A retro pulp comic-book illustration with a writer's wooden desk seen from slightly above, with an open book and a typewriter in the center. Three glowing paths of light lead out from the open pages toward three different destinations: on the left, a big friendly cloud with a robot head holding a paintbrush; in the middle, a market stall with many small framed pictures lined up on shelves like a gallery wall; on the right, a sturdy desktop computer with a glowing graphics card, wires, and a small cozy lamp.

Three Ways to Make AI Images (and which models are worth using right now…)

A few years ago, “AI image generator” meant that you typed a prompt into a Discord server and waited.

Today there are dozens of models, and the same model is often available in four or five different places at four or five different prices.

That’s good news, because it means you can almost certainly find a way to work that fits your budget and your comfort level. It’s also confusing news, because “which AI should I use?” really has two questions hiding inside it: which model do you want, and where do you want to use it?

Let’s tackle both today!

First, the three main ways to get at an image model (directly from the company, through an aggregator, or on your own computer). Then a quick rundown of the models worth knowing about right now, what each is good at, and what you can expect to pay.

(If you’d rather watch than read, we’re putting the two newest models head-to-head live on Tuesday, October 27.)

Method 1: Go Directly to the Source

The simplest approach is to use the company that built the model. You sign up, type a prompt, and get an image. There are really two flavors of “direct.”

The app or website

This is what most people start with, and for many authors it’s all they ever need.

  • ChatGPT (OpenAI) makes images inside the same chat you already use for everything else. You can describe a picture, then ask for changes in plain English (“make the sky more orange, and move the cat to the left”). There’s a free plan with limited images, and paid plans (about $20 a month for Plus) give you more.
  • Gemini (Google) works the same way, and it’s where Nano Banana lives. Free accounts get a limited number of images, and Google’s paid AI plans raise the limits.
  • Midjourney is its own website (it no longer requires Discord). It has no free plan. Plans start at $10 a month.

Good for: getting started, casual use, and anyone who wants one simple login and/or is already subscribed to a service for something else they offer.

Watch out for: limits. Subscription apps rarely tell you exactly how many images you get, and the limits can change without much notice. Also, the app sometimes picks the model for you, so double-check which one you’re actually using.

The API

An API is the “behind the scenes” door that apps use to talk to the models. Instead of paying a monthly subscription, you pay a small amount per image, and you use the model through a program that connects to it with a key.

This is how we usually work. During our last head-to-head test in April when ChatGPT Image 2 first launched, we generated a lot of images, including a number of them for people attending live, and spent less than $2 in total. We use TypingMind with our own API keys, which lets us run the same prompt through several different models without switching between websites.

Good for: people who generate in bursts rather than every day, anyone who wants to control costs, and anyone who wants to compare models side by side.

Watch out for: a bit more setup, and you’re responsible for watching your own spending. Set a spending limit on the account before you start.

Method 2: Aggregators (One Login, Many Models)

An aggregator is a service that puts a lot of different models behind a single website and a single bill. Instead of subscribing to OpenAI and Google and Midjourney, you pay one company and pick the model from a dropdown.

Leonardo is the best-known example. Here’s what it listed on its pricing page this week:

  • 🆓 Free: 150 tokens a day, with limited model access
  • 💰 Essential: $12 a month, 8,500 tokens a month, access to 80+ models
  • 💰 Premium: $30 a month, 25,000 tokens, plus unlimited “relaxed” (slower) generation on select models
  • 💰 Ultimate: $60 a month, 60,000 tokens, with the same relaxed generation, plus video

Other options in the same spirit include Adobe Firefly, which includes Adobe’s own models alongside partner models from companies like Google and OpenAI (partner models use up your plan’s credits faster), and sites like Krea and Freepik that bundle models, upscalers, and editing tools. For a more do-it-yourself version, services like OpenRouter, fal.ai, and Replicate let you pay per image with one account (and are what many other tools use behind the scenes).

Good for: trying lots of different models without juggling accounts, and for people who want editing and upscaling tools in the same place.

Watch out for:

  • Token and credit math. You’re buying a pile of credits, and different models cost different amounts. A fancy model can burn through a month of tokens quickly, and unused monthly tokens are usually forfeited. Always look at what one image costs in that service.
  • A lag on new models. The newest model from a company often shows up on the company’s own site first and in aggregators a few days or even weeks later.
  • A markup. You’re paying for the convenience. For heavy use of a single model, going direct is usually cheaper.
  • Different rules. Each model still has its own terms, and the aggregator’s terms sit on top of them. Check the licensing for the model you actually pick.

Method 3: Run It on Your Own Computer

Some image models are “open weights,” which means the company released the model itself and anyone can download it. If your computer is powerful enough, you can generate images locally. There are no per-image fees (outside of the electricity to run your computer), no monthly subscription, and nothing leaves your machine.

