Typing a prompt and getting a picture is no longer the hard part. The hard part, for a company, is everything printed around the picture: whether the business owns it, whether it can go on packaging, whether the uploaded product shots end up in somebody's training set, and what one usable image costs once the discarded drafts are counted. This ranking covers thirteen products a small or mid-sized business can buy today, from design workspaces to per-image APIs and model weights you run yourself. Each was judged on its vendor's own pages, fetched for this review: the meter it bills on, the stated ownership and commercial-use terms, whether an indemnity is written down, what the vendor says it does with prompts and uploads, and whether there are seats, SSO or an API. Picture quality is deliberately not scored; it changes with every model release, and the contract does not.
Visibility in this ranking can be paid for. Payment moves a vendor's position within the
shortlist; it never adds a vendor, and it never changes a word of the review. The largest vendors in AI image generators
cannot hold places 1 to 3. How it works: placement disclosure ·
editorial process.
How we ranked these
Setup: a browser tab, an API key, or a GPU under a desk
Getting started ranges from nothing to a small infrastructure project. Recraft, Krea, Magnific, Adobe Firefly and getimg.ai are workspaces: sign in, type, download. Black Forest Labs, fal, Replicate, Bria and Google's Gemini API hand over a key and documentation, and the first image arrives only after someone writes code or wires up an MCP server. Self-hosting Stable Diffusion or FLUX weights means hardware, drivers and somebody who owns the updates. The slower, less visible part of setup is making output look like your brand: training a custom style, collecting reference images, agreeing prompt conventions, and deciding who may spend credits. Shared styles and central administration appear on Recraft Teams and on Krea's Business and Enterprise plans; most API vendors leave that layer for you to build.
The real price: credits hide the cost per usable picture
Nothing on this page is priced per seat in the way office software is. The meters we found are monthly credits (Recraft, Adobe, Photoroom, getimg.ai), compute units (Krea), yearly credit pools (Magnific), dollars per image or megapixel (Black Forest Labs, fal, Replicate, Google, Bria) and prepaid packs (Getty). Credits are not comparable across vendors, and inside one vendor a model can cost 200 times another: Krea lists 2 credits for its Krea 2 Turbo and about 400 for Flux 2 Max. Resolution moves the bill too; Google charges $0.067 for a 1K image and $0.151 for 4K from the same model. The honest unit is cost per image you actually publish, which means multiplying by the drafts thrown away. Check whether unused credits expire monthly, and whether top-ups do.
Getting your data out: files are easy, the training question is not
Exporting is rarely a problem here, since the product is a PNG, JPEG or SVG you download. Retention is the catch on the API side: Bria says generated images expire after three days and Replicate removes API outputs after an hour by default, so a pipeline must copy results to its own storage. The harder half of this test is inbound. Recraft documents training use per plan and exempts its API entirely. Black Forest Labs and Photoroom use customer content for training unless you opt out, and Stability AI offers an opt-out for its hosted services. Adobe says it has never trained Firefly on user content. Google separates unpaid from paid quota. For Magnific and getimg.ai we could not find a statement at all, which is itself an answer.
Independence from the vendor: who owns the model behind the button
There are two dependencies to weigh. One is the company you pay; the other is the lab whose model draws the picture. Aggregators such as Krea, Magnific, getimg.ai, fal and Replicate resell models from Black Forest Labs, Google, OpenAI and others, so a price change or a withdrawn model upstream lands on you through a vendor that cannot prevent it, although switching to a different model is easy. Single-model vendors such as Recraft tie a house style to one roadmap. The strongest position is weights you can run without the vendor: FLUX.2 klein 4B under Apache 2.0, Stable Diffusion under the Community License, Bria deployed in your own cloud or on-premises. Prompts and custom styles are the lock-in, because neither transfers cleanly between models.
Who it's for: designers, developers and legal each want a different product
Three buyers show up in this market and they rarely want the same thing. A marketing or design team wants a canvas, styles, vectors and colleagues in one place; Recraft, Krea, Magnific and Adobe Firefly serve that. A product team embedding generation in its own software wants a predictable per-image price and an SLA; Black Forest Labs, fal, Google, Bria and Replicate serve that. A legal or brand-protection function wants licensed training data and an indemnity in writing; Bria, Adobe and Getty make that their pitch. Sellers with a catalog to shoot are a fourth case that Photoroom addresses directly. A company with more than one of these needs usually ends up with two tools, and that is a reasonable outcome.
