Best AI Image Generator for Business in 2026: Model, Meter, Contract, and Exit Risk

Last updated: August 18, 2026

An AI image platform is a stack: a model version, an inference configuration, a billing unit, a data route, a contract, and a continuity policy. Two services can expose the same model family and still differ in reproducibility, unit cost, confidentiality, client-use rights, and what happens when the provider replaces the model. This briefing compares the documented layers of that stack across five operating routes.

Verdict in one line: treat visual quality as a pass or fail gate that this briefing did not test, then choose the operating route whose pricing, data handling, deployment rights, and model continuity your team can actually manage.

A model name is not a service

Quality still decides whether a model enters your shortlist at all. A model that fails the visual brief is rejected, and no contract rescues it. This briefing takes that gate as given and starts after it, because the gate is the part a document cannot test and the part almost every comparison article already covers.

What documents can show is that a shared model name discloses less than it appears to. Black Forest Labs demonstrates this inside its own product. Its API reference documents named parameters including seed, guidance, steps, a safety_tolerance scale running from 0 to 6 where 0 is most strict, and prompt_upsampling, which the documentation describes as automatically modifying the prompt for more creative generation. The FLUX.2 repository shows that the prompt expansion step can be run through a local text model or through a remote model selected on OpenRouter.

Parameter exposure differs by endpoint. A platform that lists “FLUX.2” without naming an endpoint or version has disclosed less than the family name suggests, and building a parameter table by endpoint rather than by family is the only way to know what you actually control.

Adobe states the same underlying point from the opposite direction. In its Firefly product description, Adobe explains that two users entering identical prompts may receive the same or very similar output if the prompt query shares the same seed, and describes that seed as a randomness variable with billions of options. Somebody chooses the seed. On a hosted platform, that somebody is not you.

None of this proves that two platforms serve different weights. It proves something narrower and more useful: a family name does not establish the same checkpoint, endpoint version, inference settings, preprocessing, or postprocessing. Two services can expose the same name and deliver different results without either misrepresenting anything.

Sources: Black Forest Labs, flux2 prompt upsampling documentation, opened 17 August 2026 · Adobe, Firefly IP Indemnification, published 13 July 2026, opened 17 August 2026

Adobe Firefly: availability is not coverage

Adobe’s meaningful differentiator is not that it has a contract. Every paid service has one. It is that Adobe publishes a supplemental output indemnity extending to a defined subset of third-party models, which most competitors do not publish at all. The value of that document depends entirely on reading which models it reaches and which conditions it attaches.

The eligible subset is narrower than the roster

The Creative Partner Models Product Description, effective May 8, 2026 and opened by FSR on August 17, 2026, names the eligible subset: Google image and video generation models, and OpenAI image and video generation models. A footnote excludes anything designated beta, trial, preview, early access, evaluation, or similar wording indicating a pre-general-availability release. Adobe adds that it does not guarantee access to the listed models.

Adobe’s partner model roster, in the version FSR captured on August 17, 2026 and marked updated August 3, 2026, listed partner models from Black Forest Labs, Luma AI, Runway, Topaz Labs, ElevenLabs, Kuaishou, and ByteDance alongside Google and OpenAI.

So a buyer has two separate questions, and answering the first does not answer the second.

1. Is this model available to our plan and region?
2. Is output from this exact model eligible under our agreement?

What the coverage document actually says

The PSLT for Creative Partner Model Supplemental Coverage, version 2026v1, defines Covered IP Rights as patent, copyright, moral rights or similar rights. It extends the defined Infringement Claim to proceedings alleging that unmodified eligible output directly misappropriates a trade secret or infringes those rights, and only to proceedings made during the License Term. It defines a Legal Proceeding as a formal proceeding filed before a court or government tribunal, which does not on its face include a demand letter.

Its exceptions remove liability to the extent a claim arises from combining the output with any other material, content or information; from use in commerce violating a third party’s trademark or related rights; from the customer’s own input or supplied data and models, including in connection with customization or fine-tuning; from use after Adobe has instructed the customer to stop; and from disabling or circumventing Adobe’s filters or safety systems.

The document does not define “unmodified.” It does not state how cropping, resizing, color correction, format conversion, or placing an image inside a layout are classified. Each exception uses “to the extent” language, which limits its reach in a way the public text does not resolve. A buyer cannot answer these from the PSLT alone.

