Short answer: Ideogram 4.5 Edit kept the rest of the image the most consistent. After 40 chained edits its background drift averaged 0.95 on a 0–255 scale, which is invisible. FLUX 3 Edit ended at 5.64, GPT Image 2.5 Sunburst Edit at 48.44 and Nano Banana 2.1 Edit at 100.48. Two caveats: this test scores the scene, not the face, and Ideogram 4.5 replaced one subject's face late in a chain. Nano Banana 2.1 followed the instructions best.
Ask an AI image editor to change one thing and keep everything else the same. Then do it again on the result. Then 38 more times. Some models leave the rest of the photo alone. Others repaint the whole frame on every edit, and the small changes add up.
We ran that test on four image editors on October 7, 2026: Ideogram 4.5 Edit, FLUX 3 Edit, GPT Image 2.5 Sunburst Edit and Nano Banana 2.1 Edit. Three photos, 40 chained edits each, 480 edits in total. Below are the results, what they mean for faces, and how to keep the same person across images and videos.
Ideogram 4.5 and FLUX 3 kept the background at or under the point where you can see a change. On GPT Image 2.5 Sunburst and Nano Banana 2.1, the plain wall behind the subject turned into a mosaic, and the change was visible by about edit 10. Ideogram 4.5 also sometimes did nothing at all.
The results
Background drift after 40 chained edits. Lower is better. Under about 5 is invisible; about 10 is visible side by side.
| Model | Kitchen | Café | Still life | Mean |
|---|---|---|---|---|
| Ideogram 4.5 Edit | 0.00 | 2.86 | 0.00 | 0.95 |
| FLUX 3 Edit | 9.11 | 7.80 | 0.00 | 5.64 |
| GPT Image 2.5 Sunburst Edit | 41.98 | 58.36 | 44.99 | 48.44 |
| Nano Banana 2.1 Edit | 80.58 | 87.81 | 133.04 | 100.48 |

The two groups behave differently, not just by degree.
Ideogram 4.5 and FLUX 3 put the untouched pixels back. Their control areas scored exactly 0.0 for long stretches. Ideogram 4.5 stayed at 0.0 for all 40 edits on the kitchen and the still life. FLUX 3 stayed at 0.0 on the still life, and on the other two photos it moved in two jumps (edits 15 and 20 in the kitchen, 13 and 21 in the café) and then held flat.
GPT Image 2.5 Sunburst and Nano Banana 2.1 redraw the whole frame every time. Each edit adds a little noise and color shift to areas nobody asked to change. Averaged over the three photos, Nano Banana 2.1 crossed 10 at edit 8 and GPT Image 2.5 Sunburst at edit 11. By edit 40 the white kitchen wall was a pink or olive mosaic.

The hair color and the short bob in the kitchen chain are not drift. Edits 14 and 22 asked for them.
Where each model lost points
Drift is only half the job. We also checked every step by eye to see if the requested change actually happened.
- Nano Banana 2.1 followed instructions best. We found no clear misses in 120 edits. It is also the fastest, at a median of 12.6 seconds per edit.
- GPT Image 2.5 Sunburst missed about 3 to 4 edits. Median 17.2 seconds per edit.
- FLUX 3 missed or weakly applied about 5 to 7 edits. Closed eyes and a beard trim did not land, and the necklace never appeared. It was also the slowest at a median of 41 seconds, with 14 timeouts or 503 errors. All 14 succeeded on retry.
- Ideogram 4.5 sometimes did nothing. On several edits it returned a pixel-identical image: the necklace, earrings, freckles and nail polish. It turned "eyes closed" into a wink. Median 23.7 seconds per edit.
Ideogram 4.5's low café score also needs a caveat. It redrew the grout lines of the tiled wall, and the café man's face was replaced around edits 26 to 32. The café scored only 2.86 because a bright, plain wall shows little change in our measure even when its structure moves. Do not read Ideogram 4.5's numbers as a promise that faces stay the same.

