Ideogram 4.5 edits images repeatedly without drift, company says
Ideogram says its new edit model avoids the artifacts that build up over repeated edits, and fal says 94 to 99 percent of untouched pixels survive ten edits.
Ideogram released Ideogram 4.5 on Wednesday, an image-editing model it calls the most precise edit model. In its launch post the company said leading models add artifacts, pixel shifts and colour changes with each edit, and that 4.5 eliminates artifact buildup, which it says makes multi-turn editing possible. It is live in Ideogram, the API and launch partners.
Launch partner fal says that after ten edits in a row, 94 to 99 percent of everything the user did not ask to change stays identical. fal also says text is rendered exactly as written in any language, that edits can be guided by a mask or up to four reference images, and that text-to-image and editing share one endpoint. Those are vendor claims that no outside tester has confirmed.
https://x.com/fal/status/2105341213138199020
TestingCatalog reported a price of 0.8 to 22 cents an image at native 2K resolution, and said Ideogram plans to publish open weights soon. No date or licence has been given, so the weights are not available yet. Krea also lists the model, and said it is built to solve repeatable edits that usually end in artifacts, pixel shifts and colour drift.
https://x.com/testingcatalog/status/2105334260013908190
Arena, which ranks models by crowd vote, says Ideogram 4.5 entered the top 20 of its Image Edit Arena with 1351 points, 18th overall. It placed 14th in 3D imaging and modeling and 15th in product and commercial design, and Arena said its art ranking is coming. An 18th-place finish means the model does not lead the board, and a single score may not isolate the repeated-edit claim.
Justine Moore posted a test in which she took an Olipop ad and asked for a series of precise edits, like a creative team exploring variations or localising content. She called the handling of multi-turn changes without warping and artifacts incredible. That is one user's test, and the paper has not run the model itself.