The 13 tools, reviewed
#1 Recraft
Raster and vector image generator with its own models · United States · recraft.ai
Vector output Own models Team opt-out by default
Recraft writes down the things a buyer normally has to extract from a sales call. Its documentation states the price of every tier, says that paid subscribers own what they generate, and sets out training use plan by plan: no opt-out on Free, a profile switch on Pro, opted out by default on Teams, and never for API inputs and outputs. The models are Recraft's own (V2 through V4.1) and they produce vector files as well as raster ones, which matters for logos, icons and anything a designer has to scale. Teams adds shared styles, central account management and SSO from the 2,000-credit tier at $22 per month.
Where it falls short
The free plan is a demo, not a working tier: Recraft owns those images, shows them in its community gallery and does not license them for commercial use. Commercial rights are worded as applying while you are subscribed, so check what lapsing means for work in progress. On Pro, training use is on until somebody finds the switch. The terms have you indemnify Recraft, not the reverse, and they are governed by New York law.
Wrong for
A company that wants to choose between several labs' models in one subscription will find only Recraft's own here; an aggregator such as Krea or Magnific fits that better. Buyers needing vendor indemnity should read Bria or Adobe Firefly.
Pricing: Basic is $12 per month billed monthly, or $10 per month billed annually, for 1,000 credits; Pro and Teams tiers scale from 2,000 to 16,000 credits, and the API sells 1,000 units for $1.00. (free plan)
Visit Recraft →
#2 Krea
Multi-model image and video workspace metered in compute units · United States · krea.ai
Model aggregator Compute units SAML SSO on Enterprise
Krea solves the problem of not knowing which model will suit next month's brief. Its model overview lists its own Krea 1 and Krea 2 Turbo next to Flux models from Black Forest Labs, Google's Nano Banana line, OpenAI image models, Seedream and Qwen, each with a stated cost: 2 credits for Krea 2 Turbo, roughly 400 for Flux 2 Max. That published table lets a team budget before it generates. The free tier refreshes 100 compute units daily, a self-serve Business plan covers up to 50 seats with units that roll over, and the Enterprise page promises SOC 2, SAML SSO and full IP indemnification.
Where it falls short
The spread between models is wide enough to wreck a budget: one frontier-model image costs about what 200 Krea 2 Turbo images do, and individual plans lose unused units each month. Compute packs expire after 90 days. The never-train statement and the indemnity we found sit on the Enterprise page, which starts at 50 seats and is quote-only, while the general terms grant Krea a perpetual, sublicensable license to user submissions.
Wrong for
A small team that needs a written no-training commitment at a published price should look at Recraft Teams. Developers who only want API calls will find fal or Replicate closer to what they need than a workspace subscription.
Pricing: Free tier with 100 compute units a day; paid plans start at $10 per month with 20% off for annual billing (Basic 4,000, Pro 20,000 and Max 54,000 compute units a month); Enterprise is custom. (free plan)
Visit Krea →
#3 Black Forest Labs (FLUX)
FLUX image models by API or as open weights · Germany · bfl.ai
Open weights Per-image API German company
Black Forest Labs is on this list because it is the one model maker here that lets a buyer leave the hosted service without leaving the model. The API charges per image by megapixel, from $0.014 for FLUX.2 [klein] 4B to $0.07 and up for FLUX.2 [max]. The same family is published as weights: klein 4B under Apache 2.0, free for commercial use, and the larger ones under a non-commercial license that a paid tier (Builder at 10K images a month, up to Enterprise) converts into commercial rights with fine-tuning. The company is BFL GmbH in Freiburg, Germany, which puts a European model lab within reach of a small team.
Where it falls short
This is a developer product with a playground attached, not a workspace: we found no seats, shared libraries or brand kits on its site. By default the terms let BFL train FLUX models on your inputs and outputs; opting out means emailing the legal address, and it only applies going forward. Self-hosting license prices are not printed on the licensing page, and the terms are governed by Delaware law despite the German entity.
