AI image generators: a big comparison and prompting guide

An AI image generator is a tool that creates or edits an image from a text instruction (a prompt). In October 2026, according to the independent Arena and Artificial Analysis leaderboards, the best models are GPT Image 2.5 and GPT Image 2 from OpenAI, which replaced the earlier DALL·E, followed by Grok Imagine Image 2.0 from xAI, MAI-Image-2.6 from Microsoft and Nano Banana 2.1 from Google. The best value for money is Nano Banana 2.1 at $0.034 per image, while the best image from GPT Image costs around $0.21. Recraft suits vectors, Midjourney aesthetic illustrations and Adobe Firefly images that need commercial certainty. In this article you will find a comparison of 13 models with leaderboard figures, prices, recommendations for people, products, text and illustrations, a prompting guide with six pairs of bad and good prompts, and the rules that apply to AI images from August 2026.

Key takeaways

1. The Arena leaderboard (7 October 2026, 82 models, 6.6 million votes) is led by OpenAI's GPT Image 2.5 with 1,425 points.
2. Close behind OpenAI are Grok Imagine Image 2.0 (1,334), MAI-Image-2.6 (1,332) and Google's Nano Banana 2.1 (1,328).
3. The price per image through the API varies twentyfold: from $0.01 (Meta Muse Image, standard quality) to $0.21 (GPT Image, top quality). A quick draft from GPT Image 2 at the lowest quality costs $0.006.
4. A good prompt describes the scene, subject, action, composition, lighting, style and format. Phrase what you do not want positively.
5. Do not generate Czech text in the image, put it in the email editor.
6. Under the AI Act, from 2 August 2026 you must label realistic looking AI images of people and events.

In short: best overall, best value, best for text and best for vectors

• Best overall: GPT Image 2.5 from OpenAI, in ChatGPT and through the API. Expect the highest price and generation times of up to 2 minutes.
• Best value: Nano Banana 2.1 from Google, 4th to 6th place on the leaderboards at $0.034 per image, in the Gemini app.
• Best for text in images: GPT Image 2.5, outside OpenAI MAI-Image-2.6 and Nano Banana 2.1, among specialists Ideogram 4.0.
• Best for vectors and illustrations: Recraft V4.1 (true SVG), for aesthetics Midjourney V8.2, for commercial certainty Adobe Firefly.

What an AI image generator is and how it works

An AI image generator is a tool that creates a new image from a text description or edits an existing one according to an instruction. Behind it is an artificial intelligence model trained on a large number of captioned images, which has learned to connect words with shapes, colours, lighting and styles. You can use a generator in an app or through an API. An API is an interface through which another program or app orders images, and you pay for each image separately.

Current generators can do three things:

• Create an image from text (text-to-image): you write what you want to see and the model creates the image.
• Edit an image (image editing): you upload a photo and write what to change, for example the background, a colour or the season.
• Work with references (reference images): you upload a photo of a product, a character or a style and the model uses it in a new scene. Google's Nano Banana accepts up to 14 references in a single prompt.

A prompt is a text instruction for an artificial intelligence model. For images, it describes what should be in the image, what it should look like and in what format. The quality of the prompt matters more than the choice of model, which is why we devote a separate section with examples to it below.

Which AI image generators are the best: leaderboards

The most reliable comparison of AI image generators today comes from public leaderboards, where people blindly pick the better of two images for the same prompt. From thousands of these duels, the models get an Elo score, just like in chess. The largest is Arena (formerly LMArena) with 82 models and more than 6.6 million votes; the second is the Artificial Analysis leaderboard.

GPT Image 2.5 appears on the leaderboard in two versions: according to OpenAI, Sunburst is its most capable model for creating and precisely editing images, and Flare its fastest model for high quality everyday images. Both versions cost the same to generate.

Other well known models rank lower: Ideogram 4.0 in 20th place (1,205 points), Recraft V4.1 in 26th (1,169), FLUX.2 [max] from Black Forest Labs in 29th (1,162), Google's Imagen 4.0 Ultra in 37th (1,148), Leonardo Lucid Origin in 75th (1,013) and Stable Diffusion 3.5 Large in 81st place (938). The new FLUX 3 Image from 1 October 2026 is not yet on the Arena leaderboard; on the Artificial Analysis leaderboard it is in 8th place.

The Artificial Analysis leaderboard ranks the models similarly: GPT Image 2.5 Sunburst 1,198 points, Flare 1,191, GPT Image 2 1,172, Nano Banana 2.1 1,160, Grok Imagine Image 2.0 1,155 and MAI-Image-2.6 1,151. Both leaderboards therefore agree that OpenAI leads and the group of Google, Microsoft and xAI is close behind.

Note: Midjourney and Adobe Firefly are not on the leaderboards. That does not mean they are worse, only that they cannot be compared using the same method. Midjourney is known for its aesthetics and artistic style, Firefly for commercial certainty.

