I tested AI image tools for six months — this is the only one worth paying for

Last October I generated a Renaissance-style portrait of my dog using DALL-E 3 and it came out with six legs. Not like, subtly wrong. Six fully rendered legs, confident as anything, like the AI had made an artistic choice. That was month one of what became a slightly obsessive testing project across nine different AI image generators.

I wasn't doing this for fun, exactly. I needed consistent illustrations for a client's newsletter — nothing fancy, just clean visuals that didn't scream "stock photo from 2015." The promise was simple: describe what you want, get something usable. The reality was messier. Way messier.

The Six-Month Chaos

I ran the same ten prompts through every major tool: Midjourney, DALL-E 3, Stable Diffusion (multiple versions), Leonardo AI, Ideogram, Adobe Firefly, Playground AI, and a couple others I've already forgotten. Same prompts, same use cases. Product mockups. Editorial illustrations. Simple icons. Portraits with specific ethnicities and ages.

Here's what I expected going in: Midjourney would be the artistic one, DALL-E would be the accurate one, and the free tools would be garbage. I was wrong about all three.

Midjourney produced the most visually striking images, sure. But "striking" and "usable" aren't the same thing. When I needed a simple illustration of a person working at a laptop, Midjourney gave me dramatic lighting, moody atmosphere, and hands that looked like they'd been through a blender. Beautiful. Completely wrong.

DALL-E 3 was supposed to be the "follows instructions" option. And it does, mostly. But the aesthetic is weirdly sterile. Everything looks like it was designed for a corporate PowerPoint about combination. Clean lines, safe colors, zero personality. Also, approximately 40% of my generations got flagged for policy violations. I asked for a chef holding a knife. Flagged. A person at a protest. Flagged. A kid with a birthday cake and candles. Somehow, flagged.

The free tools were actually better than expected at first. Then you hit the rate limits, the watermarks, the sudden resolution drops, the "premium feature" paywalls that appear exactly when you need something specific.

The Thing Nobody Tells You About Consistency

Here's what broke my brain for the first three months: getting one good image is easy. Getting ten good images in the same style is nearly impossible.

I'd generate a perfect illustration for a newsletter header. Great. Then I'd need a matching one for the next section. Same prompt structure, same style keywords. Completely different vibe. Different color palette. Different level of detail. One looked like a children's book, the next like a movie poster.

This is the problem nobody talks about in the YouTube tutorials. They show you the hit. The one perfect generation they got after 47 attempts. They don't show the other 46, and they definitely don't show you trying to make image number 48 match image number one.

I burned through probably $200 in credits across platforms before I figured out that most of these tools are designed for one-off "wow" moments, not actual work.

Why I Landed on Leonardo AI

After six months, I pay for exactly one tool: Leonardo AI. About $12 a month. And I almost didn't try it because the name sounded like it was invented by a marketing committee.

Here's why it works for me: it has this feature called "Alchemy" that genuinely stabilizes output. Same prompt ten times, you get ten images that actually look related. Not identical, but coherent. Like they came from the same artist having the same day.

The hands are still wrong maybe 15% of the time. But that's down from the 60%+ failure rate I was seeing elsewhere. And there's a control called "Prompt Magic" (again, terrible name) that seems to actually parse what you're asking for instead of just grabbing keywords.

I asked for "a middle-aged South Asian woman reviewing documents at a modern office desk, warm afternoon lighting, editorial style." DALL-E gave me a stock photo robot. Midjourney gave me a cinematic scene with so much bokeh I couldn't see the documents. Leonardo gave me exactly what I described, looking like an actual photograph from an actual magazine.

It's not perfect. The interface is cluttered. There are like seventeen different model options and most of them are redundant. The mobile app is genuinely terrible. But the output is usable about 80% of the time, and that number holds consistent.

The Real Discovery (The Kick)

Here's the thing I only learned after month four, and I haven't seen anyone else mention it: Leonardo has a "Negative Prompt" field that actually works.

In theory, every tool lets you specify what you don't want. In practice, most ignore it completely. Tell DALL-E "no text" and you'll still get random floating letters. Tell Midjourney "no dramatic lighting" and it'll give you slightly less dramatic lighting, maybe.

Leonardo's negative prompts are weirdly specific. I type "no glasses, no jewelry, no patterns on clothing" and I get a person with no glasses, no jewelry, and plain clothing. I type "no watermarks, no borders, no text overlays" and the image is clean. It sounds basic. It's actually remarkable.

The workaround I developed: I now spend as much time on my negative prompt as my positive prompt. Usually 30-50 words of "don't do this" for every 30-50 words of "do this." My generations went from maybe 50% usable to closer to 85% usable once I figured this out.

One specific combo that solved my biggest headache: "no distorted hands, no extra fingers, no missing fingers, no merged fingers" in the negative prompt. It's not 100%. But it's close enough that I'm not regenerating the same image twelve times hoping for normal human anatomy.

I still use Midjourney occasionally for concept art, stuff where weird is okay. I still use DALL-E through ChatGPT for quick one-offs where I don't care about the aesthetic. But for anything I'm actually delivering to someone, anything that needs to look professional and consistent, it's Leonardo.

The $12 is annoying. But I spent twice that per month in wasted credits on other platforms before I stopped. And I spent maybe four hours a week re-generating failed images. Now it's closer to forty minutes. The math is boring but the math is real.

Still can't get it to draw a bicycle correctly though. Not once in six months. Every AI draws bicycles like they've only heard them described over a bad phone connection. I've given up on that one.

Heads up: Some links in this post may be affiliate links. I only recommend tools I've personally tested. Opinions are entirely my own.