What held up across edits in December's ChatGPT Images
David Gewirtz uploaded a photo of himself walking in a park and asked ChatGPT to put him in a red shirt with a printed slogan. The shirt and lettering came through. The output changed the framing to a square, altered his expression, and removed a microphone. Two follow-up instructions about aspect ratio failed in his test. The results showed how much a straightforward wardrobe edit could change elsewhere in the frame. Gewirtz's before-and-after images
The December release
On December 16, 2025, OpenAI began rolling out a new image model in ChatGPT and introduced a dedicated Images space with preset styles and example prompts. The launch emphasized narrower edits, detail preservation, and faster rendering. OpenAI released the model to developers as GPT Image 1.5 through the API. The Images workspace and the automatically saved My images gallery belonged to the ChatGPT app. Images 2.0 followed on April 21, 2026, when OpenAI introduced image generation with thinking and tool use, including web search. December launch · ChatGPT release notes · April Images 2.0 system card
From an uploaded photo to an invitation
When Gewirtz asked to move the subject to Vasquez Rocks after re-uploading the photo, the first attempt used the earlier red-shirt output. He started a new chat with the original and got the subject he intended. After several scene and costume edits, he asked for a December 17 party invitation. The first design used the requested date but felt generic to him. A follow-up asking for copy that fit the setting, followed by a name correction, produced the version he showed. The article documents an editorial trial and leaves distribution of the invitation unconfirmed. Full prompt and output sequence
How much of the image to select
DataCamp's tester gave a generated polar bear vintage sunglasses by selecting part of the image. With the head selected, the model effectively replaced it with glasses. Selecting the eyes and repeating the instruction produced the intended placement. Moving the bear into a Tokyo subway kept its pose but carried over lighting from the forest scene. Separate edits to the shadow and the reflections helped. Repeated attempts to fix who stood in front of whom added a new person to the scene. The guide shows why a recognizable subject can still look out of place after a background change. DataCamp's examples and revisions
When transparency became a separate step
On December 17, a Reddit poster attributed style drift to an early GPT Image 1.5 character-edit test and showed attempts to derive a back view from existing front views. They pointed out a hammer inconsistency in one output. After dropping the request for a transparent background, they reported better-looking results and suggested removing the background separately. The post records the change that helped in this trial. Its explanation about background post-processing remains the author's hypothesis. Original submission and follow-up
The three accounts show where review mattered after an apparently successful edit. Gewirtz checked the original framing and visible details, DataCamp checked whether the light and perspective fit the new setting, and the Reddit poster checked the character's style and prop across views. Each workflow exposed a different kind of drift within the same launch period. Photo and invitation · Selected-region edits · Character reference edits