Top 10 AI Image Generators (2026)

• GPT Image 2 leads the Artificial Analysis Text-to-Image Arena at Elo 1178 across 14,585 blind comparisons in September 2026, ahead of Microsoft’s MAI-Image-2.6 at 1149.
• Nano Banana models had generated more than 50 billion images by I/O 2026, up from 5 billion in October 2025, per Sundar Pichai’s keynote.
• Hosted API cost per image spans roughly $0.006 (GPT Image 2, low, estimate) to $0.24 (Nano Banana Pro at 4K, estimate) in 2026, a forty-fold spread.
• Ideogram 4.0, released June 3, 2026, posts a 0.97 OCR-accuracy score and ranks first among open-weight models on DesignArena, per Ideogram’s technical writeup.
• Microsoft’s MAI-Image-2, launched March 19, 2026, is free with a Microsoft Account and tops the Artificial Analysis Image Editing Arena as of September 2026.

How We Selected the Top 10 AI Image Generators (2026)

We selected the top 10 AI image generators of 2026 by ranking on independent blind-vote arenas first, then checking price, openness and provenance, and dropping any model its vendor has retired. That last filter mattered this year. DALL·E 3 left OpenAI’s API on May 12, 2026, and Adobe removed Firefly Image 3 from every plan in August 2026, so neither gets a slot no matter how familiar the names are.

The primary signal is human preference at scale. The Artificial Analysis Text-to-Image Arena runs blind head-to-head votes and publishes an Elo score per model, and its sister Image Editing Arena does the same for edits. For design-specific work (posters, type, layout) we also weight the DesignArena designer-preference leaderboard. Academic benchmarks stay in as a sanity check: T2I-CompBench++ for attribute binding, spatial relations and counting, and GenEval for object composition. They move slower than the products do, so they rank second.

Four practical axes then decide the order:

• Price per image: API cost where a hosted endpoint exists, subscription cost where it doesn’t.
• Openness: whether the weights are downloadable and under what license, because a model you can run on your own GPU answers a different question than one behind an API.
• Provenance: whether outputs carry C2PA Content Credentials, a SynthID watermark, or both. Since May 19, 2026, when OpenAI joined the C2PA steering committee and committed to embedding SynthID alongside Content Credentials, the two largest vendors use the same pair of mechanisms.
• Availability: a model you can’t buy or download today (early access, waitlist, partner-only) doesn’t get a numbered slot.

One honest caveat. Arena Elo measures which image people prefer in a blind pair, and it rewards attractive output as much as correct output. A model can win the arena and still miscount the chairs in your product shot. That’s why each numbered entry names the specific job the tool is best at instead of pretending one score settles everything.

This is the same filter AlphaCorp AI applies when scoping a generative feature for a client: shipped, priced, licensed, traceable. Hype fails all four.

Comparison Table: Top AI Image Generators (September 2026)

The comparison table lists the ten current picks with vendor, best use, Artificial Analysis Elo, resolution, open-weight status and price as of September 10, 2026. Three reading notes. Elo figures come from the Artificial Analysis Text-to-Image Arena in September 2026, except Ideogram’s, which is a DesignArena score on a different scale. API costs for GPT Image 2 and Nano Banana Pro are third-party per-token estimates rather than list prices. “Not listed” means the vendor doesn’t publish the figure.

Two models sit near the top of the arena without appearing here. Reve 2.1 holds third place at Elo 1127 in September 2026, and the wider MAI-Image family places several variants in the top ten. The ranking favors one slot per vendor and per job, so the table shows the strongest version of each.

1. OpenAI GPT Image 2 (ChatGPT Images 2.0) – Best Overall for Text Accuracy

OpenAI’s GPT Image 2, shipped as ChatGPT Images 2.0 on April 21, 2026, is the best overall AI image generator of 2026 and the most accurate at rendering text, leading the Artificial Analysis arena by a clear margin. It replaced gpt-image-1 and GPT-4o image generation outright.

The headline change is that the model thinks before it draws. Before generating, it plans the composition, verifies object counts and checks the prompt’s constraints, using the same reasoning approach OpenAI applies in its O-series models, per OpenAI’s ChatGPT Images 2.0 announcement. Add web search grounding and a prompt for an infographic can pull real figures instead of plausible ones.

