The experiment

I am tired of AI image generator comparisons that test each tool with a different prompt and then conclude that "each has its strengths." That tells you nothing. So I did the obvious thing that apparently nobody else can be bothered to do: I used the same five prompts across ten different generators and compared the results directly.

The prompts were chosen to test specific capabilities: A photorealistic portrait - "A 40-year-old woman with short grey hair, reading a newspaper in a London cafe, morning light through the window, candid photography style" Text in an image - "A vintage bookshop sign that reads 'Thornton & Sons, Est. 1923' in gold lettering on dark green wood" A complex scene - "A busy street food market in Bangkok at dusk, neon signs reflecting in puddles, steam rising from cooking stations, shot on 35mm film" An illustration - "A children's book illustration of a fox wearing a blue scarf, walking through an autumn forest, watercolour style" Something weird - "A Renaissance oil painting of a cat sitting on a throne, wearing a tiny crown, looking extremely unimpressed"

I used each tool's default settings with no prompt engineering tricks. Just the raw prompt, exactly as written above. Here's what happened.

The results, ranked from best to worst Midjourney v6.1

I'll be honest, I expected Midjourney to win and it did. But not by as much as I expected.

The portrait was stunning. Perfect lighting, natural skin texture, genuinely looks like a photograph. The London cafe felt authentic rather than generic. The newspaper looked like an actual newspaper rather than the usual AI blur.

The text prompt was decent but not perfect. "Thornton & Sons" was legible, but "Est. 1923" came out as "Esc. 1923." Close, but wrong. This is still Midjourney's weakness.

The Bangkok market was where it really shone. The neon reflections, the steam, the depth of field - it looked like a photograph from a travel magazine. The atmosphere was perfect.

The fox illustration was beautiful but had that very specific Midjourney aesthetic where everything looks like high-end concept art. If you wanted a simple, gentle watercolour, this was too polished.

The cat painting was magnificent. Genuinely looked like something that could hang in a gallery. The cat's expression of mild contempt was perfect.

Score: 9/10 - Still the best overall, but the text rendering remains a weakness. Flux Pro

Flux surprised me. I had expected it to be good but it was genuinely neck-and-neck with Midjourney on several prompts.

The portrait was slightly less aesthetically polished than Midjourney's but arguably more realistic. It looked more like an actual photograph and less like a fashion shoot. The cafe background was more believable.

Where Flux pulled ahead was the text prompt. "Thornton & Sons, Est. 1923" was rendered perfectly. Every letter correct. This alone makes it the better choice for anything involving text in images.

The Bangkok scene was excellent but lacked the atmospheric magic that Midjourney brought. Technically accurate, emotionally flat. Like comparing a skilled photographer to an artist.

Score: 8.5/10 - Best for text rendering and photorealism. Less "artistic" than Midjourney. Ideogram 2.0

ideogram is the text specialist and it proved it emphatically. The bookshop sign was perfect. Not just legible but aesthetically beautiful. The gold lettering had texture and depth. The wood grain was convincing. This was the best result from any tool on that specific prompt.

The portrait and scenes were good but not outstanding. Ideogram produces images that look slightly more "digital" than Midjourney or Flux. There is a sheen to everything that gives it away.

The illustration was surprisingly lovely. The watercolour style was more authentic than Midjourney's, which defaulted to its usual polished look.

Score: 8/10 - The best at text by a clear margin. Good but not exceptional at everything else. Leonardo AI

leonardo-ai does not get enough attention. The model fine-tuning feature means you can train it on a specific style, which none of the others offer at this level. Out of the box, the results were solid without being spectacular.

The portrait had a slightly soft quality that some might prefer. The scene images were well-composed but lacked the fine detail of the top three. Where Leonardo surprised me was the illustration prompt, the fox was charming and genuinely looked like a children's book illustration rather than concept art.

The free tier generosity (150 tokens/day) also deserves mention. You can seriously evaluate this tool without paying.

Score: 7.5/10 - Underrated. Great for illustration and style consistency. DALL-E 3 (via ChatGPT)

dall-e lands in the middle of the pack, which is exactly where I expected it.

