Before you pick a tool
There is one question worth answering in advance: what is the output, and who is going to use it? A mood image for the blog, a product shot for the webshop, a logo for print and an illustration for a slide deck are four different jobs, and no single tool is the best at all four.
The other advance question is whether AI is the right tool for the job at all — in general terms Can AI Do This Task? is the post that answers it. If the answer is yes, then the choice comes next.
Who is who in the autumn of 2026
The market looks roughly like this. I deliberately leave version numbers out, because they go stale in six months; the families and their strengths are more stable.
- OpenAI — GPT Image. It leads the independent leaderboards built on blind human votes, in text-to-image and in editing alike. It is included in the ChatGPT subscriptions, so there is nothing extra to pay for it.
- Google — Nano Banana. The strongest non-OpenAI generalist, in several sizes. Its strengths are handling several reference images at once and multilingual text.
- Microsoft AI — MAI-Image. The surprise entrant: it takes half the editing categories, and leads outright in text correction and restoration.
- xAI — Grok Imagine. Mid-table in the rankings, available from the API, and it handles reference images and batch runs.
- Midjourney. Strong aesthetics and a house visual language, though no longer at the top of the independent rankings — its following is not an accident either.
- Recraft. The only one that gives you genuine, editable vector output. For graphic work that can be the deciding factor.
- Ideogram. For a long time the benchmark for text rendering; today it no longer holds the category lead, but its typographic approach is still usable.
- Black Forest Labs — FLUX and Alibaba — Qwen-Image. The downloadable, locally runnable line. On that, Running Image Models Locally goes into far more depth.
Five use cases, five picks
- Photorealism — whatever leads the rankings, which means the OpenAI image model; the Google generalist as the alternative. If you want something that passes for a photograph, the risk is smallest here.
- Graphics, illustration, icons — Recraft, because of the vector output. Not because it is prettier, but because you can genuinely edit the result afterwards.
- Typography, lettering inside the image — the Microsoft image model leads the text and sign editing category. But the real advice is still to put the lettering on afterwards.
- Product images — the leading model for identity-preserving editing (from the ByteDance family), or the top OpenAI variant. With a caveat: in strict fidelity measurements none of them gets above 30 percent.
- Open source, local runs — the smaller, permissively licensed FLUX variant and the twenty-billion-parameter Qwen-Image model. Those are the two genuinely unrestricted choices.
What the leaderboards do not tell you
The independent Elo scores are useful because they are built on blind human votes — but there are three things they do not measure. One is the workflow: a tool that gives you twenty variants quickly and re-runs easily can be better in practice than a higher-scoring but awkward one.
The second is style: the ranking measures average preference, while your brand needs one specific visual world. The third is availability: a model you cannot get, or can only get in a corporate plan, is not an option. So treat the ranking as a starting point, and run the same ten prompts through two or three tools.
The ownership question, which differs by tool
This is the most frequently skipped consideration, and commercially the most important one. Vendor terms come in four very different shapes.
- There are vendors that assign you every right they hold in the output — with the note that they may generate something similar for someone else.
- There are ones where it is tied to revenue: above a certain annual turnover the image is only yours on the more expensive plan.
- There are ones where an image made on the free plan belongs to the vendor, goes into a public gallery, and may not be used commercially.
- And there are ones where the vendor explicitly makes no claim on the images, whatever the plan.
On top of that, a vendor term is not the same thing as copyright: the vendor "giving you" the image does not by itself create a copyright in it. On that point, Who Owns an AI-Generated Image? is the detailed treatment.
How to choose in ten minutes
1. Write down THREE real jobs you would actually do in the next month.
Not "nice images" — specifically: product shot on white,
blog cover, social post with a caption.
2. Pick two tools from the list above, whichever fit the job.
3. Run the same three jobs on both, with the same prompt.
4. Do not look for the prettiest image, look at HOW MANY OUT OF TEN are usable.
That number decides, not the one lucky hit.
5. Read the terms: who owns the image, and what you may do with it.
This step costs the least time and does the most damage when it is skipped.
Common misconceptions
- "The best model is the best at everything." The category breakdown shows the exact opposite: editing, typography and restoration each belong to a different model.
- "If it is open, it is free to use." No. Some of the downloadable models are explicitly non-commercial — and it varies from variant to variant.
- "The ranking measures image quality." It measures human preference, which is not the same thing. In a fidelity measurement the order can even reverse.
- "My old favourite is still the best." In this market the field turns over in nine months. Re-testing regularly is cheaper than staying with the wrong tool.
What to take away
Do not look for the best tool, look for the one that fits your job — and make the decision on how many out of ten prompts are usable, not on one lucky image. And the thing to give five minutes to whatever else you skip: the ownership part of the vendor terms. If you also work with old photographs, that is a separate class of tool, which Upscaling and Restoring Old Photos walks through, and for understanding the mechanism itself How Image Generation Works is the starting point.