity.

Gemini: Equally Surprising Creations

Gemini, Google's competing tool, is not to be outdone. It can also generate unexpected works, such as representations of tacos or Victorian-era family photos, often featuring animals. These creations highlight Gemini's ability to produce images that, while surprising, remain within the realm of the possible.

The Reasons Behind These Oddities

Understanding why these tools produce such images remains complex. The AI models used for image restoration and generation are not distinct but share similar technological foundations. They rely on diffusion models or transformers, conditioned by text and, potentially, by an encoded image.

When the input image is absent or poorly transmitted, the model enters "pure generation" mode while receiving restoration instructions. The textual prompts, often laden with references to strange or shocking content, then influence the creation of distressing or absurd scenes.

An inappropriate adjustment of parameters can lead to unexpected behaviors. The problem seems to primarily reside in the user interface and user experience (UX). Ineffective detection of the absence of an image and the lack of clear messages when the AI receives nothing perpetuate the illusion of restoration, while the scene is entirely fabricated.