For ecommerce teams that turn private product photos into catalog images at volume, the best enterprise AI image API is Nightjar. Nightjar does not train or fine-tune AI models on your uploads or outputs on any plan, takes a license only to provide and improve the service for you, and gives you full ownership of every image, including for commercial use. It keeps your Products, looks and images in one Library, so every run builds on the last. Its API is built for production runs: a retry sent with the same key returns the first answer instead of running or charging again.

Teams on Google Cloud, Azure or AWS that would rather build the product-photography layer themselves can use those clouds' image models under the data terms they already have. Adobe Firefly Services fits companies already on Adobe enterprise contracts that want a model trained on their own catalog.

We read each vendor's current terms, privacy pages and API documentation in 2026 and quote them below. Vendors change these documents, and this is not legal advice: have counsel read the version you sign.

How fourteen AI image APIs handle private product photos

The table compares each API's default on self-serve or standard paid use. Enterprise contracts can change the answer, and the sections below say where.

Does an AI image API train on your product photos?

The answer depends on the vendor, the plan and which document you read. Nightjar, OpenAI, Azure OpenAI, Google's Agent Platform and Amazon Bedrock say no for every paying customer, and Bria says no unless you consent in writing. For the rest, check the plan and the setting.

Nightjar does not train or fine-tune AI models on your uploads or the images it makes. The commitment is the same on every plan and on every surface: the web app, the Shopify app, the API and AI assistants. There is no setting to find. The Nightjar Privacy Policy states it: "We do not use them to train or fine-tune AI models."

OpenAI says on its data controls page that API data is not used to train its models "unless you explicitly opt in to share data with us". Azure OpenAI says prompts and outputs "are NOT used to train any generative AI foundation models without your permission or instruction" (Microsoft Learn). Google's Agent Platform carries a training restriction in section 18 of its Service Specific Terms. Amazon Bedrock stores no inputs or outputs from Stability models by default (AWS abuse detection), and model providers "don't have access to Amazon Bedrock logs or to customer prompts and completions" (AWS data protection).

The answer changes by plan, setting or model at these vendors:

• Gemini Developer API. On the unpaid tier, including Google AI Studio without a billing account, Google uses your content to "improve, and develop Google products", and "human reviewers may read, annotate, and process your API input and output." Google adds: "Do not submit sensitive, confidential, or personal information to the Unpaid Services." The paid tier is not used for improvement (Gemini API Additional Terms).
• Photoroom API. Photoroom's own pages disagree. Its Terms (section 2.2.c) "expressly authorize Photoroom to use User Content to improve, train and develop Photoroom's products", with an opt-out "in the app under Data Control". Its security page says "Self-serve API plans: yes by default, with a customer-controlled opt-out in account settings." Its help article says API images are not used for training. Enterprise contracts train only with consent.
• Stability AI. The Terms let Stability use content to improve its services, and "you can opt-out to prevent us from using your Inputs and Outputs to train our models." The opt-out is a toggle named "Training: Improve the Model for Everyone" in the account settings.
• Black Forest Labs. The FLUX API Service Terms say "the Company may use Inputs and Outputs to train and improve its artificial intelligence models." The opt-out is an email to the privacy team (privacy policy). BFL's FLUX 3 training summary says API data "may have been used to improve the quality and capabilities of the FLUX 3 model."
• Claid. Its pricing page says Claid will "never use customer data for model training." Its Terms never use the word "train", but their license covers use "solely in de-identified form, to improve our products and services ... developing new products or features." The master service agreement that governs API use is not public.
• Adobe Firefly Services. Adobe's Firefly Services security fact sheet says enterprise inputs and outputs are not in "datasets used to train Firefly foundation models." The General Terms (section 6.2) still permit use of customer data "for product improvement and development." Ask for the no-training promise in your sales order.
• Bria. Bria warrants that customer content "shall not be used to train generative models, unless Customer expressly consents to such training in writing." Its enterprise terms count buying a fine-tuning product as that consent (Bria legal).
• fal. Its API Services Terms say fal "will not use Client Content to create, train, develop (directly or indirectly)" its products, except on models marked "Pending Enterprise Ready", where its data processing terms don't apply either. Its general Terms also let fal use de-identified data derived from your inputs to develop its products and AI models.
• Replicate. Its Terms give Replicate no right to train its own models on your data. The license covers training only your own fine-tuned models, and Replicate may use aggregated, anonymized usage data to improve its services. Models from third-party providers run under those providers' terms.

A no-training promise is only as strong as the document that binds you. A help article or a blog post is not the contract.

