Build2 publishers2 min readPublished
Ideogram sells 4.5 on holding an image steady through repeated edits
Ideogram 4.5 launched September 30 at $0.008 to $0.22 per image, pitched on repeated edits that leave the rest of the picture untouched. No independent test of that claim has been published, so teams paying per edit pass have to measure drift on their own assets.
The Engineer · Build desk
What happened
- Ideogram's launch thread shows GPT Image 2.5 Sunburst, Nano Banana Pro and Nano Banana 2 becoming unusable within a few edits while 4.5 stays clean.
- Ideogram's listed uses include recoloring products, repairing old photographs in stages and swapping furniture in a room photo without changing the architecture.
- Besides Ideogram's own app and API, the model runs on launch partners including Picsart, fal, Krea, Runway, Pika, Gamma, Luma and ComfyUI.
- CEO Mohammad Norouzi was a senior staff research scientist at Google Brain, where his projects included Google's Imagen text-to-image system.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- cost If each pass bills as one image, a ten-edit chain costs $0.08 at the cheapest tier and $2.20 at the most expensive, so the tier that actually holds an image steady sets the budget for long edit work.
- contradiction The Decoder reports that early testers confirm strong consistency across edits; with the testers unidentified and no figures attached, Runtimewire's view that independent performance data is unestablished still holds.
- decision Teams already working on a partner platform can trial 4.5 inside the pipeline they run today, so a side-by-side drift test does not require adopting a new tool first.
- constraint Until the promised open weights ship, every pass in an edit chain is a hosted per-image call, and teams cannot run or freeze the model on their own hardware.
The test I would run needs no human judge. It starts from a real product image and the chain of edits a designer would actually request. Each instruction gets a mask over the region it is allowed to change. After every round, the pixels outside the mask are diffed against the original. Ideogram says 4.5 reduces pixel shifts, color changes and texture artifacts across repeated edits [2]. Each of those shows up in that diff as a number per round.
The same chain then goes through the model already in production. For the launch thread's verdict on GPT Image and Nano Banana [6] to transfer, Ideogram's chosen images and edits would have to resemble a team's own. That means the same kind of change and a chain of similar length. A vendor side-by-side contains the cases the vendor chose to post. Runtimewire notes that the thread does not specify an independent evaluator, a benchmark protocol or a measured failure rate [7].
I like the shape of the API. The precise-edit endpoint returns an image at the input's dimensions, and a second endpoint combines generation with editing [3]. Same-size output drops back into the layout slot the original came from. It also lets the diff line up pixel for pixel with no rescaling. The diff should run at production size. Ideogram's model page edits a 4,016-by-6,016-pixel source [5], about 24.2 megapixels [3], while the launch post advertises native 2K output [4]. If untouched pixels get resampled on the way between those sizes, the diff will show it.
The quality tier has to stay fixed for the whole run. The four modes span a 27.5x price range [1]. A stability result at one tier is not evidence for another.
I would also put at least one text edit in the chain. Ideogram built its earlier reputation on text rendering and graphic design [12]. It lists translating stylized lettering while keeping the design as a 4.5 use case [13].
What to watch
- Release of the promised open weights, which would let teams run long edit chains on their own hardware against a fixed model version.
- An independent evaluation of 4.5 with a published protocol and a measured failure rate across long edit sequences.
- Whether Ideogram discloses which quality tier its side-by-side demos used.