Ideogram 4.5 vs ChatGPT Images. What 100 edits show.

Is Ideogram 4.5 better than ChatGPT for editing images?

For editing the same image more than a few times, yes. GPT Image 2.5, the model behind ChatGPT Images, redraws the whole picture on every edit, including the parts you did not ask it to touch. Each redraw adds a small color shift and some noise, and when you edit the result again, those errors stack. Ideogram 4.5 restores the pixels you left alone.

One edit looks fine in either. Ten edits in a row and ChatGPT's sky starts to blotch. ChatGPT still drew printed text on the shirt more cleanly, so the better choice depends on the step: ChatGPT to create a new element, Ideogram or your own composite to keep the rest of the image intact.

Ideogram vs ChatGPT: the 100-edit test

Alec Wilcock, who runs social media at ElevenLabs, posted the comparison on X on October 1, 2026, the day after Ideogram released 4.5.

He started with one photo of himself on a balcony above a harbor and wrote 100 prompts that each change only the T-shirt, such as "Change his T-shirt to a black T-shirt with the number 42 printed large in white in the middle of the chest." Each model edited its own previous output, never the original. GPT Image 2.5 Sunburst ran at high quality and Ideogram 4.5 Precise Edit at medium, both at 2,048 pixels.

GPT Image 2.5's first edits look right, apart from slightly heavier color. By round 10 the sky has blotches. By round 15 the bushes look painted, and by round 25 the harbor is hard to read. At round 50 the background is noise and his face is breaking up. At round 100 his head is gone. Ideogram's background at round 100 is still close to the original photo.

It is one photo and one run, posted by one person, with the two models at different quality settings. OrcaRouter's launch writeup notes that no independent test has measured drift across repeated edits yet. Wilcock's own guess is that Ideogram is "likely just simple masking/in-painting."

Why does ChatGPT degrade images when you keep editing?

GPT Image 2.5 does not copy the pixels you left alone. It generates a new image that resembles the old one, with your change in it.

That new image is close, not exact. Each pass shifts color a little, sharpens or smooths texture, and fills fine detail with something plausible. On a single edit you will not see it. The next edit takes that output as its input, so it reproduces the earlier errors and adds its own. In this test, 50 rounds of that was enough to replace the background.

Ideogram's launch post names the problem directly: "With each edit, leading models add artifacts, pixel shifts, and color changes." Its answer is to change only the requested area and restore every other pixel from the input. Tibor Blaho's summary adds that it can edit a crop of a high-resolution image and stitch the result back into the original without downsizing it.

A mask does not fully solve this on OpenAI's side. The model page lists inpainting as supported, and the image edit endpoint accepts a mask, but OpenAI's image generation guide says masking with GPT Image is prompt-based: the model uses the mask as guidance and may not follow its exact shape. The pixels outside the mask can still move.

Where Ideogram 4.5 still falls short

Restoring untouched pixels keeps the background stable. The area that does change still has two problems you can see in the same test.

Printed text gets cut off

Ideogram's HELLO lost its H, its 42 became part of a 2, and its SUNDAY lost letters. In each case the print stops at a hard line partway across the chest. That looks like the edge of the area the model allowed itself to change, though Ideogram has not said how it picks that area. GPT Image 2.5 drew all three prints cleanly, on a background that was already breaking down.

The edges drift

Wilcock pointed out that the bush behind his shoulder slowly turns into a line, and his tattoos fade on the side of the arm closest to the T-shirt. The strip along the boundary of the edit area gets redrawn every round, so it degrades the way GPT Image 2.5's whole frame does, only slower and in a narrower band. By round 100 the horseshoe tattoo near his sleeve is gone.

How to edit ad creative with AI without drift

Ad creative goes through rounds: a new color, a new headline, a new product shot, a resize for each placement. Set the process up so every round starts from a clean file.

1. Make variants from the master, not from each other. Generate every variant from the approved original, not from the last variant. Wilcock chained edits on purpose to show the worst case. Most teams chain by accident, by downloading the latest output and editing it again.
2. Paste the original back. Put the AI result on a layer above the original and mask it to the area you meant to change. In Photoshop that is a layer mask. In a script it is a composite with the same mask you sent the model. Everything outside the mask is then the original file, pixel for pixel, whichever model you used.
3. Pick the model for the step. In this test GPT Image 2.5 drew printed text better and Ideogram 4.5 kept the rest of the photo. Use the first to create the new element and the second, or your own composite, to put it into the finished image.
4. Check faces, hands, text, and logos at full size every round. Drift in those areas is easy to miss in a thumbnail grid and easy to spot on a phone screen.
5. Save every round with its prompt. When a variant drifts, go back to the last clean round instead of trying to repair the broken one.

Wilcock wrote that he plans to try running the restoring approach on every frame of a video. He has not posted results yet. If you try it, expect the edge drift above to show up as flicker along the edit boundary, and test a few seconds before committing a whole cut.

Frequently asked questions

Related: Creative production · Video production · What is Jev?