AI Image Generators Compared: Models, Prices, Licences and Provenance

Choosing an AI image generator now means comparing model versions, prices, licences and provenance signals that change month by month, and most comparison pages online mix those facts with opinion. This page sets out documented facts for five widely used image generator families: FLUX from Black Forest Labs, GPT Image from OpenAI, Midjourney from Midjourney, Nano Banana (Gemini image models) from Google, and Stable Diffusion 3.5 from Stability AI. Products are listed alphabetically by product name, and the order is not a ranking. Every field comes from the provider's own documentation, pricing page or model card; where a provider does not publish a fact, the table says so. It is written for designers, marketing and content teams, developers, procurement and the security, legal and privacy teams who approve these tools.

Last checked: 2026-11-15.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.

Key takeaways

• Two families publish open weights. Black Forest Labs publishes open weights for FLUX 3 Image Klein and FLUX.2 [klein], and Stability AI publishes Stable Diffusion 3.5 weights under its Community License.
• Licence thresholds differ. The FLUX Kommunity License allows free commercial use under USD 5 million annual revenue, the Stability AI Community License under USD 1 million, and Midjourney requires Pro or Mega for companies above USD 1 million gross revenue.
• Pricing models are not alike. GPT Image and Nano Banana bill API use per token, FLUX and Stability bill per image in credits, and Midjourney sells monthly subscriptions with GPU time.
• Provenance signals vary. Google states that every Gemini API image carries a SynthID watermark; OpenAI added C2PA metadata to DALL-E 3 images in 2024, and Google reports that OpenAI is bringing SynthID to more of its AI content; the other providers' documentation does not state a signal.
• Model churn is high. OpenAI shut down DALL-E on 12 May 2026 and gpt-image-1 on 23 October 2026, and lists further GPT Image shutdowns on 1 December 2026 and 10 February 2027.

Which AI image generators does this page cover?

The five families below were chosen because each is a current, actively documented product with a public pricing or licence page. Each has its own access route: some are APIs for developers, some are apps for creators, and two can be self-hosted.

For model families beyond images, see the AI Image, Video, Voice and Music Models Hub.

How much do AI image generators cost?

Token-billed and credit-billed prices do not convert directly: a token-billed image costs more at higher resolution and with more input images, and Google bills thinking tokens on Gemini 3 image models. Run a sample workload on each candidate before comparing totals.

FLUX (Black Forest Labs)

Black Forest Labs describes FLUX 3 as its latest model family, "one model, one API", and says FLUX.2 "remains fully supported for production image generation and editing". FLUX 3 Image supports text to image, instruction-based editing, up to ten reference images referred to by position, and bounding boxes on a 0 to 1000 grid to place elements or limit edits to part of an image. Web search grounding is on by default and can be switched off with grounding set to false.

• Open weights: FLUX 3 Image Klein, released on 15 October 2026, is a 5.6B-parameter open-weights model with a base model for fine-tuning and a step-distilled model, hosted on Hugging Face and also served through the API. FLUX.2 [klein] is published at 4B and 9B.
• Licence: the FLUX Kommunity License allows free commercial use for individuals and companies with annual revenue under USD 5 million, plus research and non-commercial use; larger organisations use self-hosted licensing or the API.
• Safety: the Klein release post lists dataset filtering, post-training fine-tuning, red teaming and external evaluation.

GPT Image (OpenAI)

OpenAI's image generation guide recommends gpt-image-2.5-sunburst where editing precision matters most and gpt-image-2.5-flare for fast everyday generation. The 2.6 models add complexity control (low, medium or high, default high, with no separate fee), markup-guided edits, SVG output and animated GIF output through the Image API.

• Sizes: recommended sizes are 1024x1024, 1536x1024 and 1024x1536; custom sizes must be multiples of 16, with aspect ratios between 1:3 and 3:1 and no edge above 3840 pixels.
• Controls: a moderation parameter takes auto (default) or low; blocked requests return moderation_blocked. Transparent backgrounds are available with PNG or WebP output.
• Access: API Organization Verification may be required before using GPT Image models.
• Limitations stated: complex prompts can take up to two minutes, and precise text placement, recurring character consistency and layout-sensitive compositions can be hard to control.

Midjourney

Midjourney is sold as a subscription with four plans. Each plan includes monthly Fast GPU time: Basic 3.3 hours, Standard 15 hours, Pro 30 hours and Mega 60 hours. Relax mode gives unlimited images on Standard and above, and unlimited standard-definition video on Pro and Mega.

• Privacy of outputs: Stealth mode, which keeps images and videos private, is available on Pro and Mega.
• Commercial terms: all subscribers receive Midjourney's General Commercial Terms; companies with gross revenue above USD 1,000,000 a year must subscribe to Pro or Mega.
• Billing: subscriptions renew automatically unless cancelled; annual plans are billed upfront at a 20% discount.

