AI image generators compared: what’s good, what’s not yours
Type the same prompt into three different AI image generators and you’ll get three genuinely different pictures, not just stylistic variations. That’s the first thing worth knowing before you spend any money on a subscription: these tools are not interchangeable, and the marketing copy won’t tell you which one actually suits what you want to make. We’ve spent a fair chunk of this year testing prompts across the big names, and the differences are bigger than most reviews let on.
The second thing worth knowing, and the one nobody wants to sit with for long, is that the copyright situation underneath all of this is still genuinely unresolved. Not “there’s fine print you should read”. Unresolved, as in courts overseas are still working it out and Australian law hasn’t caught up either.
What the major tools are actually good at
Midjourney remains the one most people point to when they want something that looks deliberately composed, almost like a painting or a concept-art piece. It has a house style, a sort of lush, dramatic lighting default, and even when you fight against it the results still look like Midjourney made them. That’s fine if you want atmosphere. It’s annoying if you want something that looks like an ordinary photo.
OpenAI’s image tools, built into ChatGPT, lean the other way. They’re better at following instructions precisely, including text inside the image (signage, labels, that sort of thing), and they handle edits to an existing image more sensibly than most rivals. If you need a specific diagram-style graphic with correct-looking text, this is usually the one that gets closest.
Google’s Gemini image generation, built on its Imagen models, tends to produce the most photorealistic results out of the box, probably because Google has an enormous amount of photographic data to draw on. It’s also the one most tightly integrated into everyday Android use, which matters more than people realise, because the barrier to generating an image drops to almost zero when it’s sitting inside your phone’s keyboard.
Adobe Firefly sits in its own category because Adobe has made a specific, narrower promise: it trains primarily on Adobe Stock content and public domain material, and it indemnifies commercial users against copyright claims arising from Firefly-generated content used in Creative Cloud apps. That’s a meaningfully different pitch to “look how creative this is”, and it’s the one we’d point small businesses toward first.
Our one-line verdict: there’s no single best tool, only a best tool for the specific job in front of you.
The copyright question nobody has actually answered
Here’s the bit that gets glossed over. Every major image generator was trained on enormous datasets scraped from the internet, which included copyrighted photographs, illustrations and artwork, almost certainly without the original creators’ permission. Several major lawsuits in the United States, brought by artists, stock photo libraries and news organisations, are testing whether that training process itself was lawful. None of them had reached a final, settled conclusion as this piece was written.
Australia doesn’t have a specific carve-out in copyright law for AI training, and the Attorney-General’s Department has been consulting on exactly this question rather than legislating a clear answer. You can read the department’s own framing of the issue on the Attorney-General’s Department copyright page. That leaves Australian users in an odd spot: the tools are legal to use, but whether the underlying models were built lawfully is a separate, murkier question that rumbles on in the background regardless of what you do with the output.
Then there’s the second layer, which is who owns what you generate. In the United States, the US Copyright Office’s guidance on AI and copyright takes the position that purely AI-generated images, with no meaningful human creative input, can’t be copyrighted at all. Australia’s Copyright Act has historically required a human author too, so a prompt alone probably doesn’t cut it as “authorship” in the traditional sense. If you’re building a brand around AI-generated artwork, that’s worth sitting with before you commission a logo this way.
Why this matters for small businesses
We hear from small business owners fairly often who’ve used an AI tool to knock out a logo, a flyer, or product photography because it’s cheap and fast, sometimes free. The practical risk isn’t usually a lawsuit from an artist whose style got echoed. It’s narrower and more mundane: you might not actually own what you’ve made in a way that stops someone else using it too, and you might unknowingly produce something close enough to existing copyrighted material to cause a headache later.
If you’re setting up a small business site and need imagery for it, our view is that a tool with an indemnity, like Firefly, or stock photography with a proper licence, is the safer commercial choice over a general-purpose generator, even if the general tool makes a prettier picture. We’ve covered the broader setup process in our guide to building a small business website, and image sourcing is one of the sections worth reading slowly rather than skimming.