The PSLT is not the whole contract

Adobe’s enterprise licensing terms sit above and around the PSLT. The General Terms contain the broader technology indemnity, the procedural conditions governing notice and control of the defense, and the aggregate liability cap. A separate Generative AI Features Terms document also sits in the precedence order. The General Terms are published in regional versions, so the terms applying to a Japanese or Australian entity are not necessarily the version a North American buyer reads.

FSR did not open the General Terms for this briefing. That is a gap, and it matters, because the cap and the procedural conditions materially change what the indemnity is worth. If you are evaluating Firefly on indemnity grounds, this is the document to read next, and the four fields to find are: the aggregate liability cap, the notice and cooperation conditions, who controls the defense and settlement, and the order of precedence between your Sales Order and everything else.

The defensible conclusion is narrower than “Adobe indemnifies commercial AI images.” It is this: Adobe publishes supplemental output coverage for a defined subset of partner models, subject to model eligibility, agreement and plan requirements, output-specific exclusions, procedural conditions, and a liability cap that sits in a document outside the PSLT.

The rate card and the entitlement are different things

Adobe’s roster page carried a per-model credit consumption table when FSR captured it on August 17, 2026. The published rates spanned 5 credits per generation at the low end to 80 credits per generation at the high end, a 16 to 1 spread inside one table.

That is a rate card. It is not a monthly output forecast. Adobe’s generative credits FAQ states that Creative Cloud Pro, Firefly, and credits plans include unlimited standard generations and that credits are spent on premium features. Adobe’s plan pages have also carried promotional entitlements. Dividing a monthly allowance by a credit rate can therefore produce a calculation that is arithmetically correct and commercially wrong.

Before pricing Firefly, record the plan, billing interval, region, promotional term, the specific model, the output resolution, and whether that model sits inside any unlimited set on your plan. A credit table without that mapping is not a cost forecast.

The data path is disclosed asymmetrically

Adobe disclosed on the captured page that generating with Seedance 2.0, Kling 3.0, or Kling 3.0 Omni requires sending prompts and reference files to the Chinese company that developed them, and that those models are not available on Creative Cloud for enterprise plans. A general FAQ answer on the same page, covering all partner models, stated that Adobe shares the text prompt entered and any reference image or video files uploaded.

The structural fact is the same in both passages. Prompts and reference files leave Adobe and reach the model developer. What differs is how prominently each is said. Adobe also stated that customer content is never used to train partner or Adobe models, and that contractual protections and technical safeguards are in place. Those are Adobe’s representations about its own arrangements and about third parties. FSR reports them as such and did not verify them independently.

Sources: Adobe, PSLT Creative Partner Model Supplemental Coverage 2026v1 · Adobe, Creative Partner Models Product Description, effective 8 May 2026 · Adobe, Partner models in Adobe products, captured 17 August 2026 · Adobe, Generative credits FAQ, opened 17 August 2026 · Adobe, Enterprise licensing terms index · Adobe, Generative AI Features Terms

Black Forest Labs: control is per endpoint, not per family

Calling a model API directly gives you a calculable unit rate and named endpoints. OpenAI and Google publish developer rates too, so this is not unique. What is specific to Black Forest Labs is the pricing unit and the parameter surface.

BFL’s published pricing, opened August 17, 2026, sets one credit at one US cent and prices FLUX.2 image generation by output megapixel rather than per image. The table listed text-to-image starting at $0.014 for klein 4B, $0.015 for klein 9B, $0.03 for pro, $0.05 for flex, and $0.07 for max, with pro image editing starting at $0.045. Klein charges a flat rate for the first megapixel and adds for each additional one, illustrated as $0.014 plus $0.001 for a 2 megapixel klein 4B image.

Two lines on that page change how a budget is built. The documented batch formula multiplies the base cost by the number of images, so no discount appears at the request level. And FLUX 3 is listed as the video line, priced per second for text-to-video, image-to-video, and video continuation. A brief written around “FLUX 3 for the key visuals” is specifying a different product category.

Parameter access is endpoint-specific. The controls named above are documented for the endpoints FSR opened, and other hosted endpoints expose different subsets. Build the table by endpoint before you commit to a workflow that depends on a specific control.

There is also a refund position to read before topping up. The FLUX API Service Terms state that all credit purchases are non-refundable. FSR did not locate a published policy on whether failed or content-filtered requests are billed, and searched docs.bfl.ml and bfl.ai for it. That is recorded as a search result, not as an absence of policy.