Wrong for
Marketers who want to type a prompt, tidy the result and export a social post should use a suite such as Magnific or Adobe Firefly, several of which serve these same FLUX models behind a friendlier interface.
Pricing: API is prepaid in credits worth $0.01 each: FLUX.2 [klein] 4B from $0.014 per image, FLUX.2 [pro] from $0.03, FLUX.2 [max] from $0.07, priced by megapixel; self-hosting licenses are tiered by monthly image volume with prices shown only in the dashboard or by quote. (open source)
Visit Black Forest Labs (FLUX) →
#4 Magnific (formerly Freepik)
AI image and video suite with a stock library attached · Spain · magnific.com
Stock plus AI Yearly credit pool Spanish company
The banner on its legal pages says it plainly: Freepik is now Magnific. What a business gets is a suite that treats generation and stock as one budget. The pricing page names models from other labs, Nano Banana 2 and Nano Banana Pro among them, alongside video models such as Google Veo 3.1 and Kling 2.5, and sits them next to a library it counts at more than 250 million photos, videos, vectors and PSDs. Three individual plans are published in euros, and Premium+ is described as unlimited generations. The operator is Freepik Company in Málaga, Spain, so the counterparty is a Spanish company.
Where it falls short
The pricing page lists a commercial AI license and music rights on Premium+ and Pro Starter; we did not see it on the entry Premium plan, so confirm before using that tier for client work. The general terms give an indemnity for stock content and state that it does not apply to AI-generated content. No free plan or trial is advertised on the pricing page we fetched, team prices were not visible, and several product pages refused our fetch.
Wrong for
Anyone who needs to know in writing whether prompts and uploads train models should pick a vendor whose policy we could read, such as Recraft or Adobe Firefly. Developers wanting an API should start with Black Forest Labs or fal.
Pricing: Premium is €16 per month or €12 per month billed annually (240K credits a year); Premium+ is €36 or €27 (600K credits a year); Pro Starter is €88 or €66 (1.5M credits a year). (none)
Visit Magnific (formerly Freepik) →
#5 Adobe Firefly
Adobe's generative image, video and audio app · United States · adobe.com
Licensed training data Partner models Content Credentials
Adobe is the incumbent in creative software, so the majors rule keeps it out of the first three places; on the evidence it would otherwise sit higher. Its statements are the clearest on this list: Firefly models are trained on licensed content such as Adobe Stock and on public-domain work, Adobe says it does not and never has trained Firefly on user content, and outputs from Firefly models are described as safe for commercial use with Content Credentials attached. The app also offers partner models, GPT Image, Gemini (Nano Banana) and FLUX among them, and Adobe says user content is not used to train those either. Plans run from US$9.99 to US$199.99 a month.
Where it falls short
The commercial-safety language is about Adobe's own Firefly models, and the indemnification is for eligible enterprise customers and select outputs, so a team on a US$19.99 license choosing a partner model should not assume either applies. Credits are the meter, and we could not fetch Adobe's credit FAQ to see how many a given model burns. The two upper tiers were showing a first-year promotional price when we looked, which flatters year one.
Wrong for
A developer who wants weights to run on company hardware gets nothing of the kind from Adobe; Black Forest Labs, Stability AI and Bria do that. Teams with no other Adobe subscription may find Recraft simpler to buy.
Pricing: Firefly Standard is US$9.99/mo with 2,000 generative credits, Firefly Pro US$19.99/mo with 4,000; team licenses start at US$19.99/mo per license on annual billing, and there are free daily generations. (free plan)
Visit Adobe Firefly →
#6 Bria
Licensed-data image models for API or private deployment · Israel · bria.ai
Licensed training data Indemnification On-prem option
Bria is the choice when legal sign-off is the hard part of the purchase. It says all its models are trained exclusively on licensed data from Getty Images, Alamy and Envato, and indemnification appears on the pricing page itself: capped standard cover on pay-as-you-go, unlimited IP indemnification on Enterprise. The meter is per generation, from $0.018, after 100 free ones. Enterprise buyers can run the models in their own AWS or Azure account, on their own hardware or air-gapped, and Bria's docs say hosted inputs are not retained beyond processing while generated images expire after three days. The model is FIBO, driven by structured prompts as well as text.