The style mimicry debate is overblown, in our view
Here’s where we’ll disagree with a lot of the commentary. The loudest argument against AI image generators is that they let anyone produce art “in the style of” a specific living artist, and that this is a moral catastrophe. We think that framing overstates the actual harm in most ordinary, non-commercial use. Human artists have always learned by imitating other artists, consciously, as technique; the idea that a style itself is property has never been how copyright worked, in Australia or anywhere else. What copyright protects is a specific expression, not a vibe.
The more serious problem is the training data question itself, not the output style. A tool trained without consent on someone’s actual photographs is a different issue to a tool that merely produces images that look vaguely like Rembrandt. Conflating the two muddies a conversation that badly needs the two threads kept separate.
Short version: worry less about “it copied a style” and more about “what was it actually trained on, and did anyone agree to that”.
What this means for everyday, non-commercial use
If you’re generating a birthday card, a meme for the family group chat, or a quick mock-up of how a room might look repainted, none of this legal uncertainty should stop you. The risk profile for personal, non-commercial use is low, and the practical privacy questions matter more day to day than the copyright ones. We’ve written separately about the broader privacy trade-offs of feeding your own photos and data into AI tools in our explainer on the privacy cost of on-device AI, which is worth a read if you’re uploading family photos to any of these services for editing.
One genuinely practical tip: check what each platform’s terms say about whether your uploaded images are used to further train the model. Some let you opt out, some don’t ask, and the setting is often buried somewhere unhelpful in account preferences rather than on the main screen. The OAIC’s privacy guidance is a reasonable starting point if you want to understand what rights you actually have over data once it’s uploaded.
Pricing and what Australians are actually paying
Most of these tools now sit in roughly the same bracket, somewhere around $15 to $35 a month for the mid-tier paid plan, though free tiers with limited generations exist for most of them and are genuinely usable for occasional, casual use. Midjourney requires a subscription from the start with no meaningful free option left. ChatGPT’s image generation is bundled into its existing paid tiers rather than sold separately. Gemini’s image features come bundled into Google’s existing subscription structure too, which is handy if you’re already paying for extra cloud storage.
None of these numbers are fixed for long. Expect them to shift as the companies compete harder on bundling AI features into subscriptions you already have, the same pattern we watched play out with cloud storage tiers a decade ago, where the free allowance slowly tightened as the paid tiers became the actual product. If your household is already juggling more subscriptions than you can name, our piece on managing too many subscriptions is a useful companion read before you add another one.
A practical way to choose
Our rule of thumb: if you want something that looks hand-crafted and arty, try Midjourney first. If you need accurate text inside an image or careful editing of an existing photo, ChatGPT’s tools usually get there with less fighting. If photorealism matters most, and you’re already in the Google ecosystem, Gemini is the path of least resistance. If you’re putting the result on a real business asset, Firefly’s indemnity is worth the subscription on its own.
None of this requires deep technical knowledge, which is sort of the point of these tools existing at all. The actual decision that matters most, in our view, is less about which tool makes the prettiest picture and more about what you intend to do with the result afterward.
Our verdict on using these tools responsibly
We’d rather see Australians use these tools with a clear-eyed sense of the trade-offs than avoid them out of vague unease, or dive in assuming the output is automatically theirs to do whatever they like with. For personal use, have a crack at whichever tool suits the job and don’t lose sleep over it. For anything commercial, treat the copyright uncertainty as a real cost of doing business, not a technicality, and budget for proper licensing or an indemnified tool accordingly.
The legal landscape here will keep shifting faster than most buyer’s guides can track, ours included. Check the current terms of service before any commercial use, because what’s true in November 2026 may well not be true by the time this finishes rendering on your screen.
— Dale Whitfield, Founding Editor, Tech Geek