The main tool for this is ComfyUI (free, powerful, and a little intimidating because it works with connected blocks instead of a simple prompt box). On a Mac, Draw Things is a free app that’s much friendlier to start with.

Here’s a rough idea of what you need. These are guidelines, not guarantees, because the numbers shift with the specific version and settings:

  • About 8GB of video memory (VRAM): small, fast models such as FLUX.2 [klein] 4B, or quantized builds of Z-Image Turbo
  • 12-16GB: mid-size models like Qwen-Image (in compressed versions)
  • 24GB or more: the larger models like FLUX.2 [dev]
  • On a Mac: the memory is shared between the system and the graphics, so a Mac with 24GB or more of memory is comfortable for the smaller models. Expect about a minute per image rather than a few seconds.

Good for: people who generate a lot of images, authors who want to train a style or character of their own, privacy, and anyone who enjoys tinkering.

Watch out for:

  • The hardware cost. A good graphics card costs more than years of an ordinary subscription. If you don’t already have the computer, this probably isn’t the cheapest route.
  • The learning curve. Expect to spend an afternoon (or a weekend) getting set up.
  • A step behind. The best open models are good, but the top commercial models still tend to be better at following long, detailed instructions and at getting text right.
  • Licenses vary. “Open” doesn’t always mean “free for commercial use.” FLUX.2 [dev], for example, is non-commercial, while the smaller FLUX.2 [klein] 4B uses the Apache 2.0 license, which allows commercial use. Read the license before you put the images in a book you’re selling.

The Models: A Quick Rundown

Here are the models we’d pay attention to right now, newest first within each family. Prices are what the companies (or, where noted, third-party trackers) listed in October 2026.

🆓 = free option · 💲 = pay per image · 💰 = subscription

ChatGPT Image 2.5 (OpenAI) 🆓 💰 💲

The newest from OpenAI, available in ChatGPT and through the API in two flavors: Sunburst, which OpenAI recommends when editing precision matters most, and Flare, which is aimed at fast, high-quality everyday images.

  • Where: ChatGPT (free and paid), the API, and many aggregators
  • Cost: included in ChatGPT plans. Through the API, OpenAI bills by the amount of image “tokens” used instead of a flat price. Third-party estimates put a standard square image at somewhere between a cent (low quality) and about five cents (high quality), and the new top settings (xhigh and max) cost more.
  • Use it when: you want to describe and then edit in conversation, you need readable text in the picture, or you want to start from a photo or sketch and change it.
  • Skip it when: you need dozens of images fast on a budget, or you’re after a particular painterly, stylized look (more on that below).

Nano Banana 2.1 (Google) 🆓 💰 💲

Google’s newest model as we write this, which was released on October 6. It’s fast, handles edits well, can use up to 14 reference images, and in my testing is about 40% cheaper than the previous generation of Nano Banana. (Official estimates are 50-60% less expensive.)

  • Where: the Gemini app, Google AI Studio, the API, and many aggregators
  • Cost: free with limits in Gemini. Through the API, Google lists about 3.4 cents for a 1K image, 5 cents for 2K, and 11 cents for 4K, and half that for batch (slower, non-urgent) jobs. There’s no free tier for image output on the API.
  • Use it when: you want fast edits or character consistency from reference images.
  • Skip it when: you need the absolute best text and fine detail (see Nano Banana Pro below).

Nano Banana Pro (Google), the older one that’s still worth knowing 💲

Nano Banana Pro isn’t obsolete. It’s a heavier model that costs more (about 13 cents for a 1K or 2K image, 24 cents at 4K). Reviewers consistently point to it for complex layouts, fine typography, and images with lots of precise detail, and it’s the one to try when 2.1 can’t quite get something right.

Midjourney V8 💰

Midjourney is still the one people reach for when they want an image that looks artistic without a lot of effort. Its latest default is in the V8 series (V8.2 at the time of writing).

  • Where: midjourney.com only
  • Cost: $10 (Basic), $30 (Standard), $60 (Pro), or $120 (Mega) a month, with a 20% discount for annual billing. No free plan. Companies earning more than $1 million a year need the Pro or Mega plan for commercial use.
  • Use it when: you want a distinctive, polished look for covers, mood boards, or concept art, and you’re happy to explore a lot of variations.
  • Skip it when: you need exact text on the page, precise edits, or consistent characters. According to current write-ups, V8 doesn’t support the old Character Reference feature. Its Omni Reference works, but runs on the older V7 model. For a picture book series, that means either V7 or one of the other tools.