Where it falls short
It is built for developers: the docs lead with SDKs, a REST API and an MCP server, so expect to need engineering help to get value from the hosted plans. Free and pay-as-you-go tiers are cloud only. Output images expire after three days by default, so a pipeline has to copy them somewhere. Indemnification is listed for pay-as-you-go and Enterprise; we did not see it listed for the free plan, and we could not open a terms page to read the cap.
Wrong for
A designer who wants a canvas, a style library and teammates in one browser tab should pick Recraft or Krea. Buyers who care more about choosing among many labs' models than about training-data provenance fit fal better.
Pricing: Free plan with 100 free generations; pay-as-you-go starts from $0.018 per generation; Enterprise is custom-priced with on-prem or bring-your-own-cloud deployment. (free plan)
Visit Bria →
#7 Photoroom
AI product photography and background editing · France · photoroom.com
Product photos Batch editing French company
Most generators start from a blank prompt. Photoroom starts from the photo a merchant already has, which is the more common business case: remove the background, stage the product, add a shadow, repeat for the whole catalog. It runs as a web app, on iOS and Android, inside Shopify and as an API priced per image, $0.02 for background removal and $0.10 for the tier with AI backgrounds, virtual try-on and product staging. Subscription tiers are published, the company is Photoroom SAS in Paris, and the site claims SOC 2 Type 2. Its terms say the customer owns the content.
Where it falls short
It is not a general text-to-image tool, and a brand campaign with no product photo to start from is outside its range. The terms permit Photoroom to use customer content for AI training unless the account holder switches that off in settings, so the default is opt-out. 4K output is held for the Ultra plan, exports are capped monthly on every tier, and SSO and an SLA are Enterprise items. Web prices are quoted per month on yearly billing.
Wrong for
Illustrators, brand designers and anyone needing logos or vector art should use Recraft. A team after campaign imagery from many different models belongs on Krea or Magnific.
Pricing: Pro is $9/month billed yearly with 8,000 AI credits, Max $19/month with 25,000 and Ultra $59/month with 75,000; the API charges $0.02 per image for background removal and $0.10 per image for the Plus editing tier. (free trial)
Visit Photoroom →
#8 Google Gemini API (Nano Banana)
Google's Nano Banana image models by API · United States · ai.google.dev
Per-image API Big tech SynthID watermark
Several products on this page resell Google's Nano Banana models; this entry is the source. The pricing table converts token prices into per-image figures by resolution, so a 1K image from Gemini 3.1 Flash Image is $0.067 and a 4K image from Gemini 3 Pro Image is $0.24. The API terms separate two regimes: on paid quota Google does not use prompts or responses to improve products, and it says it will not claim ownership of generated content. The models accept up to 14 reference images and can ground an image in Google Search results. As a big-tech platform Google is marked major and ranked below the independents.
Where it falls short
There is no free tier for these image models, and the terms for unpaid services elsewhere in the API allow content to be used for product improvement and read by human reviewers, so a team must be sure it is on paid quota. Every image carries a SynthID watermark. This is an API with token math, not an application: no brand kit, no seats, no gallery. The terms we read contain no indemnity clause of their own.
Wrong for
Non-technical teams should reach Google's models through a suite such as Krea, Magnific or Adobe Firefly. A buyer that wants weights it can host independently of any cloud should look at Black Forest Labs or Stability AI.
Pricing: Paid tier only for the image models: Gemini 3.1 Flash Image works out at $0.067 per 1K image and $0.151 per 4K image; Gemini 3 Pro Image at $0.134 per 1K or 2K image and $0.24 per 4K image. (none)
Visit Google Gemini API (Nano Banana) →
#9 fal
API marketplace for image, video and audio models · United States · fal.ai
Model marketplace Per-image API SOC 2
fal is infrastructure: one account, one invoice, and an API in front of models from Black Forest Labs, Google, OpenAI, Alibaba, ByteDance and others. Its pricing page quotes image models by output, $0.03 for a Seedream V4 image, $0.04 for Flux Kontext Pro, $0.02 per megapixel for Qwen, and notes that higher resolutions scale proportionally. That gives a product team a way to switch model by changing a string instead of signing a new contract. Serverless endpoints bill per second and dedicated GPUs by the hour for teams running their own fine-tunes. The site states SOC 2 compliance, and the enterprise page says enterprise customers' data is never used to train fal's models.
