Image models compared: Nano Banana, Flux, Seedream, GPT Image and Soul (a test you can run)

By Ranjit Bhinge, founder of Blur Studio · Updated November 4, 2026 · 12 min read

Short answer: There is no single best AI image model; the one that wins depends on the job, and the published comparisons are mostly written by companies that sell access to the models. Nano Banana is Google's image model family, Flux is from Black Forest Labs, Seedream is from ByteDance, GPT Image is from OpenAI, and Soul is the image model that Higgsfield (a web platform that hosts many image and video models) makes itself. Reports say Nano Banana is strong on product fidelity and edits, GPT Image on text inside the image, Seedream on batches and Flux on photoreal shots, but these come from vendor tests with one prompt each. The reliable answer is a bake-off on your own product, which this page lays out, with prices as of November 2026.

Every roundup says a different model wins, and nearly all of them sell access to the models they rank. Here is how to run a fair test on your own product in an afternoon, and how to price it to a finished asset.

I'm Ranjit, and I run Blur Studio: 15+ years building software, 120 apps taken live, and clients who have gone on to raise $10M+. I build AI systems and processes that help businesses do a lot more with a lot less time. I also work differently from most of what else is out there, more on that further down this page. Say hello.

If you make ads, product pages or social posts, you will have seen the same five names in every comparison: Nano Banana, Flux, Seedream, GPT Image and, on Higgsfield, Soul. This page does three things. It says who makes each and what they are reported to be good at. It shows the prices I could read from the makers' own pages. And it gives you a repeatable five-prompt test so you pick a model with your own product, not someone else's skincare serum.

I use Higgsfield, which lets you run several of these models from one credit balance, and I am an affiliate for it.

This page may contain affiliate links from which Blur Studio earns commissions.

The short answer: route by job

Pick by the problem you need to solve, not by a ranking:

Words inside the image (a headline, a price, a label): test GPT Image first. In the only structured ad test available, it was the only model to pass copy-in-image, but that test was by a vendor and used one product.
The product must match your photo (label, shape, colour): test Nano Banana first, and Flux if you need to match exact brand colours.
Twenty variants that look like one family: test Seedream first.
Stylish people and mood images: test Soul (Higgsfield's own model).

Those starting points are reports, not facts about your product. Models also change fast, so a model that failed last quarter may pass today.

What each model is and who makes it

Nano Banana is the nickname for Google's Gemini image models. The family has several versions (the Google pricing page lists Nano Banana, Nano Banana 2, Nano Banana 2.1 and Nano Banana Pro). It is known for editing an image by chatting with it and for keeping a character or product consistent.
Flux is a family of image models from Black Forest Labs, a German company. FLUX.2 has versions named pro, flex, max, klein and dev, and the company sells access by the image with no subscription.
Seedream is ByteDance's image model family (ByteDance owns TikTok). It is reported to handle text-heavy layouts, batches and high resolution well.
GPT Image is OpenAI's image model, the same company as ChatGPT.
Soul is Higgsfield's own image model, with presets and a mood board feature for a consistent look. I cover it in the Soul explainer.

Higgsfield also lists other image models (Kling O1 Image and Cinema Studio Image 2.5, among them). Those are outside this comparison.

Which model suits product shots, people and text: what is reported

Treat this table as a list of hypotheses to test. Every source here is a company that sells access to image models, each ran very few prompts, and several mixed model versions.

Job	Model reported as strongest	Where it comes from
Text inside the image	GPT Image 2 (only "pass" in one test)	Kreado, an ad tool vendor, one skincare product, two reviewers, May 2026
Product fidelity to a source photo	Nano Banana 2	Same Kreado test
Variant speed and edits in conversation	Nano Banana	Kreado and Picsart (an online design company)
Batch with a consistent style	Seedream 5.0	Kreado; Krea (an AI image platform) says similar
Photoreal hero shots, character continuity	Flux	Krea
Matching exact hex colours	Flux 2 (Nano Banana reported not to)	Picsart, unverified, no outputs shown

A fair reading: the sources agree that no model wins everything, and they disagree on specifics. That is itself the reason to run your own test.

