CREATEVISION AI · WORKFLOW TESTS
Which model keeps the details you already approved?
Tested 6–7 November 2026 · three repeats per tested configuration.
For this bottle task, start with Nano Banana 2.1 when credits matter: 120 quoted credits per three-edit chain, versus 495 for each GPT Medium chain.
For the tested portrait chain, Flare had a shorter median observed wait: 55.5 seconds versus 85.5 for Nano Banana and 95.8 for Sunburst.
None earns a blanket face-consistency win. Check whether the head actually turns; a familiar-looking face can still miss the requested pose.
Look at the last image, then work backward.
A beautiful first result can hide a growing problem: each revision quietly redraws something you wanted to keep. Inspect the final face, earrings, jacket and label against the shared starting image, then open the intermediate steps to see where a change began.
- The bookstore and navy-jacket changes are visible across the final portrait repeats. The harder head-turn instruction is less reliable: Flare stays close to the starting front view, while Sunburst and Nano Banana vary in direction and amount. We would approve the background and color separately from the pose.
- In all three final bottle repeats for Flare, Sunburst and Nano Banana, NOMA, DAILY SERUM and 30 ml remain readable. Nano Banana gives the linen more space; the GPT results use smaller folds. That is a useful composition choice, not proof that one model always preserves products better.
- Our starting recommendation: try Nano Banana for inexpensive product-scene variants; try Flare when the observed wait in this task matters more than the quote. These matched Medium tests do not show a reason to pay the same 495-credit chain quote specifically for Sunburst. Inspect its outputs before preferring it.
Compare the outputs
Same brief, visible differences.
People
Starting reference
GPT Image 2.6 Flare Medium
3:4 · 1K · medium
Prompt, settings and references
Turn the same woman’s head about 30 degrees to her left, keeping her body still. Preserve her recognizable face, copper blunt-fringe bob, teal round glasses, silver star earrings, navy jacket, white top, bookstore background, framing and light. No new objects or text.
References: 1 · high
Starting reference ↗GPT Image 2.6 Sunburst Medium
3:4 · 1K · medium
Prompt, settings and references
Turn the same woman’s head about 30 degrees to her left, keeping her body still. Preserve her recognizable face, copper blunt-fringe bob, teal round glasses, silver star earrings, navy jacket, white top, bookstore background, framing and light. No new objects or text.
References: 1 · high
Starting reference ↗Nano Banana 2.1
3:4 · 1K · 1K
Prompt, settings and references
Turn the same woman’s head about 30 degrees to her left, keeping her body still. Preserve her recognizable face, copper blunt-fringe bob, teal round glasses, silver star earrings, navy jacket, white top, bookstore background, framing and light. No new objects or text.
References: 1 · —
Starting reference ↗The first repeat opens by default. Switch repeats to inspect variation; all returned images remain available. A failed task is retained in the public record and is not counted as a successful image. Loads the tested prompt, references and parameters into the generator. Review the current quote and press Generate yourself.
Product
Starting reference
GPT Image 2.6 Flare Medium
1:1 · 1K · medium
Prompt, settings and references
Add only one small orange blossom at bottom right behind the bottle. Preserve the exact bottle shape, ivory body, cobalt cap, brass collar, cream linen cloth, sage background, camera, scale and text NOMA / DAILY SERUM / 30 ml. No other changes.
References: 1 · high
Starting reference ↗GPT Image 2.6 Sunburst Medium
1:1 · 1K · medium
Prompt, settings and references
Add only one small orange blossom at bottom right behind the bottle. Preserve the exact bottle shape, ivory body, cobalt cap, brass collar, cream linen cloth, sage background, camera, scale and text NOMA / DAILY SERUM / 30 ml. No other changes.
References: 1 · high
Starting reference ↗Nano Banana 2.1
1:1 · 1K · 1K
Prompt, settings and references
Add only one small orange blossom at bottom right behind the bottle. Preserve the exact bottle shape, ivory body, cobalt cap, brass collar, cream linen cloth, sage background, camera, scale and text NOMA / DAILY SERUM / 30 ml. No other changes.
References: 1 · —
Starting reference ↗The first repeat opens by default. Switch repeats to inspect variation; all returned images remain available. A failed task is retained in the public record and is not counted as a successful image. Loads the tested prompt, references and parameters into the generator. Review the current quote and press Generate yourself.
Waiting once is not finishing the job.
Read the complete chain alongside its quoted cost. A fast image that breaks the label may need a correction; an expensive image may still miss the brief. We have not priced hypothetical retries or human retouching.
| Tested setup | Requests | Observed wait · Median | Range | Credits per three requests |
|---|---|---|---|---|
| GPT Image 2.6 Flare MediumPeople | 9 | 55.5 s | 48.3–57.1 s | 495 |
| GPT Image 2.6 Sunburst MediumPeople | 9 | 95.8 s | 93.4–98.7 s | 495 |
| Nano Banana 2.1People | 9 | 85.5 s | 80.5–98.6 s | 120 |
| GPT Image 2.6 Flare MediumProduct | 9 | 49.8 s | 45.9–101.7 s | 495 |
| GPT Image 2.6 Sunburst MediumProduct | 9 | 91.8 s | 82.8–141.1 s | 495 |
| Nano Banana 2.1Product | 9 | 78.1 s | 76.6–78.5 s | 120 |
A chain total covers three requested edits, excluding the shared starting image.
Observed waits include submission, reference transfer, queue, network and two-second polling. Up to three requests were in flight. They describe this batch, not guaranteed model speed.
These are CreateVision AI credit quotes for the tested settings, not API prices or a universal price per image. Check the current quote before generating.
How we tested
We fixed the briefs before generating and retained every first request. Each subject starts from the same image made with Z Image Turbo, a model outside the comparison. Each editing step uses that model’s previous result. We ran three independent chains per subject and model. Requested 1K tiers can return different pixel dimensions; inspect the dimensions in the public record. No output was retouched or rerolled for quality.
Visual assessments use a fixed checklist: requested change completed, protected details preserved, required words correct, no unwanted additions. They are editorial judgments, not biometric identity scores.
The identical brief is a reproducible baseline, not the best prompt possible for every model. We did not run a separate model-optimized prompt arm or a budget-matched quality sweep. Our conclusions apply to these settings and tasks.
CreateVision AI sells access to the compared models. This is an in-house editorial test, not an independent laboratory ranking. We publish the originals and our criteria so you can disagree with the assessment.
A small set of fictional subjects cannot establish reliability for every face, product, language or campaign.
Official capabilities and tested availability
An untested feature does not count as a win. GIF, SVG, markup and web search are outside the shared image-task comparison. Official GPT Image documentation and the live model pages are linked below.
Questions before you choose
Does this prove which model always preserves faces better?
No. It tests one fictional person and one product over three repeated chains. Different faces, poses and briefs can change the result. Open all three repeats before choosing.
Why compare Flare and Sunburst separately?
A generic GPT Image 2.6 label hides a meaningful setting. Both were tested on matching three-round tasks at Medium quality and High planning depth, so you can compare the actual outputs.
Can I try the same prompt with my own photo?
Yes. Load a tested setup, replace the reference, and review the model and quote. Loading a setup does not submit a generation.
Try the task with your own image.
Choose an output above to load its exact setup. Replace the fictional reference with your own and check the quote before creating.