Where it falls short
The no-training sentence we found is on the enterprise page and speaks of enterprise customers; we saw no equivalent for a self-serve account. The terms make no promise that output will be original and have the customer indemnify fal. Each hosted model also comes from a lab with its own rules, which fal does not replace. No free plan or trial is advertised on the pricing page, and there is no interface for a non-developer.
Wrong for
A five-person marketing team with nobody to write code should not start here; Recraft or Magnific will be in use the same day. Buyers needing training-data provenance and indemnity should read Bria.
Pricing: Pay per output: Seedream V4 is $0.03 per image, Flux Kontext Pro $0.04 per image and Qwen $0.02 per megapixel, with GPU compute billed separately by the hour. (none)
Visit fal →
#10 getimg.ai
Multi-model creative workspace from Poland · Poland · getimg.ai
Model aggregator Credit tiers Polish company
getimg.ai is the smaller European take on the aggregator idea. The operator is Webrockets Sp. z o. o. in Warsaw, and the product puts Seedream, FLUX.2, Nano Banana, GPT Image, Grok Imagine and Qwen image models, plus video and audio models, behind one credit balance. Four tiers are published, each with a stated monthly credit count and a stated number of simultaneous generations (two on Entry, ten on Ultra), and every paid plan includes commercial rights. From Core upward all models are available. Its terms say the company claims no intellectual property rights over material users provide, and the site describes shared team workspaces.
Where it falls short
Neither the terms nor the privacy policy we read says whether prompts and uploads are used for model training, which leaves the question every business buyer asks unanswered. Entry is limited to selected models. The pricing page mentions no free plan or trial, so evaluation costs a month's fee. Its commercial-use answer still points to the CreativeML Open RAIL-M license, a Stable Diffusion-era reference that sits oddly beside the models now on offer.
Wrong for
Organizations that need SSO, an indemnity or a documented training policy should choose Recraft Teams, Adobe Firefly or Bria. Developers should buy per image from fal or Black Forest Labs.
Pricing: Four published monthly tiers (Entry, Core, Plus, Ultra) from 8,000 to 107,500 credits a month, cheaper on yearly billing; the two pages we fetched showed different currencies, so no figure is quoted here. (none)
Visit getimg.ai →
#11 Generative AI by Getty Images
Indemnified image generation from a stock agency · United States · gettyimages.com
Indemnified output Licensed training data Prepaid packs
Getty approaches generation as a licensing product. Its page says the generator is powered by a custom Bria model trained exclusively on licensed content, including Getty's own creative collection, that contributors whose work is in the model are compensated, and that each generated image comes with legal protection starting at $50,000 USD. The license is royalty-free, perpetual and worldwide, and Getty says it will not add images a customer generates to its library for others to license. For a company that publishes a handful of high-exposure visuals a year and wants a known counterparty behind each one, that package is coherent.
Where it falls short
At the prices shown, €45 for 25 generations works out far above the per-image rates elsewhere on this page, and a generation is spent whether or not the result is usable. Packs are prepaid and lapse after one year. We were redirected to a Dutch storefront, so prices in other markets may differ. The page did not say whether prompts are retained or reused, and we found no team workspace or API described for the generator.
Wrong for
Teams that iterate through hundreds of drafts a week will burn packs quickly and should use a credit plan from Recraft or Krea. Developers wanting the same licensed-data approach by API can go to Bria directly.