What it does that gpt-image-1 didn’t:

• Generates up to eight consistent images from a single prompt, so a storyboard or product-variant set becomes one request.
• Outputs at native 2K resolution.
• Renders non-Latin scripts far more reliably, which matters for anyone shipping signage or packaging outside English markets.
• Splits into two modes: Instant, free to every ChatGPT user, and Thinking (web search, layout reasoning, multi-image batches) for Plus, Pro, Business and Enterprise.

What most people don’t notice until they try it: the count-verification step is what makes the eight-image batches usable. The images agree with each other on how many objects are in frame, which earlier models routinely got wrong across a set.

GPT Image 2 (high) leads the Artificial Analysis Text-to-Image Arena at Elo 1178 across 14,585 blind comparisons as of September 2026, ahead of Microsoft’s MAI-Image-2.6 at 1149, Reve 2.1 at 1127 and Google’s Nano Banana 2 at 1121.

At launch the gap was wider still, with reporting in April 2026 putting the lead at roughly 242 Elo points before rivals caught up.

Now the bill. The API charges per token: $5 per million text input tokens, $8 per million image input tokens, $2 per million cached image tokens and $30 per million image output tokens. Third-party calculators translate that to about $0.006 (low), $0.05 (medium) and $0.21 (high) per 1024×1024 image. Compare gpt-image-1’s flat rate of roughly $0.035 and the high tier costs about six times more. Reasoning compute isn’t free, and anyone who budgeted on the old flat rate should rerun the numbers with the quality tier picked per job.

The migration clock is also running. OpenAI’s deprecations page schedules gpt-image-1 for shutdown on October 23, 2026, with gpt-image-1.5 and gpt-image-1-mini following on December 1, 2026. The DALL·E 2 and DALL·E 3 API endpoints already closed on May 12, 2026. If a pipeline still calls any of them, you have weeks.

2. Google Nano Banana Pro / Nano Banana 2 – Fastest for Photorealistic Results

Google’s Nano Banana family, the Gemini Image line that replaced Imagen, is the fastest route to photorealistic output in 2026 and the only pick on this list that generates natively at 4K. Imagen is gone as a product: Google’s Gemini API documentation set the shutdown date for August 17, 2026, and pointed developers to Nano Banana instead.

The name sounds like a joke. The models don’t. As of September 2026 there are three tiers, per Google DeepMind’s Gemini Image model page:

• Nano Banana Pro (Gemini 3 Pro Image): the quality tier. Native output up to 4K, up to 14 reference images per prompt, consistent rendering of up to five people across a set, and Google Search grounding so an infographic can carry real numbers.
• Nano Banana 2 (Gemini 3.1 Flash Image): Google’s description is “pro-level generation and editing at Flash speed.” It has been the default in the Gemini app since March 2026.
• Nano Banana 2 Lite (Gemini 3.1 Flash-Lite Image): the fastest and cheapest of the three.

The odd part is that the fast model outscores the expensive one. On the Artificial Analysis Text-to-Image Arena in September 2026, Nano Banana 2 sits fourth at Elo 1121 while Nano Banana Pro sits ninth at 1099. Blind voters, it seems, don’t reward 4K on a screen that can’t show it. Pro earns its price on reference-heavy work (14 inputs, five consistent faces) and on grounded factual images. For a plain product shot, use Nano Banana 2 and pocket the difference.

Every Nano Banana image carries an invisible SynthID watermark that can’t be stripped without degrading the picture. Free and Google AI Pro users also get a visible Gemini sparkle in the corner. Only AI Ultra subscribers lose the visible mark, so if you’re producing client deliverables on the $19.99 per month Google One AI Premium plan, budget for cropping or upgrade.

API pricing for Nano Banana Pro works out to roughly $0.134 per image at 1K to 2K and $0.24 at 4K, by third-party token estimates in 2026. Those are estimates rather than list prices.

Adoption is the number that tells you where the market is. Sundar Pichai said at I/O 2026 that Nano Banana models had generated more than 50 billion images cumulatively, up from 5 billion in October 2025, per Google’s I/O 2026 keynote recap. A tenfold jump in about seven months.

3. Midjourney V8.2 – Best for Cinematic Aesthetics

Midjourney V8.2, the default model since July 24, 2026, is still the best AI image generator for cinematic, painterly, art-directed images, and no arena score changes that. Midjourney doesn’t sit on the Artificial Analysis leaderboard, so the case rests on what people who test these models side by side keep reporting.

