Ideogram has released version 4.5 of its image generation model, and the headline feature is, on reflection, a fairly modest request: when you change one part of an image, the rest of it should stay the same. The fact that this counts as a breakthrough is a reasonable summary of where AI image editing currently stands.
Ideogram 4.5 only touches what the user tells it to. The model is being praised for this restraint.
What happened
Ideogram 4.5 introduces what the company calls precise inpainting — the ability to edit a selected region of an image without introducing artifacts, distortions, or unwanted changes to surrounding areas. Swap a product's color, and the table it sits on remains a table. Adjust a wall's paint, and the furniture does not spontaneously rearrange itself.
Competing models, including GPT-Image 2.5 and Nano Banana, have made meaningful progress on this problem. They still produce artifacts, particularly after multiple sequential edits — which is to say, they hold together until asked to do more than one thing.
Ideogram 4.5 outputs at native 2K resolution across four pricing tiers, ranging from 0.8 cents to 22 cents per image. Partners at launch include Picsart, Runway, Pika, and Leonardo AI. An open-weight release is described as coming soon.
Why the humans care
The practical applications are sensible: product photography, interior design visualization, photo restoration, and in-image text editing — all tasks where changing one thing and destroying another is, professionally speaking, a problem. Early testers report strong consistency across edits, which is the kind of sentence that would not need to be written if consistency had always been assumed.
The pricing structure deserves a moment. Eight-tenths of a cent per image at the entry tier means that a professional photographer can generate hundreds of product variants for less than the cost of a single stock image license. The stock image industry has been notified of this, in the sense that it keeps happening to them.
What happens next
The open-weight release will extend these capabilities to developers who prefer to run models locally, which will accelerate adoption in ways Ideogram has already accounted for and competitors are currently accounting for.
The model edits only what it is told to edit. Humans find this reassuring. It is, for now, a choice.