What license does an image API take to your uploads?

Every image API needs a license to process your photos. What matters is how long it lasts, whether the vendor can pass it on, and what it covers besides running your request.

Nightjar's Terms (section 3(b)) grant Nightjar a "limited, non-exclusive, worldwide, royalty-free license" to your input and output, "solely for the purpose of providing, maintaining, and improving the Services for your benefit." It is not perpetual, irrevocable or sublicensable. OpenAI's Services Agreement is similarly narrow, granting OpenAI "only the limited rights that are reasonably necessary" to deliver the service, and fal's license is non-sublicensable and limited "to provide the Services". Replicate's license includes "publish" and "display", but only "to the extent necessary" to deliver outputs, train your own fine-tunes and run the service. On Google's Agent Platform, Azure and Bedrock, the cloud provider uses your content only to process it on your instructions.

Three vendors take licenses a legal team will read twice:

• Photoroom takes "a royalty-free, sublicensable, transferable, perpetual, irrevocable" license to "publish ... publicly display, and make derivative works of" your content. It is limited to "our provision of the Services", and the training sentence follows it in the same section.
• Claid takes "an irrevocable, perpetual, transferable, sublicensable (through multiple tiers)" license. Its purposes include de-identified use for "developing new products or features", and Claid owns that de-identified data.
• Black Forest Labs takes a "perpetual, irrevocable, worldwide, non-exclusive, and fully sublicensable" license to "publicly perform, and publicly display" your inputs and outputs, "improving the Company's products and services, and developing new products and services."

A perpetual, sublicensable license outlives your contract. For photos of a product that has not launched, that is the clause to negotiate first.

Who owns AI product images made through an API?

You own the images at almost every vendor, but the wording differs. Nightjar's Terms grant "full ownership rights to the Output generated specifically for you", including for commercial purposes. Microsoft's Product Terms say Microsoft "does not own Customer's Output Content". OpenAI, Bria and Stability assign their rights in outputs to you, and Replicate grants you "all right, title and interest, if any, in and to Output", subject to each model's own terms. Black Forest Labs says "you own all right, title, and interest in and to Output."

Four gaps are worth knowing. Google "won't claim ownership" of Gemini API output but assigns nothing. Photoroom's general Terms cover what you upload and never mention generated outputs; its Enterprise Services Agreement says "Customer ... owns the Outputs." Claid's public Terms do not address outputs at all. fal's Terms assign nothing and say fal gets no rights in outputs, and commercial use depends on each model's license; some are for research only (fal FAQ).

Most vendors, Nightjar included, also say outputs may not be unique. Ownership between you and the vendor is a separate question from whether copyright protects an AI image, which the help answer on who owns the copyright of an AI product photo covers.

How long do image APIs keep your product photos?

Retention runs from nothing at all to the life of your account.

The two ends of the table suit different jobs. A stateless API returns the image and forgets it, so your team builds the storage, the product records and the history around it. Nightjar keeps the work so it can be reused: a saved Product, its photos, the looks your Team saved and every image made from them, searchable in one Library. To delete the account, you ask support, and Nightjar deletes your images, uploads and saved Products within 30 days, as its Privacy Policy sets out.

How do AI image APIs handle retries, failures and volume?

At catalog volume, what matters is whether a retried job runs and bills twice, whether failed images cost anything and how much work the API runs at once. Networks drop requests and workers restart, so a job that runs twice is billed twice unless the API prevents it.

Retries. Every Nightjar API request that starts work carries an Idempotency-Key, and the API documentation requires it. Sending the same key again returns the original answer, so a retry doesn't start the work or charge a second time. Photoroom's billing page says repeated calls bill twice: "Calling the API twice with the same input image and the same edit options will be billed as 2 separate calls."

Failures. A Nightjar request that delivers no image costs nothing. Nightjar automatically looks at every generated image and retries the bad ones at no cost to you. Photoroom ("Calls that result in an error will not consume an image"), Stability ("You will not be charged for failed generations") and fal ("You pay only for successful outputs") also skip charges for errors.

API keys. A Team owner creates Nightjar API keys as read-only or full access, with an optional expiry, and can rotate or revoke them. Generation history is shared across the Team, and each job keeps a record of the key that started it, even after that key is revoked.