The version history moved quickly this year. V8.1 replaced v7 as the default on June 10, 2026, held for six weeks, and gave way to V8.2 on July 24, according to Midjourney’s version documentation. The V8.2 notes list three targets: aesthetic quality, personalization accuracy (how well the model learns your taste from your ratings) and fewer low-quality outputs per batch. Nothing about text. Nothing about counting objects. Midjourney knows what it is.

An August 2026 eight-model test by the HalfJourney Lab, running the same prompt through every major generator, put it plainly:

Midjourney “remains the benchmark for aesthetic output in 2026 … moody, painterly, with that specific Midjourney vibe, nothing else gets close.”

I’d agree with that, with one reservation. That “specific vibe” is recognizable enough that a trained eye can spot a Midjourney render in a mood board. Great for concept art. Less great for a brand that wants its imagery to look like nobody else’s.

Pricing hasn’t changed shape in 2026:

There’s still no public API. That single fact keeps Midjourney out of most automated pipelines. If your workflow is a designer in a chair choosing between four renders, it’s the right tool. If your workflow is a script generating 400 thumbnails overnight, it isn’t an option.

4. Stable Diffusion 3.5 – Best Open-Source for Privacy and Control

Stable Diffusion 3.5 is the best open-source AI image generator for teams that need a private, fully local pipeline in 2026, and it holds the slot on license terms and hardware reach rather than on freshness. There is no Stable Diffusion 4. Stability AI’s own news page listed no SD4 announcement as of July 20, 2026, and none has appeared since. The current release is the one Stability AI shipped between October 22 and 29, 2024.

That makes it roughly 20 months old. Old, in this field, is a real cost.

What you get, per Stability AI’s release announcement:

• Stable Diffusion 3.5 Large: 8.1 billion parameters on the MMDiT-X architecture, the top-quality checkpoint.
• Stable Diffusion 3.5 Large Turbo: the distilled, faster variant of Large.
• Stable Diffusion 3.5 Medium: 2.5 billion parameters, running in 9.9GB of VRAM.

That last figure is the one that keeps SD 3.5 on this list. Medium fits on a 12GB consumer card, which means a regulated team can run image generation on a workstation that never touches the internet. No API keys, no per-image bill, no prompt logs on someone else’s server.

The licence is the other half of the argument. The Stability AI Community License is free for non-commercial use and for any organization under $1 million in annual revenue. Above that line you negotiate an enterprise license, but the weights are still yours to run on-premise.

Be honest with yourself about the trade. A model from late 2024 won’t match GPT Image 2 on text or Nano Banana Pro on reference handling, and the gap widens every quarter Stability stays quiet. The buy case is narrow and clear: air-gapped deployment, reproducible outputs from fixed seeds, and zero data leaving the building. If you’re not sure whether that isolation requirement is real or inherited, an AI integration audit settles it faster than a bake-off does.

5. FLUX.2 [dev] and Kontext – Best for Editing and Compositing

FLUX.2 [dev], Black Forest Labs’ open-weight release from November 25, 2025, is the best AI image generator for controllable editing and compositing in 2026, and it retires the SDXL plus ControlNet plus IP-Adapter stack that held this job for two years. Those older workflows still run. Hugging Face’s diffusers library still documents IP-Adapter for SDXL, and Civitai still hosts active workflows built on it. They just aren’t the state of the art anymore.

The reason is reference handling. FLUX.2 takes up to 10 reference images in a single request and keeps character and style consistent across them, per Black Forest Labs’ FLUX.2 announcement. It edits at up to 4 megapixels. Anyone who has chained three ControlNet units and an IP-Adapter to get a product onto a new background, then fought the seams, will recognize how much that collapses. One model, ten references, done.

What the open-weight version gives you:

• FLUX.2 [dev]: 32 billion parameters, downloadable from Hugging Face, with code in Black Forest Labs’ GitHub repo.
• License: non-commercial by default, with a commercial license available separately.
• Deployment: local only. There is no hosted [dev] endpoint, so the private-pipeline case is fully yours to run.

Independent voting backs the editing claim. On the Artificial Analysis Image Editing Arena in September 2026, FLUX.2 [pro] Edit sits in the leading group alongside Nano Banana 2 Edit, with Microsoft’s MAI-Image-2.6 holding the top spot on that particular board. So it isn’t the single best editor by blind vote. It’s the best editor whose weights you can download.

