Ideogram 4.5 claims to fix the quality drift that plagues AI photo editing

Ideogram released version 4.5 of its image model on September 30, built around a problem anyone who’s edited an AI-generated image repeatedly already knows: quality degrades fast. Ask a model to recolor an object, then relight it, then fix something else, and most tools compound small errors with every pass until the image looks noticeably worse than where it started. According to the company’s own announcement, 4.5 is built specifically to avoid that drift, changing only the part of the image a user asks it to touch while leaving everything else untouched at full sharpness.

The feature list reads like a working designer’s wish list rather than a tech demo: recoloring a product with accurate new shadows and reflections, swapping furniture in a room while keeping the architecture intact, translating text on a sign or menu while preserving the original font and layout, and restoring old photographs in stages rather than one irreversible pass. Ideogram’s side-by-side comparisons claim GPT Image and Google’s Nano Banana models accumulate visible artifacts and color shifts within a handful of edits, while 4.5 holds up across the same sequence. That’s Ideogram’s own framing of the comparison, not an independently run benchmark, so the specific margin deserves some skepticism even though the underlying problem, multi-turn editing drift, is real and well known across the field.

The model is live now on Ideogram’s own site and through its API, priced across four quality tiers running from under a cent to 22 cents per image at native 2K resolution, and it’s already rolled out to launch partners including Picsart and Krea AI. Open weights are coming later, though Ideogram hasn’t given a firm date beyond saying it’s planned. The company was founded by four former Google Brain researchers who worked on Imagen before leaving to start it in 2022, a lineage that shows up in how much of Ideogram’s pitch is built around solving a specific technical failure mode rather than chasing more impressive first-generation output.