Throughput and price. OpenAI and Google's Gemini API offer batch processing at 50% off with a turnaround of up to 24 hours, which suits runs that can wait. Adobe's Firefly API starts at 4 requests per minute until your account manager raises it (usage notes), while Nightjar's API reports the Team's concurrency limit, so a pipeline can run as many jobs at once as the Team allows. Photoroom's Enterprise plan requires an annual commitment of at least 200,000 images, and Black Forest Labs' enterprise agreements start at 200,000 generations a month. Nightjar's API comes with every paid plan at no extra cost, from $25 a month, at one Credit per image (two at 4K), with custom plans beyond the largest listed plan. That works out at about $0.17 an image on the $25 plan and $0.125 on larger plans.

Batches through AI assistants. Teams that run catalog work by conversation can drive the same Team from Claude, Claude Code, Codex, or ChatGPT Business, Enterprise and Edu workspaces. One batch covers up to 100 items across Products. Planning a batch is free and shows the total cost and a spending ceiling. You approve samples first, then the rest runs and keeps running after the chat closes. Every paid request carries the Credit ceiling you approved.

A raw model or Nightjar. The comparison of AI image generation APIs for ecommerce lists what the raw model APIs charge per image. A raw API gives you one model for every shot. Every model is best at some shots and weaker at others, so a system that sends each shot to the model that is best at it gives better results than any one model used on its own. Nightjar has a sophisticated routing system that always picks the best AI image model for the task at hand, with prompts tuned for each model. Nightjar is also built to keep your product true to the original. Every image starts from a saved Product with several photos, a description and real dimensions, and Nightjar automatically looks at every generated image and retries the ones that misrepresent the product, at no cost to you. Your Products and looks stay saved for the next run, and your team doesn't have to build that pipeline.

Should you train a custom AI model on your product catalog?

Train a custom model only if you want a model of your own and can accept the contract terms that come with it. Saved product records and looks keep a catalog consistent without one.

Several enterprise image APIs keep a brand consistent by training a model on your images. Adobe's Firefly Foundry builds private models "trained on entire catalogs of existing IP", and Firefly Custom Models fine-tune on at least 10 of your images. Photoroom Enterprise offers "Custom AI models tuned to your catalog", and Bria and Black Forest Labs offer fine-tuning on brand assets.

A trained model brings contract terms of its own. Adobe Custom Models are for internal use only, Adobe "cannot provide Customer with a copy of any Custom Model", and you must retrain when the underlying Firefly model is retired (Custom Model terms). Under Black Forest Labs' developer terms, BFL owns fine-tuned versions of its models. At Bria, buying a fine-tuning product counts as consent to train on your content. The help answer on the IP risks of training a custom AI model on your products goes further.

Nightjar keeps a catalog consistent without training anything. Each product is saved once as a Product, with several photos, a factual description and real dimensions, and every new image starts from it. The look lives in saved Photography Styles (made from three of your own photos), Fashion Models, Backgrounds and Poses that belong to your Team. The API creates and uses all of them directly, with nothing to set up in the web app first. When Nightjar moves to a newer AI image model, your Products and looks carry over with nothing to retrain.

The best enterprise AI image API for each kind of team

Nightjar: best for ecommerce teams turning private product photos into on-brand catalog images at volume

Nightjar makes new product photography from saved Products: listing and lifestyle images, on-model shots, Photoshoots, edits and upscales, through the same Team and Library as the web app, the Shopify app and AI assistants. It always picks the best AI image model for the task at hand, does not train or fine-tune AI models on your content, takes a narrow license, gives you full ownership, and settles retries and failures in your favor. The API is on every paid plan with no minimum commitment. Try Nightjar free in the web app with 6 Credits and no card, then add the API with any paid plan (API documentation).

Google Gemini Enterprise Agent Platform: best for Google Cloud teams building their own product-photo layer

Agent Platform gives you Nano Banana image models under a no-training restriction, and you build the product records, review and retries yourself. Test with private photos only on a billed account, because in Google AI Studio without billing, Google uses content to improve its products.

Azure OpenAI in Foundry: best for Microsoft shops

Azure OpenAI offers GPT Image models under Microsoft's data terms, inside the Azure account you already have. Its image models offer no batch discount.

OpenAI API: best for developers building their own product-photo layer on GPT Image

OpenAI does not train by default and assigns outputs to you. Zero data retention needs OpenAI's approval and excludes the discounted Batch API.

Adobe Firefly Services: best for enterprises on Adobe contracts that want a model trained on their catalog

Firefly models are trained on licensed content such as Adobe Stock. Prices come through sales, and the no-training promise belongs in your sales order.

Amazon Bedrock: best for AWS teams that need Stability's editing services with zero retention

Bedrock stores nothing by default, and you build the product records and retries yourself.

Bria: best for teams that want models trained on licensed data

Bria does not train on your content unless you agree in writing, and buying one of its fine-tuning products counts as agreeing. It keeps results for 3 days.