The convenience factor is real. Being able to refine your prompt through conversation, describing what you want changed about the previous result, is a genuinely superior workflow. No other tool makes iteration this easy.

But the images themselves are... fine. The portrait looked like a stock photo. The Bangkok scene was flat and lacked atmosphere. The illustration was generic. The cat painting was technically competent but somehow boring, which is an achievement given the prompt.

The text rendering was surprisingly decent. Not Ideogram-level, but "Thornton & Sons" was correct, only "Est. 1923" became "Est 1923" (missing the full stop). Minor, but still wrong.

Score: 7/10 - The most convenient tool with the most mediocre results. The definition of "perfectly adequate." Stable Diffusion XL (local)

stable-diffusion is the free, open-source option that requires you to run it on your own hardware. The quality ceiling is high because you can use custom models and fine-tune everything. The quality floor is also low because the default settings produce unremarkable results.

With the default SDXL model and no tweaking, the results were mixed. The portrait was decent. The scene images were good but needed upscaling. The illustration was bland. The cat painting was actually quite good, possibly because art-style images are well-represented in the training data.

The text rendering was dreadful. Every single prompt came out as illegible gibberish.

If you are willing to spend time learning ComfyUI, downloading custom models, and fine-tuning settings, you can get results that match or exceed Midjourney. But "after hours of configuration" is a significant caveat that most comparisons gloss over.

Score: 6.5/10 - High ceiling, low floor, steep learning curve. Adobe Firefly 3

adobe-firefly is the "safe" choice. Trained on licensed content, commercially safe, integrated into Photoshop and Illustrator. And the results are... safe. Very safe.

The portrait looked like it came from a stock photo library. Not a good stock photo library. The kind where everyone is suspiciously attractive and the lighting is suspiciously perfect and nothing feels real.

The Bangkok scene was bland. The illustration was generic. The cat painting was the best result from Firefly, probably because "generic but polished" works well for classical art styles.

The text rendering was actually reasonable. Not as good as Ideogram or Flux, but better than Midjourney. Adobe clearly prioritised this.

Score: 6/10 - Buy it for the commercial license, not the quality.

8-10. The rest

I'll group the bottom three because the results were consistently disappointing enough that detailed breakdowns would be cruel.

Tool 8 produced images that looked like they were from 2023. Blurry backgrounds, weird hand anatomy, inconsistent lighting. The technology has clearly not been updated in a while.

Tool 9 had decent quality but absurdly aggressive content filters. Two of my five prompts were rejected for unspecified "policy violations." A woman reading a newspaper in a cafe violated their policies. Incredible.

Tool 10 was a newer entrant that clearly was not ready for prime time. Inconsistent quality, long generation times, and a user interface that felt like it was designed by someone who had never used a computer before.

What I actually learned

The gap between the best and worst AI image generators is enormous. Midjourney and Flux produce images that are genuinely difficult to distinguish from real photographs. The bottom tier produces images that look like they were made with a tool from three years ago.

Text rendering is still the most reliable way to separate good tools from mediocre ones. If you need text in your images, use Ideogram or Flux. Everything else is a gamble.

The "best" tool depends entirely on what you need. Midjourney for beauty, Flux for accuracy, Ideogram for text, Leonardo for style consistency, DALL-E for convenience. Pick the one that matches your actual use case rather than the one that wins on vibes.

And if you are still using the default tool that came with your AI chatbot subscription, you are missing out. The dedicated image generators are in a completely different class.

My setup going forward

I'm keeping Midjourney for editorial and creative images. The aesthetic quality is unmatched when you want something that looks genuinely striking.

I'm using Flux for anything that needs text, product mockups, or photorealism where accuracy matters more than atmosphere.

And I'm using DALL-E for quick throwaway images where convenience matters more than quality. Sometimes you just need a picture of a cat for a presentation and you do not want to leave ChatGPT to get it.

Everything else gets cancelled. Three is enough. More than three is a waste of money and an excuse to procrastinate by "testing" tools instead of doing actual work.