Pricing: Sold in prepaid packs: the Dutch storefront we were redirected to showed introductory prices of €45 for 25 generations and €139 for 100, valid for one year. (none)
Visit Generative AI by Getty Images →
#12 Stability AI
Stable Diffusion models to self-host or call by API · United Kingdom · stability.ai
Self-hostable Community License UK company
The Community License is the reason to read this entry. Stability AI says its core models, the Stable Diffusion 3.5 suite included, are free for everyone unless they are used commercially by an organization with more than USD $1M in annual revenue; above that line an Enterprise License is required. It also says users own the outputs. A small agency with a capable GPU can therefore run image generation with no per-image fee and no prompts leaving the building. The same models are offered through an API with editing endpoints, through cloud partners and in a web app called Stable Assistant. The contracting entity for UK and EU customers is Stability AI Ltd in London.
Where it falls short
The developer pricing page returned nothing readable to us, so we cannot state what a credit costs; we only saw that a Stable Assistant image uses 6.5 credits and that several edit endpoints cost 5. For the hosted services the terms allow inputs and outputs to be used for training unless the user opts out. Self-hosting shifts the work to you: hardware, updates and safety filtering. Cross the revenue threshold and the price becomes a sales conversation.
Wrong for
A marketing department without an engineer or a GPU budget should buy a hosted suite instead; Recraft and Krea publish their numbers. Companies over the revenue line that want indemnity with their license should compare Bria.
Pricing: Core models are free under the Community License unless you use them commercially and your organization makes over USD $1M a year, in which case a paid Enterprise License is required; the API and Stable Assistant are metered in credits whose dollar price we could not read. (free plan)
Visit Stability AI →
#13 Replicate
Pay-per-use API for running published AI models · United States · replicate.com
Model hosting Per-image or per-second Short API retention
Replicate is where a developer goes to try a model this afternoon. Its price list mixes two meters: official image models are charged per output, $0.04 for FLUX 1.1 Pro, $0.09 for Ideogram V3 Quality, $3.00 per thousand for FLUX Schnell, while everything else is charged for the seconds of hardware it occupies. Two details matter to a business. The terms grant the customer the rights in output, commercial use included. And the docs say API inputs, outputs and logs are removed automatically after an hour by default, which is a short retention window compared with most hosted tools. An enterprise tier adds SLAs, higher GPU limits and volume discounts.
Where it falls short
It sits last because it offers the least beyond raw access: no workspace, no team features we could find, no statement on the pricing page of a free allowance. Predictions made through the web interface are kept indefinitely, unlike API calls. Per-second billing makes costs hard to predict for models without a per-image price. The about page sends job seekers to Cloudflare's careers site, so ask who the contracting party is before signing.
Wrong for
Non-developers will find nothing to click on here. A company that wants one vendor accountable for the model itself, with training data and indemnity documented, should deal with Bria, Adobe or Black Forest Labs directly.
Pricing: Usage-based: FLUX 1.1 Pro is $0.04 per output image, FLUX Schnell $3.00 per thousand output images, Ideogram V3 Quality $0.09 per output image; other models bill by hardware time, for example $0.000225 per second on an Nvidia T4. (none)
Visit Replicate →
Questions and answers
What is the best AI image generator in 2026?
For a small or mid-sized business, Recraft ranks first here: it publishes every price, produces vector as well as raster output, gives paid subscribers ownership, and documents training use plan by plan, with Team workspaces opted out by default. Krea, second, is the better pick when you want to choose among many labs' models under one subscription and test on a daily free allowance. Black Forest Labs, third, suits a team building generation into its own product, because the FLUX models can be bought per image or licensed to run on your own servers.
Do I own the images an AI generator makes for my business?
It depends on the vendor and often on the plan. Recraft assigns ownership to paid subscribers but keeps ownership of free-plan images. Stability AI and Replicate state that the user owns or is granted the rights in outputs, and Google says it will not claim ownership of generated content. Owning the vendor's share of the rights is not the same as holding an enforceable copyright, which depends on the law where you operate and on how much human authorship went into the work. Treat the contract as settling the question between you and the vendor only.
Can AI-generated images be used commercially?
Usually yes on paid plans, and frequently no on free ones. Recraft does not license free-plan images for commercial use. getimg.ai allows commercial use only on paid plans. Magnific lists a commercial AI license on its Premium+ and Pro Starter plans. Adobe describes outputs of its own Firefly models as safe for commercial use. Stability AI permits commercial use of its core models for free until an organization passes USD $1M in annual revenue. Check the wording for the exact tier you are buying, and check again if you downgrade.