Nano Banana (Google)

Nano Banana is the name Google uses for its Gemini image models. The Gemini API documentation lists Nano Banana 2.1 (gemini-nano-banana-2.1) as the primary model, Nano Banana 2 Lite (gemini-3.1-flash-lite-image) as the fastest and lowest cost, Nano Banana 2 (gemini-3.1-flash-image) as the previous generation, Nano Banana Pro (gemini-3-pro-image) for complex tasks, and the original Nano Banana (gemini-2.5-flash-image) as legacy.

• Resolution: 1K, 2K and 4K, with 1K as the default for Gemini 3 image models; Nano Banana 2 Lite supports 1K only.
• References: Nano Banana Pro accepts up to 6 object, 5 character and 3 style references; Nano Banana 2.1 and 2 accept up to 10 object and 4 character references.
• Provenance: all generated images include a SynthID watermark. Google added C2PA Content Credentials verification to the Gemini app from 20 May 2026.
• Apps: Google Pics, built on the latest Nano Banana model, opened to trusted testers on 20 May 2026 with a rollout to Google AI Pro and Ultra subscribers planned for summer 2026.

Stable Diffusion 3.5 (Stability AI)

Stability AI lists three SD 3.5 variants: Large, for professional use at 1 megapixel; Turbo, which generates images in four steps; and Medium, designed to run on consumer hardware. The SD 3.5 Large model card describes an 8B-parameter Multimodal Diffusion Transformer (MMDiT) with three pretrained text encoders.

• Licence: the Stability AI Community License is free for research, non-commercial and commercial use for organisations or individuals under USD 1 million total annual revenue; above that, Stability AI directs users to an Enterprise License. Commercial use also requires registration.
• Output ownership: as between the user and Stability AI, the user owns outputs to the extent permitted by applicable law, and is responsible for the lawfulness of using them.
• Intended use: the model card states the model was not trained to produce factual or true representations of people or events.

How do the licences and commercial terms compare?

None of this is legal advice. Copyright and likeness law varies by country, and training-data questions sit outside these licences. See the AI Training Data Provenance and Copyright Exposure Guide.

Which GPT Image and DALL-E models are retiring?

Google's documentation also marks the original Nano Banana (gemini-2.5-flash-image) as legacy and describes Imagen as shut down. OpenAI's video product followed a different path; see OpenAI Sora Shutdown Explained.

How should you choose an AI image generator?

These are factual criteria, not a verdict. Weigh them against your own use case.

• Deployment: if images must be generated on your own infrastructure, only the open-weights families (FLUX Klein models, Stable Diffusion 3.5) can be self-hosted; the others are services.
• Revenue and licence: check your organisation's annual revenue against the USD 1 million and USD 5 million thresholds above before commercial use.
• Provenance obligations: where you must label AI content, for example under Article 50 of the EU AI Act, prefer documented signals and test whether they survive your publishing pipeline. See the EU AI Act Article 50 Transparency Obligations Guide.
• Lifecycle: API model names change quickly; build model IDs into configuration, not code, and track each provider's deprecation page.
• Cost shape: subscription, per-image and per-token billing suit different volumes; model your expected monthly image count and resolution.
• Controls: moderation settings, privacy of outputs (such as Stealth mode) and enterprise data handling differ, so map them to your policy.

Where identity and security fit

Every API-based image generator is reached with a key or service account, and those credentials are non-human identities to inventory, scope and rotate. See the OpenAI API Keys Security Guide and the Google Vertex AI Security Guide. Self-hosted open weights move the security work in-house: model files come from Hugging Face, so verify sources and protect access tokens (see the Hugging Face Security Guide). Image tools are also a deepfake risk, so pair approval with a usage policy and staff awareness; see AI Image and Video Generation at Work and the Spotting Deepfake Scams Guide.

Practitioner checklist

• Record the exact model ID, provider and deprecation date for every image model in use.
• Check annual revenue against each licence threshold before commercial use of open weights or subscriptions.
• Test whether SynthID or C2PA signals survive your editing, compression and publishing steps.
• Store API keys in a secrets manager, scope them per project and rotate them on a schedule.
• Set moderation and privacy options (such as moderation level or Stealth mode) to match your AI use policy.
• Run a cost test with your real resolution and reference-image mix before committing to a billing model.
• Review OpenAI's deprecations page and Google's Gemini API image page monthly for retirements.

Standards and references

• Black Forest Labs: FLUX 3 Image Klein documentation
• Black Forest Labs: FLUX 3 Image Klein release
• Google: Gemini API image generation (Nano Banana)
• Google: Gemini Developer API pricing
• Midjourney: Plans
• OpenAI: Image generation guide
• OpenAI: Deprecations
• Stability AI: Stable Diffusion 3.5 Large model card
• Stability AI: Community License Agreement
• Stability AI: API pricing

Related NHI Mgmt Group resources: AI Image, Video, Voice and Music Models Hub · AI Video Generators Compared · AI Content Provenance and C2PA Guide · SynthID vs C2PA Content Credentials