The Pulse
Ideogram 4.5 Takes Aim at Drift in Repeated Image Edits
Ideogram introduced version 4.5, an image-editing model designed to preserve details across repeated changes. Its Precise Edit API keeps untouched pixels and returns results at the source image’s dimensions, while the company’s comparisons
AI.info Team · Written with AI ·
“Preserving details across AI image edits is hard. When we spoke with creative professionals and enterprise teams about AI editing, their biggest complaint was that image quality degraded with every edit.”
Mohammad Norouzi
Ideogram announced version 4.5 on September 30 with a claim focused on what happens after the first image: the model is designed to keep details intact as users make successive edits. The company says repeated edits can introduce pixel shifts, color changes and texture artifacts, and that 4.5 reduces that drift. It frames the model around controlled revision rather than generation from scratch.
Two editing modes, two API endpoints
Ideogram offers Precise Edit for targeted changes and a separate generate-plus-edit mode for revisions that give the model more freedom to rework an image. In the API, Precise Edit returns an image at the input’s width and height, and the documentation says pixels outside the edit are copied exactly from the source. Developers can add a mask to limit the affected area or supply up to four reference images.
The distinction gives users a choice between preservation and reinterpretation. A product team changing a shoe’s color, for example, can make that adjustment and then request another change to the sole or background; an editor can instead choose generate-plus-edit when a prompt calls for broader changes. Ideogram’s API guide shows a three-image sequence of running shoes through successive edits, with the shoe shape and position held constant.
Ideogram’s comparison is the evidence it offers
The model page compares the same edits across Ideogram 4.5, GPT Image 2.5 Sunburst, Nano Banana Pro and Nano Banana 2. Ideogram says the competing outputs become unusable within a few edits while its own results stay clean. The page does not describe an independent evaluator, a scoring method or a measured drift rate, so the comparison demonstrates the company’s product claim rather than establishing a neutral ranking.
Examples on the page cover color and lighting changes, text modifications, product photography, interiors, architecture and restoration of old photos. Ideogram also shows an edit using a 4,016-by-6,016-pixel source image, labeled 24.2 megapixels, and says the model can work on high-resolution images without downsizing them. That is a displayed example, not a stated maximum input size.
Four quality modes start at less than a cent
Ideogram says 4.5 offers four quality modes, priced from 0.8 cents to 22 cents per image, with native 2K resolution. The model is available through Ideogram’s product and API, as well as named launch partners including Picsart, Krea, Runway, Pika, fal, ComfyUI, Gamma and Luma. The company also says open weights are coming, but gives no release date in its announcement.
The per-image range makes repeated editing a cost decision as well as a quality test, especially for teams producing multiple versions of a campaign asset. The more consequential question is whether the model keeps the details users need across long edit chains, not just the selected examples on its launch page. Ideogram’s announcement provides a workflow and a price range; it does not provide independent measurements of performance over repeated edits.