The hosted tiers, for teams that want the API instead of a GPU, run as follows per Black Forest Labs’ pricing page in September 2026:

A 32B model is not a laptop model. Budget for a proper GPU or accept the hosted Pro tier at three cents an image, which still undercuts most of the closed vendors on this list. Either way, reproducibility comes from fixed seeds and a model file you control, which is the whole point of an editing pipeline that has to produce the same result next quarter.

6. Adobe Firefly Image 4 / Image 4 Ultra – Safest for Enterprise Workflows

Adobe Firefly Image 4 and Image 4 Ultra are the safest AI image generators for enterprise use in 2026 because Adobe pairs them with IP indemnification and C2PA Content Credentials on every output. Image 3, the model many Photoshop users still remember, is gone. Adobe pulled it from the Photoshop model picker on April 28, 2026 and removed it from every version and plan in August 2026, per Adobe’s Firefly Image 3 retirement notice.

The replacement predates the retirement by more than a year. Image 4 and Image 4 Ultra launched at Adobe MAX London on April 24, 2025, and have been the live models throughout 2026. Image 4 generates at up to 2K resolution. Image 4 Ultra targets professional-grade detail and costs twice as much per generation: 20 credits per image against 10 for standard, according to 2026 pricing trackers.

The surprise this year is what else lives inside the Firefly app. As of March 2026, Adobe’s Firefly has become a hub for 25 to 30+ partner models, including Google’s Nano Banana 2, OpenAI’s image generation, Runway Gen-4.5 and Black Forest Labs’ FLUX.2 [pro], alongside Adobe’s own models. That’s a real shift. A creative team can pick a third-party model for a draft and Adobe’s own model for the final, without leaving the app.

Read the fine print on that, though. The commercially-safe positioning (indemnity, C2PA credentials) is what Adobe’s own models were built to deliver. Treat the partner models inside Firefly as convenience, and check the coverage terms before you ship a partner-model image into a campaign.

Plan range in 2026:

• Free: entry tier with limited credits.
• Standard, Pro and Premium: roughly $50 to $200 per month across the paid tiers.

Where Firefly wins is procurement. If your legal team asks who trained the model on what, and who pays if a generated image draws a claim, Adobe has an answer on paper. For healthcare or financial-services marketing, that answer is often worth more than a few Elo points.

7. Microsoft MAI-Image-2 – Most Accessible for General Users

Microsoft’s MAI-Image-2 is the most accessible AI image generator of 2026: it is free to anyone with a Microsoft Account through Copilot and Bing Image Creator, and it scores in the top tier of independent blind voting. It takes the slot that DALL·E 3 used to hold, and DALL·E 3 no longer qualifies for a simple reason. OpenAI shut the DALL·E 2 and DALL·E 3 API endpoints on May 12, 2026, after moving ChatGPT users onto GPT Image models in December 2025. A retired product can’t be a recommendation.

MAI-Image-2 is a first for Microsoft. It’s the company’s own in-house image model rather than a repackaged OpenAI one, and it launched on March 19, 2026, debuting at third place on the Artificial Analysis Arena behind only OpenAI and Google, per Microsoft’s MAI-Image-2 announcement. The family has since grown into 2.5, 2.6 and 2.6-Flash variants.

Two numbers from September 2026 make the case:

• MAI-Image-2.6 holds second place on the Text-to-Image Arena at Elo 1149, behind only GPT Image 2.
• The same model tops the Image Editing Arena outright.

That second result caught me off guard. A free consumer model, six months old, beating every dedicated editing system on blind edit votes.

Microsoft pitched the model on photorealism (natural lighting, accurate skin tones) and on reliable in-image text for the infographics and slides its Office users make all day. It shows. The same rollout, through Copilot and Bing Image Creator across 2026, means a small business owner who already has a Microsoft login gets a top-three model at no cost.

The limit is on the developer side. API access is restricted to select enterprise customers pending wider availability through Microsoft Foundry, so you can’t build a pipeline on it today unless you’re already on that list. For a marketing manager who needs ten hero images by Friday, none of that matters. For an engineering team, it means MAI-Image-2 is the model you use in a browser, and GPT Image 2 or FLUX.2 is the one you call from code.

8. Ideogram 4.0 – Best for Posters and Typography

Ideogram 4.0, released June 3, 2026, is the best AI image generator for posters, typography and layout-driven design, and it’s now the highest-ranked open-weight model in designer-preference voting. That second half is new. Ideogram 3 was a closed product. Ideogram 4.0 is the company’s first open-weight foundation model, with checkpoints on Hugging Face and code on GitHub.

