OpenAI claims ChatGPT Images 2.5 cuts generation latency by up to 50%
- OpenAI says ChatGPT Images 2.5 is its new state-of-the-art image model, with sharper details, more precise editing, and faster generation than earlier releases.
- OpenAI says it reduced image-generation latency by up to 50% versus Images 2.0, while improving natural lighting, textures, and preservation of subjects in reference photos.
- OpenAI says the model follows multi-turn editing instructions more reliably, aiming to retain requested changes across successive image edits.
- ChatGPT is adding Sketch, a feature that lets users draw directly in the chat as a reference for image creation.
- OpenAI says people create more than 3 billion images per week across ChatGPT Images and the GPT-Image API models.
Hacker News opinions
The motocross shot at 0:19 still seems to have the noise-gradient problem, so I do not think they fixed it.
I see AI-generated menu posters from local shops all the time now, and I hate how generic they look.
I actually miss the bad Photoshop work. These AI signs look worse, even when the old food photos did not resemble the dish.
The problem is not only small shops. Companies that can afford ad designers are also publishing error-filled AI images.
One fried-rice place near me posts simple pencil cartoons of themselves cooking. It is a welcome break from identical AI posters, and it has earned my business.
The composite party-photo example loses tiny details: the middle man's teeth change, the left man's grip has odd fingers, and there is a disembodied hand by his shoulder.
OpenAI's image models have lost fine details during edits for me, unlike Nano Banana or Flux, which preserve the input geometry better. This release sounds similar.
The unchanged dog shadow is another obvious error. The dog's body changes substantially while the shadow nearly keeps its original shape.
I am surprised OpenAI does not address how strongly ChatGPT is associated with AI menu slop. Many generated menus and posters have the same immediately recognizable look.
Small restaurants were often using stock kebab shots or terrible flash photography before this. Their old alternatives were not necessarily more authentic.
I worry less about people faking a party photo than about fabricated images with a 2006 timestamp. Printed fakes could get mixed into old family photos and become hard to verify later.
The examples themselves seem aimed at deception: three selfies become a fake party, an unmade bed becomes a clean one, and a property listing can conceal problems.