Sources: Black Forest Labs, API pricing, opened 17 August 2026 · Black Forest Labs, pricing page, opened 17 August 2026 ·Black Forest Labs, FLUX API Service Terms

Open weights: one phrase, three different rights regimes

“Open weights” does not identify the license that governs deployment. Three distinct regimes hide behind the phrase, and buying the wrong one is how agency work gets delivered under terms that did not permit it.

Regime 1: permissively licensed local deployment

Black Forest Labs states on its FLUX.2 klein 4B model card that it approved release of that checkpoint under Apache 2.0. The card lists roughly 13GB of VRAM and an RTX 3090 or 4070 and above. Alibaba’s Qwen-Image is published under Apache 2.0 and HiDream-I1 under MIT.

A fully local deployment can keep prompts and reference files inside infrastructure you control, but only if the surrounding pipeline has no external dependency. Remote prompt expansion, hosted storage, telemetry, and third-party logging each reintroduce an egress path that the model license says nothing about. Audit the pipeline, not the checkpoint.

Regime 2: non-commercial checkpoints from the same vendor

The same model card states that BFL released the FLUX.2 klein 9B checkpoints under a non-commercial license. BFL’s pricing documentation lists the 32B dev checkpoint as local only, open weights, non-commercial, with no hosted API. The card also notes that filters or manual review must be used with the 9B models under the terms of that license, while deployers of 4B are encouraged to implement the same mitigations.

Tencent’s HunyuanImage 3.0 uses the Tencent Hunyuan Community License, which is its own instrument rather than a standard open-source license.

The rule that follows is unglamorous and load-bearing: inspect the license file for the exact variant you intend to deploy. Rights vary by checkpoint inside a single vendor’s own family, so nothing about the family name, the vendor, or the download button tells you what you may do.

Note also that a non-commercial model license restricts commercial deployment of the model. It does not automatically restrict commercial use of the images you generate. Those are two different permissions and conflating them in either direction produces a wrong purchase.

Regime 3: paid commercial self-hosting

BFL’s open weights licensing page, captured August 17, 2026, listed a Builder tier covering klein models at 10,000 images per month, one domain, and ten licensed users, described as not meant for client use or downstream applications. Platform covered klein 9B and dev at 100,000 images per month, one domain, ten users. Professional covered dev at 100,000 images per month across up to three domains. Rights to use outputs as AI training data appeared as a separate Synthetic Data tier rather than as an included right.

Two further clauses belong in an agency’s evaluation. The Self-Hosted Commercial License Terms state that the vendor will have no obligation to provide updates, maintenance, or support. And self-hosting under a paid commercial agreement is not automatically perpetual: commercial terms of this kind commonly reserve removal, reporting, and deletion rights, so read yours rather than assuming that a downloaded file is permanent.

Provenance changes with the route

BFL states that its API applies cryptographically signed C2PA metadata to downloaded output. The open-weight inference code, by contrast, includes an example of pixel-layer watermarking. If a client asks you to prove which model produced an asset, the route you used determines what the file can show.

Sources: Black Forest Labs, FLUX.2 klein 4B model card, opened 17 August 2026 · Black Forest Labs, API pricing, opened 17 August 2026 · Black Forest Labs, Open Weights Licensing, captured 17 August 2026 · Black Forest Labs, Self-Hosted Commercial License Terms · Black Forest Labs, Non-Commercial License Terms · Alibaba, Qwen-Image model card · HiDream, HiDream-I1 repository · Tencent, HunyuanImage-3.0 repository

Midjourney: a closed system billed in GPU time

Midjourney sells access to a closed creative system. FSR did not test whether its visual output is distinctive, which is a genuine question this briefing cannot answer. What the documentation does answer is how it bills, what it publishes by default, and how you would leave.

Midjourney’s plan comparison, opened August 17, 2026, lists four plans, automatic monthly or yearly renewal, and a 20 percent discount for annual payment. What the plan buys is a set amount of GPU time. The GPU speed documentation supplies one conversion figure and hedges it: a typical image prompt usually takes about one minute of GPU time. Unused Fast time does not carry over.

That approximation is the only bridge Midjourney publishes between its meter and anyone else’s. Comparison tables that print a Midjourney price per image are computing from an estimate the vendor describes as typical.

Public is the default and privacy is prospective

Midjourney’s terms of service, dated June 23, 2025, state that content is publicly viewable and remixable, and that users grant Midjourney a license. Its privacy documentation states that the community gallery can contain every image and video you create. Stealth Mode is available only to Pro and Mega plan members and does not hide creations made in public Discord channels or web Rooms.