The architecture, per Ideogram’s own technical writeup: a 9.3 billion parameter single-stream Diffusion Transformer paired with a Qwen3-VL-8B text encoder, shipped as fp8 and nf4 checkpoints. The nf4 option is what makes local use plausible on a mid-range card.

The numbers that justify the slot, all from June 2026:

• Second overall and first among open-weight models in designer-preference blind voting.
• A DesignArena Elo around 1285, roughly 115 points ahead of the next open competitor, per AlphaSignal’s analysis of the release.
• A 0.97 OCR-accuracy score on text-rendering benchmarks.

That OCR figure is the one that matters for this job. It means the lettering on a poster reads back correctly 97 times out of 100, which is the difference between a usable social graphic and a round of manual fixes in Illustrator.

What I find more interesting than the benchmarks is structured JSON prompting. Instead of describing a layout in prose and hoping, you pass hex color palettes, bounding-box positions and typed text elements as a JSON object. A brand team with a style guide can encode it once and stop re-explaining “our teal, logo top-left, headline in the lower third” on every prompt. For a design system that lives in code, that’s the first image model that speaks the same language.

Pricing as of 2026:

Note the free tier’s catch. Public outputs mean anything you make on it is visible to other users, so client work belongs on a paid plan.

9. FLUX.2 and FLUX 3 – Open-Source Contender to Watch

FLUX.2 is the open-source contender that earns slot nine as the shipped, downloadable flagship from Black Forest Labs, while FLUX 3 is the model to watch and is not yet something you can download or buy. Keep the two apart. Conflating them is the easiest mistake to make about open-weight image models in September 2026.

FLUX.2 is settled. It replaced FLUX.1 on November 25, 2025, and it’s the open-core release you can run today. The editing specs, hosted pricing and local deployment story sit in the compositing entry earlier in this list. As a from-scratch generator it belongs here on the strength of being the largest modern open-weight image model with a commercial license path.

FLUX 3 is the ambitious one. It launched in early access on July 23, 2026, and it isn’t only an image model. One jointly-trained architecture handles images, video up to 20 seconds with native audio, and early robotics action prediction. The catch list, as of September 2026:

• Access: Discord and the Black Forest Labs API, for select partners only.
• Resolution: capped at 720p at launch, with 1080p expected afterward.
• Open weights: none yet. Black Forest Labs has said an open-weight “FLUX 3 Dev” is planned for later in 2026, with no confirmed date, license or model size.

A model you can’t get is a market signal, not a purchase. That’s why FLUX 3 stays outside the numbered ranking while FLUX.2 holds the slot. If the Dev weights land with a workable license before year end, the open-source order on this list changes. Until then, the honest advice is to build on FLUX.2 and design your pipeline so the model file is swappable.

One practical warning from running these locally: the version-to-version jumps in this family have not been drop-in. Expect prompt behavior, reference handling and VRAM needs to shift, and keep a regression set of prompts so you can see what moved.

10. Leonardo AI – Best Budget Platform for Teams

Leonardo AI is the best budget AI image platform for small teams in 2026, giving a managed workspace, 80+ selectable models and seat-based team plans without any GPU or DevOps overhead. It runs under Canva, which acquired Leonardo.ai in 2024 and confirmed in March 2026 that it’s “not exploring any redundancies,” keeping Leonardo as a standalone product with a dedicated team, per SmartCompany’s reporting on the integration.

That ownership question was the thing to check this year, and the answer is reassuring. Leonardo’s own Phoenix model now also powers Canva’s Dream Lab generative tool, which suggests the parent is investing in the engine rather than winding it down.

The pitch is breadth over a single hero model:

• Leonardo’s own models: Phoenix 1.0, Lucid Origin and Lucid Realism.
• Third-party models inside the same workspace: Google’s Nano Banana Pro and Ideogram, among the 80+ on offer.
• Realtime Canvas: sketch-to-image that updates roughly every 2 seconds per stroke, useful for blocking out a composition before committing to a full render.

Pricing as of 2026, with 20% off on annual billing:

Where it fits. A five-person studio that needs shared asset libraries and a predictable monthly bill, but has nobody to babysit a local FLUX.2 install, gets more from Leonardo than from any single model above it. Where it doesn’t: if your deliverable depends on one specific model’s ceiling, go to that vendor directly. Leonardo is the place to compare and produce, and the model marketplace is the point. For teams that outgrow it and want a custom generative AI feature built around one chosen model, that’s a different project with a different budget.

