I'm a developer, not a designer. My artistic ability peaks at drawing stick figures. So when I spent last weekend generating images with Midjourney, the results genuinely shocked me. I went from "this is a cool toy" to "wait, this changes everything about how I think about visual content" in about two hours.
Getting Started
Midjourney runs through Discord, which felt weird at first. You type prompts into a chat channel and a bot sends back images. It's not the most elegant interface, but it works. I subscribed to the Basic plan at $10/month and got about 200 image generations to play with.
My first prompt was embarrassingly simple: "a futuristic city at sunset." The result was beautiful. Not "good for AI" beautiful. Actually beautiful. The kind of image I'd use as a desktop wallpaper without thinking twice.
What I Learned About Prompts
The prompt is everything. Vague prompts give you generic results. Specific prompts give you stunning, precise results. Here's what I figured out over the weekend.
Style references matter. Adding "in the style of Studio Ghibli" or "cinematic lighting, 35mm film" dramatically changes the output. You're essentially telling the model which corner of its training data to pull from.
Negative prompts are powerful. Adding "--no text, blurry, distorted" cleans up the output significantly. Midjourney v4 still struggles with text and sometimes with hands, so explicitly excluding those problems helps.
Aspect ratios change composition. The default square output is fine, but switching to "--ar 16:9" for landscapes or "--ar 9:16" for portraits gives much more natural-looking results. The model adapts its composition to the canvas shape.
Stack descriptors. Instead of "a house," try "a weathered Victorian house, overgrown garden, misty morning, volumetric lighting, hyper-detailed, 8K." Each descriptor nudges the output in a specific direction. I spent an hour just adding and removing descriptors to see how each one changed the result.
The Results That Blew My Mind
A few prompts that produced genuinely amazing results:
- "Abandoned space station orbiting a gas giant, interior shot through broken window, volumetric god rays, concept art style" - This looked like it could be concept art for a AAA game.
- "Japanese garden in autumn rain, shallow depth of field, Fuji film simulation" - Looked like an actual photograph taken by a professional photographer.
- "Developer's desk at 2am, multiple monitors with code, coffee cups, warm lamp light, isometric view, detailed illustration" - I immediately wanted to use this for my blog.
Where It Fails
Hands. Always the hands. I generated a portrait and the person had seven fingers. It's a known issue and it's still not fully solved in v4.
Text is also unreliable. If your image needs readable text, Midjourney isn't the tool. Every attempt at generating an image with text produced gibberish that looks like letters but isn't.
Consistency is hard too. If you want multiple images of the same character, you'll get variations that are close but never identical. This limits its usefulness for things like comic strips or brand assets that need consistency.
What This Means for Developers
I've already used Midjourney to generate placeholder images for a project, a hero image for a landing page, and icons for a side project. The quality is high enough for production use in many cases. Previously I'd spend time on Unsplash trying to find something close to what I imagined, or I'd pay for stock photos. Now I just describe what I want.
For developers who build products but can't afford a designer, this is a genuine superpower. Blog post headers, app screenshots, marketing materials, social media images. All of these are now trivially easy to generate.
I'm planning to dive deeper into Stable Diffusion next, which you can run locally. But for now, Midjourney's quality-to-effort ratio is the best I've found. The $10/month subscription has already paid for itself in time saved searching for stock images.
We're living in a strange and exciting time. A developer with zero artistic skill can now generate professional-quality images by typing a sentence. I don't know where this is going, but I'm along for the ride.