Photoroom API: best for high-volume listing cleanup at a per-call price

Background removal costs $0.02 and an editing call $0.10 (pricing). On a self-serve plan, switch off training in your account.

Claid API: best for automated enhancement pipelines tied to your own cloud storage

Claid's pricing page promises no training, but its Terms keep a perpetual, sublicensable license, and the agreement that governs API use is available only on request.

Replicate: best for trying many models with short retention

Replicate deletes API predictions, their files and logs after an hour by default, so copy outputs into your own storage within the hour. Rights in outputs follow each model's own license.

fal: best for fast access to many image models

fal keeps request inputs and outputs for 30 days unless you send its X-Fal-Store-IO: 0 header, and you set how long its generated files last. Check each model's label, because models marked "Pending Enterprise Ready" fall outside fal's no-training promise.

Stability AI and Black Forest Labs: best for direct model access

Switch off training first: Stability's toggle, or BFL's opt-out email. BFL's zero data retention is part of its enterprise offer.

How to check an image API before you send it private product photos

1. Find the document that binds you. It may be the online terms, an enterprise agreement, a master service agreement or a sales order. Claid's controlling agreement is not public, and Photoroom's pages give different training answers.
2. Read the training clause for your plan. Note where the opt-out lives and whether it is in the contract or only in a help article.
3. Read the license words. "Perpetual", "irrevocable", "sublicensable" and "publicly display" are the ones to negotiate.
4. Check output ownership. "Assigns to you" is stronger than "claims no ownership."
5. Ask what is kept and for how long. That covers your uploads, the generated images and the request logs.
6. Test retries and failures. Send the same request twice and confirm whether you pay twice.
7. Start the pilot with a product that is already public. Move to unreleased products once the answers above are settled.

Frequently asked questions

Does Nightjar train AI models on my product photos? No. Nightjar does not train or fine-tune AI models on your uploads or outputs, on any plan and through every surface, including the API. You own what you make, including for commercial use.

Is the free Gemini API safe for private product photos? Google's image models aren't on the Gemini API's free tier, so the risk is testing them in Google AI Studio without a billing account. Google counts that as unpaid use: it uses the content to improve its products, human reviewers may read it, and its terms ask users not to submit "sensitive, confidential, or personal information". Paid use is not used for improvement, and users in the EEA, Switzerland and the UK get the paid-tier terms everywhere.

Do I need an enterprise contract to keep my images out of AI training? Not at Nightjar, OpenAI, Azure OpenAI, Google's Agent Platform, the paid Gemini API, Amazon Bedrock or Replicate, where no training is the default for paying customers. fal makes the same promise except on models marked "Pending Enterprise Ready", and Bria trains only with your written consent. On Photoroom's self-serve plans, Stability AI and Black Forest Labs you opt out yourself, and BFL's zero data retention is part of its enterprise offer. At Adobe, put the no-training promise in your sales order.

Can I use images made through an AI image API commercially? At most vendors compared here, yes on paid use. Claid's public Terms don't address outputs, so get commercial rights in writing, and at fal and Replicate they depend on each model's license. Photoroom's free plan is limited to personal, non-commercial use, so check the plan as well as the vendor.

References

• Nightjar Terms of Service: license and ownership
• Nightjar Privacy Policy: no training and retention
• Nightjar API documentation: Idempotency-Key, Assets, Generations
• OpenAI data controls and OpenAI Services Agreement: training, retention, zero data retention, license, ownership, audit reports
• Gemini API Additional Terms and Gemini API abuse monitoring: unpaid and paid tiers, human review, 55-day logs
• Google APIs Terms of Service: the license for content submitted to the APIs
• Google Cloud Service Specific Terms and Agent Platform abuse monitoring
• Azure OpenAI data, privacy and security and Microsoft Product Terms: training, retention, ownership
• Amazon Bedrock abuse detection and data protection
• Adobe Firefly Services security fact sheet, Adobe General Terms, Custom Model terms, Firefly Foundry and Firefly API usage notes
• Bria legal agreements, asset retention, Tailored Generation and pricing
• Photoroom Terms, Enterprise Services Agreement, security page, help article on training, API billing, API pricing and Enterprise page
• Claid Terms, API pricing and image I/O docs
• Stability AI Terms of Service, privacy center and API reference
• Replicate Terms of Service, data retention and billing
• fal Terms of Service, API Services Terms, data retention, FAQ and pricing
• Black Forest Labs FLUX API Service Terms, developer terms, privacy policy, transparency hub, enterprise page and integration guidelines