The best AI image tools for ad agencies in 2026

At a glance

• Top pick: Adobe Firefly. Its own model plus 30-odd partner models, Custom Models trained on 10 to 30 of a client's images, and the only IP indemnity anyone here offers. Read the fourth bullet before you relax about that last part.

• Best for volume: Melius. An agent builds the node graph and you can open any step it built. Ten variants in parallel per node, a documented ten-market localization pass, and around 30 models behind one subscription.

• Free tiers differ more than the prices do. Ideogram grants commercial rights on every tier including free. Recraft's free output is owned by Recraft and can't be sold at all.

• Every indemnity in this category has a hole where agencies stand. Adobe's caps at $10,000 and comes with no Firefly plan. OpenAI's excludes trademark claims in trade or commerce. Midjourney's runs the other way.

• No vendor can make raw output copyrightable. The US Copyright Office holds that prompts alone don't give you enough control to be the author. What you add is protectable; the generation isn't.

These tools stopped competing on the same thing

Two years ago, choosing an AI image tool for agency work meant choosing whichever model made the nicest picture. That contest is close to settled and it stopped being the interesting question. The six here now do genuinely different jobs, and the distance between those jobs is wider than the distance between their outputs.

Firefly trains a model on your client's own product shots. Melius runs thirty other companies' models on a graph you can open and edit step by step. Recraft emits vector you can hand to a printer, Ideogram sets legible type inside the image, and GPT Image 2.5 is an API you build a pipeline against. Midjourney still looks better than any of them and is close to impossible to buy. Which of those you need is the first decision, and it matters more than which demo reel you liked.

The second decision is what the paperwork says, because somebody's legal team will ask. Every vendor here advertises an answer to "what if we get sued", and almost none survive being read. Adobe's indemnity caps at $10,000 and comes with no Firefly plan. OpenAI's excludes trademark claims arising from use of the output in trade or commerce, which is the definition of advertising. Midjourney's runs the other way. Terms break ties here. They don't decide the ranking.

How we picked

• What the tool actually does. Generation, orchestration, vector output and typography are four different jobs. A tool that's excellent at one can be useless at another.

• Throughput. Unilever reported going from 20 assets per campaign to an average of 400 per product. Different denominators, so not a clean multiple, but a tool that makes one nice image is no longer the job.

• Brand control that holds. Style references, brand kits and trained models, judged across a run of variants, because holding a style for one generation is the easy part.

• Terms and procurement fit. Who owns the output, what the vendor licenses back from you, and whether they'll engage with a security review at all. Several will not.

• How we compared. Nothing was tested hands-on. This rests on published pricing, terms and license text, vendor documentation and what practitioners report, checked 17 September 2026.

Tool comparison

1. Adobe Firefly - the widest toolkit, and the only indemnity

Firefly is the broadest toolkit here. It runs Adobe's own Firefly Image Model 5, trained on Adobe Stock and licensed content, and brokers more than 30 partner models alongside it, including Nano Banana 2, FLUX, GPT Image and Runway. Custom Models train on 10 to 30 of a client's own images for style or character consistency, and the whole thing sits inside the Creative Cloud apps your retouchers already work in. Adobe named Dentsu, Publicis, Stagwell, IPG Health and Monks among its customers in March 2025.

The indemnity is the other reason it wins procurement, and the fine print is worth your time. Adobe caps liability for indemnified Firefly outputs at $10,000 per output or related set of claims, which will not cover a national campaign. Coverage doesn't attach to any Firefly plan either, including Firefly Premium at $199.99 a month; it rides on a Creative Cloud Pro Plus or Edition 4 agreement. It then excludes outputs you modified, outputs combined with other material, the context you use them in, and anything marked beta. Generate, composite in Photoshop, set type over it, and you've left the covered path three times.

Here's the catch, and it lands on the two features you'd actually buy Firefly for. Adobe says plainly that deciding whether a partner model is safe for commercial use is the creator's job, and no partner model output is indemnified. Neither is Custom Models output: it's still in beta, beta is excluded by name, and Custom Models has never appeared on Adobe's list of eligible features anyway. Use either one and the commercial-safety story you bought Firefly for is gone, with nothing flagging it at the moment you choose.

Best for: Agencies whose client contracts warrant clean IP, who will read the entitlement before relying on it.

Price: Standard $9.99/mo, Pro $19.99/mo, Pro Plus $49.99/mo, Premium $199.99/mo, list prices as of September 2026. Firefly Pro for teams is $19.99 per license.

Platforms: Web, Creative Cloud apps, API via Firefly Services.

Commercial use: You own the output. Indemnification comes with no Firefly plan, rides on a Creative Cloud agreement, caps at $10,000, and excludes modified and combined output, the context of use, every partner model, and anything in beta, which includes Custom Models.