Two other points recur across comparison pages. Picsart reports that Flux 2 can take up to 8 reference images and exact colour codes, while Nano Banana can search the web and reason before drawing. WaveSpeed's roundup also names models outside this page: Flux Klein (a smaller, cheaper Flux that its page says has open weights) and Qwen Image (from Alibaba, praised there for text in two languages). Check their makers' prices, and if text-heavy or very cheap volume work is your need, add one to your bake-off.

What most comparison pages get wrong

They are sellers. Picsart, Krea, Kreado and WaveSpeed all sell access to models. That does not make them wrong, but it shapes what they test.
One prompt, one product. A single run shows luck as much as skill. Image models vary run to run.
Mixed versions. WaveSpeed's price table puts different versions side by side, and one source tests "GPT Image 2" while another prices "GPT Image 1.5".
Price per image, not per finished asset. Krea makes this point itself: judge cost to a finished asset, not per attempt.
No rights notes. Few say what the licence or watermark rule is.

Prices I could read from the makers' pages

Figures are as of 4 November 2026. Check each page before relying on any figure.

Model	Price shown	Source
Nano Banana 2.1 (Google API)	$0.0336 per image at 1K, $0.0504 at 2K, $0.113 at 4K	Google AI Studio pricing
Nano Banana 2 (Google API)	$0.045 at 0.5K, $0.067 at 1K, $0.101 at 2K, $0.151 at 4K	Same
Nano Banana Pro (Google API)	$0.134 at 1K or 2K, $0.24 at 4K	Same
Original Nano Banana (Google API)	$0.039 at 1K; Google lists a shutdown date of 15 March 2027	Same, plus Google's deprecations page
GPT Image 2 (OpenAI API)	Priced per token: $8 per million for image input, $30 per million for image output. No per-image table	OpenAI API pricing
Flux (Black Forest Labs)	Pay as you go, no subscription. A calculator example showed $0.048 for one image, with the model and size not stated; per-model rates are in its calculator	Black Forest Labs pricing
Seedream	Maker's per-image price not read	Not available

"1K, 2K, 4K" are image sizes: roughly 1,000, 2,000 and 4,000 pixels on the long side. Note on OpenAI: its page lists tokens, and a token is a chunk of data the model processes, so there is no official per-image price on that page. Reseller sites publish per-image figures for GPT Image that are conversions of the token rates, and they vary widely across versions and sizes. I do not use them. Use OpenAI's own calculator for your image size and quality. The same warning applies to any Seedream per-image number you see quoted: I could only find reseller and blog figures, so I leave it out.

Credits on Higgsfield

Higgsfield charges in credits, one balance across its models. As of 4 November 2026 its pricing page shows Starter at $19 a month for 270 credits, Plus at $59 monthly or $47 annual for 1,200, and Ultra at $129 monthly or $99 annual for 3,000, so a credit costs roughly $0.033 to $0.07 depending on the plan (my arithmetic). Check the pricing explainer and the live page.

Credit costs per image printed on Higgsfield's pricing page, with my dollar estimate at those credit values:

Model	Credits per image	Roughly, in dollars
Soul 2.0	0.12	under a cent
Nano Banana	1	$0.03 to $0.07
Nano Banana 2	2	$0.07 to $0.14
Nano Banana Pro	2 (4 at 4K)	$0.07 to $0.14 ($0.13 to $0.28 at 4K)
Seedream (printed row)	1	$0.03 to $0.07
GPT Image (printed row)	1	$0.03 to $0.07
FLUX.2 Pro	1	$0.03 to $0.07
FLUX.2 Flex or Max	3 or 4	$0.10 to $0.28

The Seedream and GPT Image rows do not name a version, and Higgsfield's model list has newer ones (Seedream 5.0 and GPT Image 2.5). Their cost shows on the Generate button, so read it there. Subscription credits expire at the end of each billing cycle, which raises your real cost if you do not use them. Compared with Google's API prices above, Nano Banana Pro on Higgsfield (2 credits) works out similar to the $0.134 API price on a mid plan and cheaper on a big annual plan, but you also pay for the subscription.