Which AI model keeps the same face across edits?
None of the four editors earned a face-consistency win in this test, because the test measured the background, not the face. Our score compares three untouched background patches with the original photo. It says nothing direct about the person.
What the data does show about faces:
- The background winner still lost a face. Ideogram 4.5 Edit had the lowest drift (0.95), yet it replaced the café man's face around edits 26 to 32.
- Whole-frame redraws touch everything. GPT Image 2.5 Sunburst and Nano Banana 2.1 redraw the full frame on every edit, so the face is redrawn too, 40 times in a row. Their background drift crossed 10 by edits 11 and 8.
- Long chains are the risk. Drift builds up with each round on models that redraw the frame, and the café face change came after more than 25 edits.
So if the face matters, do not depend on any editor to hold it over a long chain. Edit from the original or an approved version each time, and give the model the same reference photos of the person on every generation, as described next.
Which editor should you use?
Many edits on the same scene: Ideogram 4.5 or FLUX 3. Product colorways, outfit swaps, prop changes, ad revisions. Anything where the background has to survive round after round. Check each step, because both will sometimes skip or soften an edit.
One edit where the instruction matters most: Nano Banana 2.1. It got the change right more often than any other model here. Use it for a single edit, not a 40-step chain.
GPT Image 2.5 Sunburst sits in between on instructions and drifts like Nano Banana 2.1, only more slowly.
On any model, branch instead of chaining. Make each edit from the original, or from a version you already approved, instead of from the latest result. A model that redraws the frame adds drift every round, so fewer rounds means less drift.
| Model | Morphed credits per edit (setting we tested) | Median seconds per edit |
|---|---|---|
| FLUX 3 Edit | 8 (1K) | 41 |
| Ideogram 4.5 Edit | 9 (Medium) | 23.7 |
| Nano Banana 2.1 Edit | 11.5 (1K) | 12.6 |
| GPT Image 2.5 Sunburst Edit | 12 (Medium) | 17.2 |
How do I keep the same character across images and videos?
Save the person once as a fixed set of reference images and send those same references with every image and every video, instead of describing them again or editing the last result. Reference images give the model the face to copy each time, so errors do not stack up the way they do in a 40-edit chain.
On Morphed this is the Characters feature:
- Create the character. Describe a new person and pick traits (gender, age, heritage, build, and a style: realistic, 3D animated, anime or illustrated), then choose one of the generated portraits. Or upload up to four clear photos of one face.
- Let setup build the references. Morphed makes two images from your pick: a main photo on a plain white background and a character sheet showing the front, a three-quarter view and the full profile. Those two are the references every later generation sends. Characters are created with GPT Image 2.5 Sunburst reference editing, with no training step.
- Mention them in any prompt. Type @ and the character's handle in an image or video prompt. The references go to models that accept reference images.
- For video, approve a still first. Put the character in the scene as an image, check the face, then animate that still. For a shot list, start each shot from an approved still of the same character rather than from the last frame of the previous clip. Video models must support reference images to use a character, and Seedance 1 cannot use characters.
We have not run a drift test on video, so we can't tell you which image-to-video model holds a character best across shots. The method above is the part you control on any platform.
How do I keep the same AI creator or influencer across variants?
Generate every variant from the same saved references, and change one thing per prompt. A variant made from the previous variant inherits its drift. A variant made from the saved character starts clean every time.
- One anchor, many variants. Outfit, location, product and format changes all start from the character, not from yesterday's post.
- One change per prompt. In the drift test, each edit asked for one change. That made misses easy to spot and easy to redo.
- Review before you reuse. If a variant changes the face, don't save it as a new reference. Throw it away and regenerate from the character.
- Pick the editor for the job. For many small changes to one finished scene, Ideogram 4.5 or FLUX 3 kept the background best in this test. For a single edit that has to land, Nano Banana 2.1 followed instructions best.
How we tested
This test copies the drift benchmark Stav Zilbershtein (@mightyking) posted on X. Credit for the method goes to him.
- Photos. Three generated photos: a woman in a kitchen, a man in a café and a milk-bottle still life. We made them with Seedream 5 Pro, a model outside the test, so no model edited its own image.
- Edits. 40 edits per photo, each asking for one change, followed by "Keep everything else in the image exactly the same." Each edit ran on the previous step's output. Each chain ran once.
- Control patches. Three background areas per photo that no edit touches, such as an empty stretch of wall.
- Alignment. We resized every output to the original size and aligned it to the original by the best whole-pixel shift within ±12 px, measured on the patches. Whole-image phase correlation found shifts that were not there, so we dropped it.
- Score. The mean absolute RGB difference between each step and the original image, over the patch pixels, on a 0–255 scale. Under about 5 is invisible; about 10 is visible side by side.
- Settings. The defaults Morphed runs: Ideogram 4.5 Edit at Medium with high edit precision, FLUX 3 Edit at 1K, GPT Image 2.5 Sunburst Edit at Medium (1024 × 1024), Nano Banana 2.1 Edit at 1K.
Read the 40 kitchen edits
Each edit ended with: "Keep everything else in the image exactly the same."
1. Make the tank top red
2. Make the jeans black
3. Make the sneakers bright yellow
4. Give her round tortoiseshell glasses
5. Put her hair up in a high ponytail
6. Make her smile with her teeth showing
7. Add a thin gold necklace
8. Make the tank top navy blue with thin white horizontal stripes
9. Add a silver wristwatch to her left wrist
10. Make her hold a white coffee mug in her right hand
11. Change the black jeans to khaki chinos
12. Add small gold hoop earrings
13. Change the sneakers to black leather ankle boots
14. Make her hair dark brown
15. Give her red lipstick
16. Make her hold a green apple in her left hand
17. Add a thin brown leather belt to her trousers
18. Make the tank top emerald green
19. Make the coffee mug yellow
20. Add a few freckles across her nose and cheeks
21. Roll the trouser legs up to just above the ankle
22. Cut her hair into a short chin-length bob
23. Change the gold necklace to a pearl necklace
24. Change the ankle boots to white canvas sneakers
25. Add a black baseball cap on her head
26. Make her eyes closed
27. Make the glasses black-framed
28. Add a beaded wooden bracelet to her right wrist
29. Change the trousers to a knee-length denim skirt
30. Make the baseball cap red
31. Add a light-grey cardigan worn open over the tank top
32. Replace the green apple in her hand with a banana
33. Give her pink nail polish
34. Make the cardigan mustard yellow
35. Make the sneakers green
36. Remove the glasses
37. Make her raise her eyebrows in surprise
38. Make the coffee mug white with a black stripe around it
39. Make the denim skirt black
40. Remove the baseball cap
This is three photos and one run per chain, not a statistical benchmark. A plain, bright wall also under-counts some changes, as the café result shows. For an earlier, shorter test with changed-pixel maps and text edits, read Ideogram 4.5 vs GPT Image 2.5, Nano Banana Pro and Seedream 5.
Try it on your own image
Open the AI image editor, upload a photo and pick the model for the job: Ideogram 4.5 or FLUX 3 for a chain of edits on one scene, Nano Banana 2.1 for a single edit that has to land. Ask for one change per prompt, and go back to your original when the result starts to drift.