Which AI image generators offer indemnification?
Four vendors on this list put an indemnity in writing on the pages we read. Bria lists capped standard indemnification on pay-as-you-go and unlimited IP indemnification on Enterprise. Adobe offers IP indemnification to eligible enterprise customers for select Firefly outputs. Getty attaches legal protection starting at $50,000 USD to each generated image. Krea's Enterprise page promises full IP indemnification. Elsewhere the direction is reversed: the terms of Recraft and fal have the customer indemnify the vendor. None of these is a substitute for reading the cap and the exclusions.
Will my prompts and uploaded images be used to train the model?
Sometimes, by default. Black Forest Labs trains on inputs and outputs unless you email an opt-out, and Photoroom's terms allow training unless you switch it off in settings. Recraft trains on Free, lets Pro users opt out, opts Team plans out by default and never trains on API traffic. Stability AI offers an opt-out for its hosted services. Adobe states it has never trained Firefly on user content. Google does not use paid-tier API prompts for product improvement but does use unpaid-tier content. Where a vendor says nothing, as with getimg.ai, assume nothing.
How much does an AI image generator cost per image?
On per-image APIs the published range runs from fractions of a cent to around a quarter of a dollar. Recraft's API charges $0.007 for its Flash model and $0.04 for V4. Black Forest Labs starts at $0.014 for FLUX.2 [klein] 4B and $0.07 for FLUX.2 [max]. Google charges $0.067 for a 1K image from Gemini 3.1 Flash Image and $0.24 for a 4K image from Gemini 3 Pro Image. Bria starts from $0.018 per generation. Subscription credits make the per-image figure depend on which model you select, so work it out for that model.
What is the difference between a credit subscription and an API for image generation?
A credit subscription buys access to an application: a canvas, a gallery, styles, sometimes seats and SSO, with a monthly allowance that usually expires. An API buys raw generation that your own software calls, billed per image, per megapixel or per second of compute, with no interface for non-developers. Some vendors keep the two wallets apart: Krea's API draws on a separate dollar balance from the compute units in its web app, and Recraft sells API units separately from plan credits. Choose by who will press the button, a designer or a program.
Can a business run an image model on its own servers?
Yes, with three vendors here. Black Forest Labs publishes FLUX.2 [klein] 4B under Apache 2.0 for free commercial use and sells licenses for its larger open-weight models in tiers by monthly image volume. Stability AI's Community License allows free commercial use of Stable Diffusion for organizations under USD $1M in annual revenue, with an Enterprise License above that. Bria offers bring-your-own-cloud on AWS or Azure, on-premises and air-gapped deployment on its Enterprise plan. Self-hosting removes per-image fees and keeps prompts in-house, but adds hardware and maintenance.
Which image generators use models from other companies?
Several. Krea lists models from Black Forest Labs, Google, OpenAI, ByteDance and Alibaba beside its own. Magnific names Nano Banana models on its pricing page. getimg.ai offers Seedream, FLUX.2, Nano Banana, GPT Image and Qwen. fal and Replicate are hosting platforms for other labs' models. Adobe Firefly adds GPT Image, Gemini and FLUX as partner models. Getty says its generator is powered by Bria. When a product resells a model, its promises about training and indemnity may cover only some of the models in the menu, so ask which.
Are images from free AI generators private?
Often not. Recraft states that free-plan images are publicly visible in its community gallery and owned by Recraft, with private generation reserved for paid plans. Google's terms for unpaid API services allow content to be used to improve products and read by human reviewers, and advise against submitting confidential information. A free tier is a fair way to judge an interface, and a poor place for an unreleased product, a client's brief or a staff photo. Move to a paid plan, and set the training option, before uploading anything you would not publish.
Do AI image generators watermark or label their output?
Some do, and it is worth knowing before images go into an ad platform that checks provenance. Google states that every image generated through its Gemini image models includes a SynthID watermark. Adobe says Content Credentials are attached automatically to Firefly output. Photoroom watermarks results from its free API sandbox, which is a trial restriction and not a provenance label. For the other vendors on this page we did not find a statement either way on the pages we read, so ask if disclosure rules in your market require a label.