Watch out for: A generation that completes but ignores half your prompt still costs you the credits (outright failures are refunded). Content Credentials are embedded on qualifying output, can't be switched off, and the terms forbid removing them.

Link: adobe.com/products/firefly

2. Melius - an agent that builds a graph you can open

Melius is a canvas of nodes wired together by edges. A node is one model call, and an edge feeds its output into the next node's input, so a campaign becomes a graph you can look at instead of a conversation you have to remember. You don't have to build that graph, though. An agent writes the prompts, creates the nodes, wires them up and picks the models for each step.

The pairing is the real idea. In a chat box the reasoning is hidden, so a wrong result means a new prompt and the loss of whatever was working. Here every step is an object with an editable prompt on it: you open the node that went off and re-run that one. Failure is local, which is a real difference. 

The agent lets you create faster and at greater scale than other tools, and editing nodes in the draft give control over the fine details. This helps agencies focus more on ideation and design while saving time on execution work. 

The main caveat here isn’t about the software. Melius is a seed-stage company of eight to fifteen people, launched publicly in 2026. General Catalyst is among the named investors, so the backing isn't thin, but the paperwork isn't there yet.

Best for: Building and localizing large variant sets across several models. Around 30 image models sit behind one subscription, and ten variations run in parallel on a single node.

Price: Creator $20/mo, Growth $50/mo, Professional from $110/mo on a credit slider, list prices as of September 2026, 15% off annual. Seats are unlimited on every paid tier.

Platforms: Web, desktop, Slack, API, CLI, MCP. Slack starts on Professional.

Commercial use: Granted on every tier including free, and you own the output. There's no IP indemnity, and you indemnify Melius against third-party claims.

Watch out for: Uploads are JPEG, PNG, WebP and HEIC only.

Link: melius.com

3. OpenAI GPT Image 2.5 - production engine with an API

OpenAI shipped ChatGPT Images 2.5 on 8 September 2026, with two models on the API side: Flare for speed, Sunburst for tighter control on detailed edits. The API is why it's here. It composites multiple reference images into one shot, generates on transparent backgrounds for compositing, and supports multi-turn editing, so you refine one asset across passes instead of regenerating it and losing what worked.

It's a strong production engine and a mediocre art director. OpenAI's own docs concede the model can struggle to hold brand elements consistent across generations, and to place elements precisely in layout-sensitive compositions. Those are the two things campaign work leans on hardest. Earlier GPT Image models also carried a warm color cast strong enough that people built correction tools for it; whether 2.5 still does is untested at nine days old.

Best for: Teams with an engineer, wiring generation into an existing asset pipeline.

Price: A standard ChatGPT Business seat is $25 monthly or $20 annually, two-seat minimum, as of September 2026; premium seats are $125 and $100. The API bills $30 per million output tokens, roughly $0.006 for a low-quality 1024px image.

Platforms: ChatGPT web and apps, REST API.

Commercial use: OpenAI assigns you its interest in the output "if any", a quitclaim, not a warranty that copyright exists to assign. The output indemnity covers API customers and ChatGPT Enterprise, Edu and Healthcare; on the current wording it doesn't reach ChatGPT Business, and it excludes trademark claims from use of output in trade or commerce.

Watch out for: Rate limits start at five images a minute on Tier 1 and reach 250 only at Tier 5, slow for a variant sweep on a new account. Signups to the $200 Pro plan have been paused since 10 September 2026; the $100 plan is unaffected.

Link: openai.com

4. Recraft - vector output for print and packaging

Everything else here makes pixels. Recraft's V4 models generate vectors directly, separately from its raster-tracing Vectorize tool, which matters the moment a deliverable goes to print, packaging or out-of-home, where a 2K raster is a dead end. Its V4 Styles feature takes one to ten reference images and learns line weight, color logic and rendering style with no training step, and on Teams those styles are shared across the workspace with SSO.

Recraft is unusually candid about where that breaks: mixed references don't blend, they collide. It works when your references already agree with each other, meaning pages from one consistent brand book, and one reference is the safe default.

Best for: Deliverables that need to scale cleanly, and brand systems run across a campaign.

Price: Basic $12/mo, Pro $20/mo, Teams $22 per seat monthly, as of September 2026, about 20% off annual. Vector generations cost far more than raster: 4 credits on the V4 vector models, 12 on the Pro vector ones.

Platforms: Web, REST API, Figma, Framer, Chrome extension.

Commercial use: Paid plans assign you the copyright. Free-tier assets are owned by Recraft, can't be sold, and are public. On every tier Recraft takes a perpetual license over your inputs and the assets you make, with model training on by default and an opt-out in settings.