Cost to a finished asset, not cost per image

Formula: (attempts needed to get an approved image) times (cost per attempt), plus the cost of your time to fix it. In a made-up example, Model A costs 2 credits and needs 3 attempts to get one approved ad image (6 credits), and Model B costs 4 credits and gets it in one go (4 credits). The pricier model is cheaper per approved image. But the time matters more than either number: if each image then needs 10 minutes of cleanup in a design tool and your time is worth $30 an hour, that is $5, which dwarfs the credits. A model that gives you an image needing 2 minutes of cleanup beats one that is half the price and needs 10.

So record three numbers per model per job: attempts to approval, credits per attempt, and minutes of cleanup. For text-heavy ads, plan a human typography pass anyway for price lines and legal text, in a design tool such as Canva or Figma (see getting text on images that reads right).

A 5 prompt bake-off you can run in an afternoon

A plan you can run, not a record of results. You need one good photo of your product as a reference, the models you want to compare, and a scoring sheet.

Label legibility. "Use the attached product photo. Show the product on a plain white surface, front facing, label fully readable. Keep the label text exactly as in the photo."
Headline plus price. "Create a square ad using the attached product. Headline: [your real headline]. Price: [your real price]. Place the text at the top, clean sans-serif, high contrast."
Lifestyle scene. "Place the product on a kitchen counter in morning light, shallow depth of field, no people."
Person holding the product. "A person holding the product at chest height, smiling, natural hands, label facing the camera."
One conversational edit. Take your best result and ask: "Change the background to a pale blue wall and keep everything else the same."

Run each prompt 3 times on each model, so you see the spread. For 5 models that is 75 images. A rough cost on Higgsfield, using the printed credits: about 15 images on Nano Banana Pro (30 credits), 15 each on Seedream, GPT Image and FLUX.2 Pro (about 45 credits at 1 each, if the newer versions cost the same, which you should check), and 15 on Soul 2.0 (under 2 credits), roughly 77 credits, or about $2.50 to $5.40 at the plan prices above. Newer versions may cost more, and the button shows the real figure before you confirm.

Scoring sheet and pass or fail rules

Check	Pass	Partial	Fail
Label and logo match the photo	Every letter correct	One small error	Wrong or invented text
Headline and price	Spelled right, in place	Right but needs a nudge	Misspelled or wrong number
Product shape and colour	Matches	Close	Different product
People and hands	Natural	One flaw you can retouch	Distorted hands or face
Edit keeps the rest	Only the background changed	Small drift	Whole image changed

Rules: score each of the 3 attempts, and count a model as passing a prompt only if at least 2 of 3 are Pass or Partial. A model with a Fail on label or price text is out for that job, whatever else it does well. Pick a default model per job, and write down the model name and version with every saved asset, because names and versions change (Google has already announced the original Nano Banana will shut down on 15 March 2027). For logging assets, see where to keep AI images and prompts.

Licence, watermark and disclosure notes

Higgsfield: its Terms say it does not restrict commercial use of outputs, but the pricing table marks commercial use as not included on the Free plan, and free output carries a Higgsfield watermark. Use a paid plan for client work. Outputs may not be unique across users.
Provenance marks: the Terms say Higgsfield may embed machine-readable marks (such as C2PA Content Credentials) and that you must not remove them. One source (Picsart) says every Nano Banana image carries a SynthID watermark (an invisible mark Google adds to its AI images). Google's own page does not confirm that.
Disclosure: platform rules for labelling AI images vary and change. Check each platform's current rule and local law. See should agencies disclose AI use. This is information, not legal advice.
Flux weights: Black Forest Labs' self-hosted licences are tiered, and one model type was reported as non-commercial. If you run Flux yourself, read the licence for the specific model.

Using these models through Higgsfield versus going direct

Higgsfield's value for this job is one account and one credit balance across several of these models, so you can run the bake-off in one place and switch model per job. According to its model list in the product, the image models include the Nano Banana family, Seedream 4.5 and the 5.0 versions, GPT Image 2.5, Flux.2 Pro, Kling O1 Image and Soul 2.0. The pricing page has also shown time-limited "unlimited" blocks for some of these on paid plans, with rules (web only, personal use, no automation), so check the live page.

Going direct to Google, OpenAI or Black Forest Labs bills per image or per token and needs a developer, but it suits high volume, automated pipelines, and cases where you want a specific model version the day it appears. If you already have a developer and one model does the job, direct is often simpler and cheaper. The comparison with another subscription tool is in Higgsfield vs Midjourney, and the wider tool list is in best AI image generators for business.

