Ideogram 4.5 Editing Guide: Three Edits in a Row
Ideogram 4.5 is built to edit one thing and leave the rest alone. We ran three edits in a row against Nano Banana 2.1 and measured how far each image drifted.
Ideogram 4.5 is an image model from Ideogram, released on September 30, 2026, and aimed at editing: you change one thing and the rest of the picture should stay as it was. On GPT-IMG it costs 20 credits for a text-to-image render and 15 credits per edit. This guide runs three edits in a row on one image, compares the result with Nano Banana 2.1, and shows how to write edit prompts that hold up.
Created: October 10, 2026. Updated: October 10, 2026.
What Ideogram 4.5 is for
Most image editors redraw the whole frame when you ask for a change. A small request like "swap the cup" can shift the colors of the street behind it, soften the grain, or nudge a plant. Repeat that a few times and the picture wanders away from the original. Ideogram describes the new model as its answer to this: it is meant to leave untouched areas alone across repeated edits, and it also does text-to-image with accurate lettering.
That is the vendor's claim, so we tested it. The rest of this post is one scene, three edits, and a measurement.
The test
We generated one source image, then edited it three times in a row. Each edit used the previous result as its input. We ran the identical chain on Nano Banana 2.1, starting from the same source image. Settings: 1K, 3:2, one image per request.
The source prompt (20 credits, 11.5 seconds):
A small cafe window table photographed from inside, a neon sign in the window reading "OPEN LATE", a white ceramic coffee cup on a wooden table, a potted fern on the windowsill, rainy street outside, 35mm photo
The three edit prompts, sent one at a time:
1. Change the sign text from "OPEN LATE" to "CLOSED". Change nothing else.
2. Replace the coffee cup on the table with a glass of orange juice. Change nothing else.
3. Replace the potted fern on the windowsill with a small cactus. Change nothing else.
Each edit cost 15 credits on both models. Ideogram finished each one in about 11 seconds (10.8, 10.9, 10.8). Nano Banana 2.1 took 14, 17.2 and 14 seconds. Those are single runs, not averages.
What came back
Ideogram after three edits:
Nano Banana 2.1 after the same three edits:
Both models spelled CLOSED correctly and made all three changes. The difference is in everything nobody asked them to touch.
- The sign. Ideogram kept the pink and blue neon and only changed the letters. Nano Banana 2.1 redrew the sign in red after the first edit, and it stayed red.
- The scene. Ideogram kept the street, the reflections, the window frame, the lettering on the glass and the newspaper rack in place. Nano Banana 2.1 pushed the whole frame warmer and added visible film grain by the third edit, and the table and sugar pot moved slightly.
- The plant. Both swapped the fern for a cactus. Nano Banana 2.1 also changed the pot from dark to terracotta.
Measuring the drift
Eyes miss small shifts, so we measured them. After each edit we compared two areas that no prompt mentioned, the street seen through the window and the newspaper rack, against the source image. The numbers are the average difference per pixel on a 0 to 255 scale, so lower means closer to the original. A value around 1 is the floor, because every image here went through the same resize and compression.
| Model and region | Edit 1 | Edit 2 | Edit 3 |
|---|---|---|---|
| Ideogram, street | 1.1 | 1.6 | 1.8 |
| Ideogram, newspaper | 0.9 | 1.1 | 1.3 |
| Nano Banana 2.1, street | 3.7 | 5.8 | 7.1 |
| Nano Banana 2.1, newspaper | 6.4 | 7.3 | 8.4 |
It stayed close to the floor through all three edits. Nano Banana 2.1 started higher and kept climbing, which matches what the pictures show. Run the chain longer and the gap would likely grow, but we stopped at three.
Where it fell short
The model is not a free win. In the second edit we asked it to replace the coffee cup, and the saucer under the cup went with it. That is a fair reading of the prompt, but if you want the saucer, say so. The glass also landed in the same spot as the cup, which is what we wanted here and not something to count on.
This is also one scene and one run per model. We did not test people or faces, long blocks of text, busy product photos, or a fourth and fifth edit. The measurement covers two regions of one image. Treat it as evidence that Ideogram held still better in this case, and try your own picture before you rely on it.
What you can set on GPT-IMG
| Setting | Ideogram 4.5 |
|---|---|
| Text-to-image | 20 credits, 1K |
| Edit with a reference image | 15 credits |
| Aspect ratios (text-to-image) | 1:1, 2:3, 3:2, 3:4, 4:3, 9:16, 16:9 |
| Reference images | Up to 4, the first is the image being edited |
| Resolution | 1K only |
| Quality tiers | Not configurable |
| Masks | Not available yet |
An edit keeps the proportions of the image you upload. Nano Banana 2.1 goes up to 4K and takes up to 5 reference images, so it is the better pick when you need a larger file. The full price list is on the models page, and the pricing page explains credit packs and plans.
How to write an edit that holds
These habits are what we used in the test above, and they are the ones to copy first.
- Change one thing per prompt. Run the sign, the cup and the plant as three requests, as we did, and check each result before the next.
- Name the object and what replaces it. "Replace the coffee cup on the table with a glass of orange juice" gives the model a target and a result.
- End with "Change nothing else." It costs five words and states the rule you care about.
- Spell out text in quotes. For sign and label edits, write the exact new words inside quotation marks.
- Mention what must survive. If the saucer matters, write "keep the saucer".
Upload the picture on the edit an image without changing the rest page, which lists Ideogram 4.5 among its models. The general image to image editor works the same way, and the reference image workflow guide covers how to combine several uploads.
FAQ
What is Ideogram 4.5?
Ideogram's image model released on September 30, 2026. It handles text-to-image and, mainly, editing, with a focus on leaving the unedited parts of a picture unchanged across repeated edits. On GPT-IMG it runs at 1K.
How much does Ideogram 4.5 cost on GPT-IMG?
20 credits for a text-to-image render and 15 credits for an edit with a reference image. The price does not change with aspect ratio. New accounts get free credits to try it, and generation failures are refunded as credits.
Does Ideogram 4.5 support masks?
Not on GPT-IMG yet. You describe the change in words and name the object, and the model decides what to alter. In our test that was enough to change a sign, a cup and a plant while the background held steady.
Is Ideogram 4.5 better than Nano Banana 2.1 for editing?
In our one three-edit test it drifted less: the unedited regions stayed much closer to the original. Nano Banana 2.1 offers 2K and 4K output and one more reference image, so the better choice depends on the job. Our Nano Banana 2.1 test covers that model on text-heavy prompts.
If your edits keep shifting parts of the picture you did not mention, try Ideogram 4.5 on the edit an image without changing the rest page, and compare it with Nano Banana 2.1 when you need more resolution. Browse the gallery for edit examples from other models.