Seedream 5 Pro vs Nano Banana Pro vs GPT Image 2

23:53

Three-way model comparisons usually end in a shrug: they are all good, pick whichever you like. This one did not. Across eight tests run back to back on the same prompts, one model kept winning rounds the other two did not even understand.

What follows is the result of each test, what current pricing looks like, and an honest note about how much weight eight prompts can carry.

Which model won overall?

Seedream 5 Pro, by a clear margin — six rounds to GPT Image 2's two, with Nano Banana Pro taking none. That was not the expected result going in, and the reason is less about raw image quality than about the model interpreting what was actually asked.

The pattern across the eight tests was consistent. Seedream 5 Pro composes with restraint: it leaves breathing room around a subject instead of filling the frame, which is why it won the design-led rounds. GPT Image 2 produces technically excellent work that runs crowded. Nano Banana Pro makes attractive images and then breaks a constraint.

None of that means Nano Banana Pro is a bad model — it is not, and it wins on other axes elsewhere. It means that on these eight tasks, it lost on instruction-following rather than on looks.

Which model is best at text in images?

GPT Image 2, unambiguously. On the poster test it was the only one that got the text right, the design right and the requested square format right at the same time. Nano Banana Pro rendered the text acceptably and then failed the aspect ratio, returning a square image with a vertical poster sitting inside it.

That aspect-ratio failure is worth dwelling on because it recurred. Asking for a square and receiving a square canvas containing a differently-shaped design is not a rendering error, it is the model treating your format as a suggestion — and it happened twice across the eight rounds.

If your work involves posters, packaging, thumbnails or anything where the words have to be correct, GPT Image 2 remains the default. That conclusion matches the wider guidance in the guide to making AI images with readable text.

Which one actually understands negative prompts?

Only Seedream 5 Pro. The brief asked for a night city with no people and signs carrying no text. GPT Image 2 returned a clean image with no people and no signs at all — it deleted the objects rather than the lettering. Seedream 5 Pro kept the signs and left them blank, which is what was asked for.

This is the single most interesting result in the whole test, because it separates two very different abilities. Removing a thing is easy. Understanding that an attribute of a thing should be removed while the thing itself stays is a comprehension problem, and most models solve it by deleting the whole object.

It matters well beyond night-time cityscapes. "A desk with no clutter", "a shelf with unlabelled bottles", "a street with unbranded shopfronts" are all the same shape of instruction, and they are extremely common in commercial work where you cannot show real brands.

Which restores an old photo best?

Seedream 5 Pro, because it changed the least. It stayed closest to the original face and kept photographic grain rather than the smooth plastic finish that gives AI restoration away. Nano Banana Pro returned a clean result that had altered the face, flattening the cheek line, which for a family photo is a failure however good it looks.

Restoration is the one category where "better looking" and "correct" pull in opposite directions. The person in the photograph is the point; a more attractive stranger is not an improvement. Seedream 5 Pro did add a slight glow, so it was not perfect, but it erred on the side of leaving the subject alone.

The broader workflow for this, including when to stop, is in the guide to restoring old photos with AI.

Can a model follow instructions drawn on the image?

One of them can. A portrait was marked up by hand — gold eyeshadow, golden earrings, a glossy berry lip, a stronger contour — with the whole prompt being "follow instructions on the image". Seedream 5 Pro did every edit and removed the handwriting. The other two did the edits and left the notes printed on the image.

This is the test that best explains the overall result. Every model could read the annotations. Only one worked out that the annotations were instructions rather than content, and therefore should not survive into the output.

As a workflow it is genuinely useful: sketching changes directly onto an image is faster and less ambiguous than describing them, particularly for retouching where "a bit more contour, here" is hard to write and trivial to draw.

What do the three cost per image?

At 2K on the cheapest configured provider, GPT Image 2 is about five cents, Seedream 5 Pro about seven, and Nano Banana Pro about nine. The video quoted nine cents for Seedream 5 Pro through Replicate, which is correct for that provider — routing the same model through Kie brings it down to about seven.

Rates from Rangy's live pricing tables at the cheapest configured provider, checked 15 August 2026. Providers change rates, and the same model can differ by several cents between them.

The prices are close enough that cost should not decide this. Two cents a generation only matters at volume, and at volume you are better off picking per job.

Which should you actually use?

All three, for different jobs. GPT Image 2 whenever words appear in the image or the layout is specified precisely. Seedream 5 Pro for design-led work, restoration, and anything where the instruction is subtle. Nano Banana Pro when an existing subject has to survive an edit unchanged.

• Posters, packaging, thumbnails, infographics with real copy — GPT Image 2. Text accuracy beats everything else here.
• Ads, editorial images, UI mockups, anything where composition carries it — Seedream 5 Pro. The restraint it shows with space is the difference.
• Restoring or retouching a real person — Seedream 5 Pro for fidelity to the original, Nano Banana Pro when you need a referenced face held across new scenes.
• Negative or attribute-removal instructions — Seedream 5 Pro, on this evidence by some distance.

Which is a slightly unsatisfying answer, and also the true one. The models have diverged enough that "best" is now job-shaped rather than a ranking.

How do you run this comparison yourself?

Send one prompt to several models at once and compare the outputs side by side. That is the only way to answer the question for your own work, and at five to nine cents a generation a full three-model round costs about twenty cents.

The tests in the video were run through a Photoshop plugin. If you are not in Photoshop — or you are on Affinity, which has no third-party generative panel — the same models are available in Rangy as a standalone desktop app, selectable together so one prompt fans out across all of them in parallel, with each model's price shown before you spend anything.

Either way, the discipline matters more than the tool: use your own prompts, on your own kind of work, and judge the output at full size rather than in a grid of thumbnails.

How much should you trust eight tests?

As a strong signal, not a benchmark. This is one generation per model per prompt, judged by eye, on eight tasks chosen by one person. It reflects real work rather than a synthetic scoring set, which is its strength, but a single generation can be a lucky or unlucky draw.

Stated plainly, because this guide is published by a company that sells access to all three models:

• Sample size is one per cell. Re-rolling any round could change it. The pattern across eight rounds is more reliable than any single result.
• Judging is subjective for composition and "which looks better", though the instruction-following results — signs kept, handwriting removed, aspect ratio wrong — are objective.
• It was filmed on 20 July 2026. These models are updated without version numbers, so behaviour drifts. Anything here is a snapshot.
• Your prompts are not these prompts. The result that matters is the one you get on your own work.

Frequently asked questions

The bottom line

Use GPT Image 2 when the image contains words, and Seedream 5 Pro when the instruction contains judgement. The eight rounds did not really measure image quality, which is now broadly comparable — they measured which model understood the brief.

That is the shift worth taking away. A year ago these comparisons turned on fingers, faces and artefacts. This one turned on whether a model could tell that a sign should lose its lettering but keep its post, and that handwriting on a photo was a note to the artist rather than part of the picture.

Run your own eight tests. It costs about a dollar and it will answer the question better than any comparison written by somebody else, including this one.