Watch out for: Specified dimensions are unreliable and output sizes shift after generation, and targeted edits often fail to register. Teams is less shared than it looks: credits are per seat, only top-ups pool, and the terms set a three-seat minimum the pricing page doesn't mention.

Link: recraft.ai

5. Ideogram - when the copy is inside the image

If the deliverable has words in it, Ideogram 4.0 is the one to reach for, and it isn't close. On Ideogram's own benchmarking it scores about 0.97 on in-image text accuracy, and because it was trained exclusively on structured captions you can specify bounding boxes, per-element styling and color palettes. Multi-line, multi-font layouts hold together.

The cost is a prompting model that reads like filling in a form, and it goes deeper than the interface: the pipeline validates every prompt against a schema. On a team that isn't technically inclined that's a real adoption tax, and it's the most common complaint from people who otherwise rate the output highly.

Best for: Headlines, posters, packaging comps and anything where legible typography is the point.

Price: Plus $20/mo, Pro $60/mo, Team $30 per user monthly with a two-seat minimum, as of September 2026. Annual cuts those to $15, $42 and $20. The API runs $0.03 Turbo, $0.06 Balanced and $0.10 Quality per image.

Platforms: Web, REST API.

Commercial use: Granted on every tier including free, and Ideogram assigns you any rights it holds. Generations are public by default; private starts on Plus. The 4.0 weights carry a non-commercial license even though the code is Apache-2.0, so self-hosting for client work needs a separate license from $300/mo. Wikipedia says the whole thing is Apache-2.0, and it's wrong.

Watch out for: On paid plans, reference and edit tools run on metered priority credits, outside the unlimited slow queue, which undercuts the unlimited headline for exactly the reference-heavy work agencies do. Batch starts at Pro; private brand models are Enterprise only.

Link: ideogram.ai

6. Midjourney - the best looking, the worst fit

V8.2 still makes the most attractive images in this category, and for concepting, moodboards and pitch visuals that counts for a lot. Style references carry a look across a set, moodboards turn a curated group of images into a reusable personalization code, and the rebuilt Edit Model absorbed what character reference and Omni Reference used to do.

It is also, uniquely, a vendor that tells you in writing it won't meet your procurement requirements. Its own documentation lists what it does not offer: seat or admin management, terms negotiation, SSO, and participation in vendor or security review processes. Consolidated billing needs fifty or more annual Pro or Mega plans at no discount, and the recommended way to track your team is a spreadsheet. There's no API, and the terms prohibit automated access.

Best for: Concepting, art direction and pitch visuals that never ship as finished creative.

Price: Basic $10/mo, Standard $30/mo, Pro $60/mo, Mega $120/mo, as of September 2026, with 20% off annual. Extra GPU time is $4 an hour on every plan.

Platforms: Web and Discord. No API.

Commercial use: You own your output, but anyone at a company over $1M in revenue needs Pro or Mega to do so. You grant Midjourney a perpetual, irrevocable license over your inputs and outputs that survives termination, worth reading before uploading a client's confidential assets.

Watch out for: Your work is public and remixable by default. Stealth mode is Pro and above, and Midjourney commits only to best efforts on it. Upgrading lets you make earlier work private, but you do it image by image, and anything made in a public Discord channel stays visible either way.

Link: midjourney.com

Which one fits your desk

• Client legal reviews every asset: Firefly, with the Creative Cloud entitlement actually purchased and the exclusions read. Keep the partner-model dropdown off limits.

• You need many variants with attention to detail: Melius for the parallel batching and localization, or GPT Image through the API if you have an engineer and would rather own the pipeline.

• The work goes to print or out-of-home: Recraft, for vector output, and check the dimensions on every export.

• The headline lives inside the image: Ideogram, and accept the schema-shaped prompting.

• Still pitching, nothing shipping yet: Midjourney, on a personal seat, with nothing confidential uploaded.

[H2] What to watch over the next six months

Two things are moving faster than the tools. The EU AI Act's transparency obligations applied from 2 August 2026, and the duty falls on whoever runs the system under their own authority, which on most agency engagements points at the client. That's fact-specific enough to settle in the contract before a campaign ships. New York's synthetic performer disclosure law has applied since 9 June 2026, where the advertiser knows a synthetic performer is in the ad.

The awkward part is what disclosure costs. In a field experiment by researchers at NYU Stern and Emory, fully AI-generated ads beat human-made ones by up to 19% on click-through, while labeling an ad as AI-made cut click-through by 1.17 percentage points, a relative drop of 31.5%. The human-plus-AI hybrid did worst of the three. The creative performs better, and saying so out loud hands the gain back.

Worth reading next: our roundup of open-weight image models, for work that has to stay on your own hardware, and our breakdown of what shipped in GPT Image 2.5.