The line that matters for anyone upgrading mid-project: Midjourney states that Stealth Mode does not automatically make existing public creations private, and that the upgrade lets you make already-generated work private with bulk changes from the Organize page. Privacy applies going forward. Earlier work needs a manual pass.

Midjourney also states that companies with more than $1,000,000 USD in annual gross revenue must use the Pro or Mega plan.

Automation route

FSR searched midjourney.com for an official API using the queries “API,” “official API,” “developer API,” and “API documentation,” and located no official API offering. That is recorded as not found in the sources searched rather than as a confirmed absence. If your workflow depends on programmatic generation, repeat that search and get a written answer before committing.

Sources: Midjourney, Comparing Plans, opened 17 August 2026 · Midjourney, GPU Speed, opened 17 August 2026 · Midjourney, Stealth Mode, opened 17 August 2026 · Midjourney, Keeping Your Creations Private, opened 17 August 2026 · Midjourney, Using Images and Videos Commercially, opened 17 August 2026 · Midjourney, Terms of Service, 23 June 2025

Bundled assistant plans: check what you already hold

Before buying a separate platform, check what your team’s existing assistant subscription already permits. Buyers with an eligible ChatGPT or Google AI plan may already have image generation, and skipping that check is how teams pay twice for the same capability.

The constraint is that the consumer-side entitlement is not quantified in the pages FSR opened. OpenAI’s plan comparison labels free-tier image generation “Limited” and elsewhere describes it as limited and slower. The free tier FAQ, dated August 13, 2026, states that image generation has separate usage limits and that ChatGPT will notify you when a limit is reached. No figure appears. Google’s plan benefits documentation states that each product has its own AI usage limits, that limits depend on which features you use, and that AI credits can be purchased. Also no figure.

The developer documentation from the same two companies is granular by comparison. OpenAI’s model page for chatgpt-image-latest listed 1024×1024 generations at $0.009, $0.034, and $0.133 across three quality settings when FSR opened it on August 17, 2026. Its API pricing page listed gpt-image-2 at $8.00 per million image input tokens and $30.00 per million image output tokens. Google’s Gemini API pricing listed image output at $60 per million tokens with worked per-image equivalents of $0.045 at 0.5K, $0.067 at 1K, $0.101 at 2K, and $0.151 at 4K.

Both vendors publish granular API rates while describing some consumer-plan image limits qualitatively. For a team trying to forecast production cost, that difference is the whole point: an allowance can be adjusted without changing a published price.

Sources: OpenAI, ChatGPT plans, opened 17 August 2026 · OpenAI, Free Tier FAQ, 13 August 2026 · OpenAI, API pricing, opened 17 August 2026 · OpenAI, chatgpt-image-latest model page, opened 17 August 2026 · Google, Gemini Developer API pricing, opened 17 August 2026 · Google, Use Google AI Pro benefits, opened 17 August 2026

Retirement and version control

Model continuity is a procurement variable, and it is the one comparison articles omit most consistently.

Adobe maintains a public retirement page for partner models. The version FSR captured recorded eight dated entries covering ten named models between April 22 and July 23, 2026. Seven entries were labeled Deprecated and one, Pika 2.2, was labeled Removed. Those two labels are not equivalent and the page distinguishes them.

The clock also runs upstream, on a page most platform users never see. OpenAI’s deprecations page, dated July 20, 2026, listed December 1, 2026 shutdown dates for the model identifiers gpt-image-1-mini, gpt-image-1.5, and chatgpt-image-latest. Google’s Gemini pricing page carried at least one row designated as not available and shut down, which FSR did not resolve to a specific model.

Set that beside the platform layer. Adobe’s captured roster listed a partner model displayed as “GPT Image 1.5.” Whether that display name and OpenAI’s model identifier refer to the same artifact is not established by either page, and this briefing does not assert that they do. What is established is that a buyer reading only the platform page would not encounter the December date at all.

The same roster showed the gap in a second place. It listed “Veo” by Google with no version. Its credit table listed Google Veo 3.1 and Veo 3.1 Fast. Its retirement table listed Google Veo 2.0 and Veo 3, Standard and Fast, as deprecated. Three Adobe pages, three levels of version specificity, one model line.

The practical consequence is that a display name is not a version, and a workflow built on a display name has no anchor. If reproducibility matters to your brand, capture the model identifier, the parameters, and the output at delivery time, because the platform may not preserve them for you.