FLUX (Black Forest Labs) 💲 💰 and 🆓 locally

FLUX is the big name in open models. FLUX.2 has a range of sizes: [klein] runs on a typical gaming computer, [dev] is the large open-weights model, and [pro] and [max] are the paid, hosted versions and cost between $0.03 and $0.07 per image. Black Forest Labs’ also has the new FLUX 3 as well, which can handle multi-modal media such as video and audio as well as images. Images cost $0.008 to $0.036 per image using their [klein] implementation, or $0.041 to $0.607 per image with the normal Flux 3 Image (depending on final resolution, up to 4k!)

  • Where: Black Forest Labs’ own API, most aggregators (including Leonardo), or your own computer
  • Cost: pay-as-you-go via the API (a few cents per image), or free if you run an open version yourself
  • Use it when: you want photorealism, or you want to run things locally
  • Skip it when: you want simple. Setting up a local version is a project.

Ideogram 4.0 💲 💰 and 🆓 locally

Ideogram built its reputation on putting readable text in images. Version 4.0 came out in June as an open-weights model, and text is still its main selling point.

  • Use it when: your image is mostly words: posters, quote graphics, signs, a cover mockup with a title
  • Cost: Ideogram’s pricing page lists 4.0 at 3, 6, or 10 credits per image (Turbo, Balanced, or Quality). Pay-as-you-go credits are 1 cent each, so that’s about 3 to 10 cents an image. Subscriptions start at $20 a month (Plus, 2,400 priority credits), with a $60 Pro plan for heavy use.

Adobe Firefly Image 5 💰

Adobe’s own model, trained on licensed content (Adobe Stock and public domain works) rather than whatever was found online.

  • Use it when: you’re working with clients or publishers who care where the images came from, or you already live in Photoshop and Illustrator
  • Cost: Standard is $9.99 a month (2,000 generative credits) and Pro is $19.99 a month (4,000 credits), with pricier plans above that. There are also free daily generations, including some partner models like Nano Banana. Partner models (Google, OpenAI, Flux, and others) draw from your credits.
  • Keep in mind: the “commercially safe” claim only applies to Adobe’s own models, not the partner models in the same app.

Also worth a look

  • Recraft V4.1: designed for designers. It’s a good fit if you need vector images (logos, icons, simple illustrations that can scale to any size). Third-party listings show vector images at around 8 cents each.
  • Seedream 5.0 (ByteDance) and Grok Imagine (xAI): both show up near the top of public rankings, at around 5 cents an image through aggregators. We haven’t tested either for book projects yet.
  • Z-Image and Qwen-Image: open models that are popular with the local crowd.

So, Which Should You Choose?

There’s no single answer, but here’s a simple way to narrow it down:

  • “I just want to make some pictures.” Start with the free version of ChatGPT or Gemini. Both are good enough to find out whether this is useful for you.
  • “I need a lot of images, but not every day.” Go API, so you pay per picture. It’s the cheapest way to make a picture book’s worth of images (and you can set a spending limit).
  • “I want to try everything.” An aggregator like Leonardo is a good fit. One bill, lots of models.
  • “I want the prettiest, most artistic covers.” Midjourney.
  • “I need text in the image.” ChatGPT Image 2.5, Nano Banana Pro, or Ideogram.
  • “I make a lot of images and have a good computer.” Try local with Draw Things or ComfyUI.
  • “My publisher is nervous about AI.” Look at Adobe Firefly, and be upfront about how you work.

And in all cases, test your own idea. A model that wins on a benchmark may still stumble over your specific cover, character, or scene.

Before You Publish Anything

A few quick reminders:

  • Read the terms. Every service has its own rules about commercial use. Those rules can differ between plans and can change.
  • Disclose AI use where it’s required. For example, Amazon KDP asks you to tell them when your book includes AI-generated images.
  • Keep your prompts and your originals. If a question about an image ever comes up, you’ll be glad to have a record.
  • Think about copyright. In the U.S., works made without meaningful human authorship are difficult to protect. If an image is central to your brand, consider editing or combining it by hand.

(We cover the licensing basics in our guide to sourcing legal, eye-catching art for your book.)

See Two of Them Go Head to Head

Reading about models only gets you so far. Next Tuesday, October 27, at 1:00 pm Eastern, Erin and I will run ChatGPT Image 2.5 and Nano Banana 2.1 side by side, live, on the same prompts. We’ll look at speed, quality settings, and what each image costs.

And we want your ideas. When you sign up, you can send in a prompt (or a reference image you own) that you’d like us to try. The more specific the better: a cover style, a picture book scene, a social media graphic, or something you’ve struggled to get an image generator to do.

==> Save your spot for the live head-to-head

It’s free and open to everyone, and members can watch the replay if they can’t make it live.

Similar Posts