Try Higgsfield with a discount

I use Higgsfield a lot and I love it, which is why I partnered with them to bring you a discount. Get started through this link and you can get up to 50% off for the first 3 hours after you log in. The size of the discount and the plans it covers can vary, so check the offer on screen.

Get started with the discount

Who should skip Higgsfield for this

You use one image model all day and a developer can call it directly. Paying the maker per image may cost less.
You need a model version the day the maker releases it. Platforms often add new versions later.
You only need a few images a month. A subscription whose credits expire each cycle can cost more than pay-per-image.
You need a legal, contract-level promise about training data or indemnity. The Terms give users no IP indemnity, and Higgsfield may train on content except under an Enterprise agreement.
Your main need is photo retouching of real product shots rather than generation. Use a retouching tool, and see best AI tools for product photography.

Decision guide

If your bottleneck is	Test first	Then
Text inside the image	GPT Image	Add a human typography pass for prices
Matching your real product	Nano Banana (Pro or 2)	Check labels letter by letter
Brand colours exactly	Flux, and Soul's HEX colour feature	Verify with a colour picker
20 consistent variants	Seedream	Fix the style prompt, then batch
People and mood	Soul	Check hands and faces
Lowest cost for volume	The cheapest model that passes your sheet	Price per approved asset

How I'd approach this

I would not choose a model from a roundup. I would run the five prompts on three models, score them on the sheet, and keep a one-line rule per job: "text ads use X, product shots use Y, batches use Z." Then I would repeat the test every few months, because models and prices move.

Keep the log of model, version, prompt and credits with each asset, so a client can ask for "the same again" next month and you can do it. If you want that handled for your team, with an approval step and a shared library, Blur Studio builds it; you can see how it works. And if you want a single account to run the test, a month on a Higgsfield plan is enough to try all five.

Questions people ask

Is Nano Banana better than Flux for product photos?

It is commonly reported that Nano Banana holds a product closer to a source photo and edits well in conversation, while Flux is reported to suit photoreal hero shots and exact colours. Those reports come from vendors with one-prompt tests, so run the label-legibility and person-holding-product prompts from the bake-off on your own product.

Which model is best for text in images?

In the one structured ad test available (by a vendor, one product), GPT Image 2 was the only model to pass copy inside the image. Treat that as a starting point. Whatever model you use, check every letter and plan a typography pass in a design tool for prices and legal lines. See getting text on images that reads right.

Is Seedream good for ads?

Vendors report Seedream is good for batches with a consistent style, text-heavy layouts and high resolution. Per-image prices vary by platform, so confirm the cost on the platform you use. Run the batch test: 20 variants of one ad, then count how many look like the same family.

Is GPT Image cheaper than Nano Banana?

OpenAI prices GPT Image 2 per token, not per image ($8 per million input tokens and $30 per million output tokens as of 4 November 2026), so any per-image figure is a conversion, often from a reseller. Google lists Nano Banana Pro at $0.134 per image at 1K or 2K. Compare on your own image size, using each vendor's calculator.

Are these models free to use commercially?

Each maker has its own terms. Higgsfield's Terms do not restrict commercial use of outputs, but its pricing table marks commercial use as not included on the Free plan and free output is watermarked, so use a paid plan for client work. Check Google's, OpenAI's and Black Forest Labs' terms if you go direct, and clear rights and likenesses. This is information, not legal advice.

Does Higgsfield have all of these models?

According to the product's model list on 4 November 2026, it has the Nano Banana family, Seedream 4.5 and 5.0 versions, GPT Image 2.5, Flux.2 Pro and its own Soul 2.0. The list changes, so check the model picker. See the Soul explainer for Soul.

Do I need to label AI images?

Often yes, depending on the platform and where you are. Higgsfield's Terms say you must disclose AI generation where law requires it and must not remove its provenance marks. Check each platform's current rule. See should agencies disclose AI use. This is information, not legal advice.

How many attempts should I budget per approved image?

Start with 3 per prompt in the bake-off, then use your own ratio of attempts to approvals. Budget differently for text-heavy ads, which usually need more tries. See the credit guide for ways to keep draft costs down.

More in this topic: Higgsfield compared: X vs Y (19 guides) →