Comparison of AI image generators: what each one can do

GPT Image 2.5 can do the most, Nano Banana 2.1 is the cheapest quality choice, and for special tasks Recraft (vectors), Ideogram (text) and Midjourney (aesthetics) are a good fit. The table summarises the 13 most important models based on the makers' documentation, leaderboards and prices as of October 2026.

Source: documentation and price lists of OpenAI, Google, xAI, Recraft and Adobe, the Arena and Artificial Analysis leaderboards, prices of MAI-Image-2.6 according to RuntimeWire, Seedream 5.0 Pro according to OpenRouter and FLUX 3 Image according to Treza Labs, as of 12 October 2026.

How much AI images cost

One AI image costs from $0.01 at standard quality to $0.21 at GPT Image top quality through the API, and a typical app subscription from $5 to $20 a month. A quick draft from GPT Image 2 at the lowest quality comes to $0.006. For a regular newsletter, where you need one to three images a week, the difference is negligible. At larger volumes, for example in an online shop with dozens of product variants, it is decisive.

Most people, however, do not use generators through the API but in an app with a monthly subscription. Approximate prices as of October 2026:

Source: service price lists and price comparisons by CometAPI (ChatGPT), Engadget (Gemini), eesel.ai (Midjourney) and Adobe, October 2026. Prices change, so check them with the service before buying.

Which AI generator to choose for which image

The Arena leaderboard also rates models in five categories. GPT Image 2.5 leads in all of them, but the differences between the others show where a cheaper model pays off.

People and portraits

For realistic people, GPT Image 2.5 is the best: in the Portraits category it has 1,467 points and is more than 100 points ahead of the best model from another company. A cheaper choice with good results is Nano Banana 2.1. With people, however, remember two things: under the AI Act you must label a realistic looking image of a real person, and real people from your team are better photographed. Check hands, teeth and text on clothing, as even the best models occasionally get them wrong.

Product photos

For products and brand, GPT Image 2.5 leads (1,419 points), followed by Grok Imagine Image 2.0 and MAI-Image-2.6. AI, however, cannot faithfully preserve a specific product: it may change the shape, logo or colour. A more reliable approach is to photograph the product and use AI only for the background and setting, with the instruction “leave the product unchanged”. For precise edits of one part of an image, FLUX 3 Image is a good fit, as it can change just a defined area.

Images with text and banners

For text in images, GPT Image 2.5 leads with 1,472 points; outside OpenAI the best are MAI-Image-2.6 (1,371) and Nano Banana 2.1 (1,363). The typography specialist is Ideogram 4.0. However, we have not found a systematic test of Czech diacritics, and OpenAI itself lists text rendering among the limitations of its model. In emails, therefore, put text in the editor, not in the image.

Illustrations, vectors and brand style

For illustrations in a brand's style, Recraft V4.1 is unique, because it is the only one that creates true SVG vectors, which you can enlarge without loss of quality. For artistic looking illustrations, Midjourney V8.2 is popular, but it has no API and is not on the leaderboards. For images where you need commercial certainty, Adobe Firefly offers indemnification in intellectual property disputes in business contracts.

Editing existing photos

For photo editing, GPT Image 2.5 Sunburst leads the Arena leaderboard with 1,524 points. It also leads the Artificial Analysis editing leaderboard (1,184 points), followed by MAI-Image-2.6 and Nano Banana 2.1 with an identical 1,140 points. For precise local edits, FLUX 3 Image and generative fill in Photoshop are a good fit.

How to write a prompt for an AI image for a newsletter

You write a prompt for an AI image in six steps: scene and subject, action and details, composition, lighting and style, format, and gradual edits. The steps are based on guides from OpenAI, Google, Black Forest Labs and Adobe.

Step 1: Start with the scene and the main subject

Write where the scene takes place and what the main thing in the image is, in one specific sentence.

In its guide to GPT Image, OpenAI recommends always writing the prompt in the same order: background and scene, main subject, key details, constraints. The model is then less confused about what matters most.

Step 2: Add action and details

Describe what the subject is doing, what it is made of, what colours it has and what surrounds it.

In its prompting guide for Nano Banana, Google recommends the formula subject, action, setting, composition, style. The more specific the description, the more control you have over the result. Adobe advises using at least three words and leaving out openings such as “generate” or “create”.

Step 3: Set the composition and camera angle

State the viewpoint, the distance and where free space should remain in the image, for example for text in the email.

A photographer's vocabulary helps: top view, close up, wide shot, low angle, shallow depth of field, 50 mm lens. For a newsletter, a main subject off centre and a calm background on one side work well.

Step 4: Set the lighting and style

Describe the lighting, mood and style, for example a photo in natural daylight or a flat illustration.