Sources: Adobe, Generative AI model retirement, captured 17 August 2026 · Adobe, Partner models in Adobe products, captured 17 August 2026 · OpenAI, Deprecations, dated 20 July 2026 · Google, Gemini Developer API pricing, opened 17 August 2026

Eight checks before you pay

What FSR could not establish

This section exists because a briefing that only reports findings is hiding its own shape. These are the gaps that would change a purchase decision, listed so that you can close them before you sign.

FAQ

Methodology and change log

Evidence cutoff: August 17, 2026. Article published: August 18, 2026.

This is a Tier C briefing. Future Stack Reviews did not open a paid account, generate an image, run a checkout, or measure output for this article. Every claim is sourced to a vendor page, contract document, license file, or model card, with the date FSR opened or captured it.

Source hierarchy applied. Vendor contract documents and product descriptions first. Official API and pricing documentation second. Model cards and license files third. Everything else was treated as a lead and is not cited here.

Documents opened. Adobe: PSLT Creative Partner Model Supplemental Coverage 2026v1, Creative Partner Models Product Description effective May 8 2026, partner models roster, Firefly IP Indemnification, generative AI model retirement, generative credits FAQ. Black Forest Labs: API pricing documentation, public pricing page, API reference, open weights licensing page, FLUX.2 klein 4B model card. Midjourney: plan comparison, GPU speed, Stealth Mode, privacy, commercial use, terms of service. OpenAI: API pricing, chatgpt-image-latest model page, deprecations, consumer plan page, free tier FAQ. Google: Gemini Developer API pricing, Google AI plan benefits. Model repositories for Qwen-Image, HiDream-I1, and HunyuanImage 3.0. ByteDance Seedream and Seedance product pages.

Documents not obtained. Adobe’s General Terms and the Generative AI Features Terms, both of which sit in the same contract stack and were located but not read for this briefing. Any customer Sales Order. Black Forest Labs’ login-gated commercial order forms. Adobe checkout terms for each plan.

Recorded as not found in the sources searched. An official Midjourney API, searched on midjourney.com with four query variants. A stated failed-generation billing policy, searched across Adobe’s credit documentation and Black Forest Labs’ documentation. Prompt retention periods, searched on each of five vendors’ own domains. These record what a search returned, not that no policy exists.

Contradictions preserved rather than resolved. Adobe’s statement that partner models are not developed by Adobe against its statement that partner models are considered part of Adobe’s products and services. Adobe’s answer that there is no additional cost to using partner models against the per-model credit consumption table on the same page. Both pairs appeared on the version FSR captured.

Volatile fields. Prices, credit rates, plan allowances, model rosters, and retirement tables change without notice, and several Adobe credit rates carried a “for a limited time” footnote with no stated end date. Treat every figure here as the value the page displayed on the recorded date, and open the linked source before acting on it.

Not claimed. No image quality ranking, benchmark, or measured speed. No assertion that platform display names and provider model identifiers refer to the same artifacts. No legal conclusion about any clause. This briefing is not legal advice.

Change log. Version 1, August 18, 2026: first publication. Future corrections will be numbered here with the date, the claim affected, and what triggered the change.

Verdict by route

No route wins every layer of the purchase. Pick the one whose unresolved field your team can actually manage.

Adobe Firefly. Shortlist it for governed design workflows where admin control and a published output indemnity matter more than endpoint-level control. Before you commit, confirm your SKU, that your specific model sits in the eligible subset rather than merely in the product, and the liability cap in the General Terms. Unresolved: what “unmodified” covers, and what your Sales Order changes.

Black Forest Labs direct. Choose it when automation, usage logs, endpoint identity, and a computable unit cost matter more than a finished creative application. Build the parameter table by endpoint. Unresolved: failed-request billing, and credit purchases are non-refundable.

Local Apache-licensed weights. Choose it when infrastructure control justifies the operational overhead of hardware, security, moderation, provenance, and updates. Unresolved: whether every layer of your pipeline is actually local.

Paid commercial self-hosting. Choose it when fine-tuning or embedded deployment is required, and buy the tier that matches your client and domain structure rather than the cheapest one. Unresolved: whether your tier permits client work, and what removal or deletion provisions your agreement carries.

Midjourney. Choose it after your visual workflow clears testing and you accept public-by-default behavior and the absence of a located API. Unresolved: how GPU time converts to your actual output volume.

Bundled assistant plan. Use it first for low-volume work already covered by an existing subscription, and do not treat it as a predictable production meter. Unresolved: the per-image entitlement, which the consumer pages do not publish.

If one thing survives from this briefing, make it the second check. The name on your dashboard and the identifier on the provider’s deprecations page are maintained by different teams on different pages, and only one of them carries a date.