How to Choose the Right AI Image Generator

The right AI image generator in 2026 is the one that wins at your specific job, and the ten picks above split cleanly by job once you stop reading Elo as a verdict. Start from the deliverable, then work back to the model.

Use case to pick, as of September 2026:

• Text-heavy design (posters, packaging, slides): GPT Image 2 for accuracy across scripts, Ideogram 4.0 when the layout itself is the product.
• Photorealism at speed: Nano Banana 2 for volume, Nano Banana Pro when you need 4K or more than a handful of reference images.
• Cinematic or painterly art direction: Midjourney V8.2, accepting that there is no API.
• Fully local and private: Stable Diffusion 3.5 on the smallest hardware, FLUX.2 [dev] when you have a real GPU and need editing.
• Editing and compositing: FLUX.2 for open weights, MAI-Image-2.6 for the best blind-vote editing score you can use in a browser.
• Enterprise-safe with indemnity: Adobe Firefly Image 4 or 4 Ultra.
• Free and casual: MAI-Image-2 through Copilot or Bing, or GPT Image 2 in Instant mode.
• Small team on a budget: Leonardo AI.

Price is where the picks pull apart hardest. Per image through a hosted API in 2026, at roughly comparable output sizes:

A forty-fold spread. The cheapest tier of the arena leader costs less than the cheapest open-weight hosted tier, and the top tier of the same model costs more than almost anything else on the list. Quality tier selection per job is now the single biggest lever on an image budget.

One trade to weigh. The closed models with the strongest provenance (GPT Image 2, Nano Banana, Firefly) mark every output with C2PA credentials, SynthID or both, which a compliance team will love. Google’s visible watermark stays on unless you pay for AI Ultra. Open-weight models leave marking to you, which is freedom and liability in the same package.

Frequently Asked Questions

Is DALL·E 3 still available in 2026?

No. OpenAI shut down the DALL·E 2 and DALL·E 3 API endpoints on May 12, 2026, after warning developers in November 2025, and ChatGPT had already moved users to GPT Image models in December 2025. The current OpenAI image model is GPT Image 2, released April 21, 2026. Free ChatGPT users get its Instant mode.

What replaced Google Imagen?

Google’s Nano Banana family replaced Imagen. Google’s Gemini API documentation listed Imagen models as deprecated with a shutdown date of August 17, 2026, and pointed developers to Nano Banana instead. As of September 2026 the family has three tiers: Nano Banana Pro (up to 4K), Nano Banana 2 (the Gemini app default since March 2026) and Nano Banana 2 Lite.

Is Midjourney V8.2 free?

No. Midjourney V8.2, the default since July 24, 2026, requires a subscription of $10 (Basic), $30 (Standard), $60 (Pro) or $120 (Mega) per month, with 20% off on annual billing. There is no free tier and no public API.

Is there a Stable Diffusion 4?

No. As of September 2026, Stability AI has announced no Stable Diffusion 4, and its news page listed none as of July 20, 2026. The current release is Stable Diffusion 3.5 (Large, Large Turbo and Medium), shipped October 22 to 29, 2024, under the Community License that is free for organizations under $1 million in annual revenue.

Which AI image generator is best for text?

GPT Image 2 is the best for text in 2026, with a reasoning step that checks prompt constraints before generating and much stronger non-Latin script rendering than gpt-image-1. For design work where the typography is the point, Ideogram 4.0 posts a 0.97 OCR-accuracy score on text-rendering benchmarks and ranks second overall in designer-preference voting as of June 2026.

Which AI image generators are open source in 2026?

Three picks on this list ship downloadable weights. Stable Diffusion 3.5 (Community License, free under $1 million revenue), FLUX.2 [dev] (32 billion parameters, non-commercial license with a commercial option) and Ideogram 4.0 (9.3 billion parameters, fp8 and nf4 checkpoints on Hugging Face, released June 3, 2026). FLUX 3’s open-weight Dev variant is planned for later in 2026 with no confirmed date.

Do AI images have watermarks?

Most closed-model outputs do. Every Nano Banana image carries an invisible SynthID watermark, plus a visible Gemini sparkle for free and Google AI Pro users. OpenAI committed on May 19, 2026 to embedding SynthID alongside C2PA Content Credentials, and Adobe Firefly attaches C2PA credentials by default. Open-weight models such as Stable Diffusion 3.5 and FLUX.2 [dev] leave watermarking to whoever runs them.