For realistic images, OpenAI recommends writing the word photorealistic directly. For illustrations, name the technique: flat vector illustration, watercolour, 3D render, paper collage.

Step 5: Specify the format

Choose the aspect ratio and resolution according to the image's place in the email, for a full width header image for example 16:9 at 2K resolution.

1K means a longer side of around 1,024 px, 2K around 2,048 px and 4K around 4,096 px. Nano Banana handles ratios from 1:1 to 21:9, GPT Image from 1:3 to 3:1 and images up to 4K. Use lower quality for first drafts; OpenAI recommends this for speed and cost.

Step 6: Refine in small steps

Fine tune the result with one change in each follow up instruction and state explicitly what should stay unchanged.

OpenAI advises starting with a clean basic prompt and refining with small changes. For edits, write “change only X and keep everything else the same”. Google recommends stating explicitly what should stay exactly the same.

Tip: Describe what you do not want in the image positively. Instead of “no cars”, write “an empty street”. FLUX models do not support negative prompts and, according to Black Forest Labs, a negation can actually draw the model towards the unwanted thing. The exception is Midjourney with the --no parameter.

My tip: a guided prompt. I don't write prompts from scratch. In Claude I created my own skill that works as a guide: it asks me one by one about the points of this guide, that is the scene, the main subject, the action, the composition, the light, the style and the format. I just answer, and at the end I get a finished prompt that I paste into the image generator. This is also how the images for the articles on our blog are made. You can create such a skill too, just describe the questions from steps 1 to 6 to the assistant.

Prompt examples: bad and good

Six pairs show how to turn a generic instruction into a prompt that produces a usable newsletter image. You can write prompts in English or in your own language, most current models understand both.

1. Header image for an autumn promotion

• Bad: “autumn sale”
• Why: the model does not know what should be in the image, and the word sale tempts it to invent text.
• Good: “Photorealistic shot of a wooden table by a window with a mug of tea, a knitted sweater and a few yellow maple leaves. Warm afternoon light from the left, shallow depth of field, slightly elevated view. The right third of the image stays calm and blurred. Aspect ratio 16:9.”

2. Product photo

• Bad: “our coffee machine in a kitchen”
• Why: the model does not know your coffee machine and will invent a different one.
• Good: upload a photo of the product and write “Place the coffee machine from the attached image on a light stone worktop in a modern kitchen, morning light from a window on the right. Keep the coffee machine exactly as it is, with no change to its shape, colour or logo. Aspect ratio 4:5.”

3. Portrait for a team email

• Bad: “smiling woman in an office”
• Why: a generic shot looks like stock photography and cannot replace a real person from your team.
• Good: use a photo for real colleagues. For an illustration: “Flat vector style illustration: a woman with glasses at a laptop replying to an email, a mug of coffee and a plant beside her. Turquoise, cream and dark blue palette, white background, square format.” An illustration is not a deepfake, so you do not have to label it.

4. Banner with text

• Bad: “Black Friday banner 30% off everything”
• Why: text with diacritics and numbers often gets garbled, and recipients who do not load images will not see it.
• Good: “Dark blue background with a subtle paper texture, a gift box with a turquoise ribbon on the right, the left half of the image empty for text. Studio lighting, aspect ratio 3:1.” Then write the Black Friday headline and the discount in the email editor.

5. Editing an existing photo

• Bad: “make it wintry”
• Why: the model may also change the people, the product and the composition.
• Good: “Change only the season to winter: fresh snow on the roofs and trees in the background, overcast sky. Keep the people in the foreground, their clothing, the composition and the lighting unchanged.”

6. Webinar illustration

• Bad: “an image for an email marketing webinar, make it not boring, no stock photos, no robots”
• Why: negations and vague wishes give the model nothing it could draw.
• Good: “Paper collage: an open envelope with colourful paper arrows and small charts flying out of it, tiny paper figures around it. Pastel colours, soft shadows, front view, aspect ratio 16:9.”

How to upscale and refine an AI image

Most generators create an image at a resolution of around 1,024 px, but GPT Image, Nano Banana and FLUX 3 Image can also do 2K and 4K. For email, you need twice the width of the image, around 1,240 px with Quanda's default settings, so 2K resolution or slight upscaling is enough.

• Upscaling: tools such as Magnific (formerly Freepik) or Topaz Gigapixel enlarge the image and add detail. Topaz Gigapixel costs $149 a year in the personal version.
• Transparent background: GPT Image can create an image with a transparent background directly; save it as PNG.
• Expansion and fill: generative fill and expand in Photoshop, Nano Banana, GPT Image and FLUX 3 Image can extend an image sideways or replace part of it according to an instruction.
• Format for email: generators often save WebP. Classic Outlook does not display it, so convert the image to JPG, and graphics with transparency to PNG. The article JPG, PNG or GIF? explains why.

Rights to AI images and mandatory labelling

Use an AI image with three rules in mind: who holds the rights to it, when you must label it and what you must not upload to the generator.

Who owns the image. Under its terms of use, OpenAI assigns you the rights to the output. Midjourney grants rights to the images, but companies with annual revenue over 1 million USD must have the Pro or Mega plan. Adobe describes Firefly as commercially safe and offers indemnification in intellectual property disputes only to business customers with a contract.

It is not a copyright work in the Czech Republic. In a judgment of 11 October 2023 (case no. 10 C 13/2023), the Municipal Court in Prague ruled that an image created by artificial intelligence from a prompt is not a copyright work, because only a natural person can be an author and a prompt is merely an idea. Slovak law also requires the author to be a natural person, and Polish legal doctrine starts from the same principle. In practice, this means nobody can stop you from using a purely AI image, but you cannot stop others from copying it either.

When to label an image. From 2 August 2026, Article 50 of the EU Artificial Intelligence Act (AI Act) applies. Whoever publishes a deepfake, i.e. an image that resembles existing people, places or events and could appear authentic, must disclose that it was created by artificial intelligence. For evidently artistic, satirical or fictional works, a label that does not spoil the impression of the work is enough. Illustrations, collages or obviously stylised images do not fall under this obligation. A realistic photo of a “satisfied customer” or “our shop” that does not exist does.

Source: Content Credentials overview, Google and Meta documentation, as of October 2026.

What not to upload to a generator. Unsplash, Kaboompics and Canva prohibit using their photos for AI, and Kaboompics explicitly for image generation too. Use your own photos as references. Where to find legal photos is described in the article Free photos for emails: where to find and how to use them.

Personal data. A photo of an employee or customer is personal data. To upload it to a generator you need a legal basis, typically the consent of the person concerned, and you should know what the service will do with the photo:

• OpenAI: it may use content from personal ChatGPT plans (Free, Plus, Pro) to train its models unless you turn off the Improve the model for everyone option in the Data controls settings. By default, it does not use content from the Business and Enterprise plans or from the API for training.
• Google: the Gemini app uses uploaded images to improve its services when you have activity saving (Gemini Apps Activity) turned on, and some conversations are read by people. Business Google Workspace accounts do not use data for training.
• Midjourney: under its terms, you grant it a licence to uploaded content and outputs, and these are public by default. Only Stealth mode in the Pro and Mega plans can hide them, and we did not find an option to turn off training in the terms.

How to get an AI image into an email in Quanda

You insert an AI image into an email in Quanda just like any other. In the email editor, you insert a file from your computer or an image URL into the Image component. Quanda hosts the image itself.

If you create the campaign with an AI assistant in Claude or ChatGPT via the Quanda MCP server, you have three options:

• Link to the image: give the assistant the public URL of the image; Quanda downloads and hosts it when the campaign is saved.
• File from your computer: the assistant can upload it to Quanda itself, or you can upload it on the Images page in the Quanda MCP app, where you sort images into folders.
• Earlier images: the assistant can list previously uploaded images, such as a logo, and reuse them.

JPG, PNG and GIF are the formats that belong in email. Always add alternative text to the image, because classic Outlook does not download images from the internet by default. The article Can't see images in an email? Styled alt text shows a headline. shows how to style alternative text as a headline in Quanda.

How marketers use AI images

Most marketers already use AI for content creation. According to a HubSpot survey of more than 1,500 marketers in the USA, updated in August 2026, 80% of them use AI for content creation and 75% for media creation. The Litmus State of Email 2026 survey of 500 marketers in the USA, the UK, Australia and New Zealand from the turn of 2025 and 2026 shows how deeply AI has penetrated email marketing.

Checklist: an AI image for a newsletter

Before you send an AI image, check the model, prompt, details, text, format, labelling and references.

• The model suits the type of image: GPT Image or Nano Banana for people and products, Recraft for vectors.
• The prompt describes the scene, subject, composition, lighting, style and aspect ratio.
• I have inspected the hands, faces, text and product in the image up close.
• The text is in the email editor, not in the image.
• The image is at least twice its width in the email and saved as JPG or PNG.
• A realistic image of a person, place or event is labelled as AI.
• I have used only my own photos as references.

AI image generators: summary

The best AI image generator in October 2026 is GPT Image 2.5 from OpenAI, the best value for money is Google's Nano Banana 2.1, Recraft is the choice for vectors and Midjourney for aesthetic illustrations. More than the choice of model, however, the prompt decides: a specific scene, composition, lighting, style and format, positive phrasing and edits in small steps. Put text in the email editor and label realistic images of people and events.

AI model leaderboards change every month, so we update this article regularly. You can learn how to work with images in email in the free online course Email Copywriting and Design That Gets Read. And if you want to create campaigns and images with an AI assistant, get in touch and we will be happy to show you how.