Someone types 'beautiful moody portrait of a samurai, cinematic, highly detailed, 8k, masterpiece' into their model and hits generate. What comes back looks like a random fantasy book cover. Nothing like the reference photo open in the next tab.
That botched samurai is where one Reddit user's month-long experiment started, and it turned into one of the sharper breakdowns I've seen on r/PromptEngineering. The plan was simple: find an image worth learning from, describe what you see, generate, compare, repeat. The first batch came back as what the original poster called 'adjective soup,' stacks of vague words that told the model nothing.
Here's the number that matters: this creator went from about 1 decent recreation out of 20 attempts to close to 1 out of 4 by the end of the month. No secret model, no plugin, no fine-tune. Just five habits you can copy today, plus a two-second test that flags a weak prompt before you burn a single generation.
100+ ChatGPT Prompts to Revolutionize Your Day
Supercharge your productivity with HubSpot's comprehensive guide. This free resource is your fast track to AI mastery:
• Industry-Specific Use Cases: 15+ real-world applications across various sectors
• Productivity Guide: 21 best practices to 10x your efficiency with AI
• Prompt Powerhouse: 100+ ready-to-use prompts for immediate implementation
• Challenge Buster: Overcome common AI hurdles with expert strategies
Plus, in-depth sections on email composition, content creation, customer support, and data analysis.
Find Your Guide Here
*Ad
Habit 1: Why adjective soup never works
Piling on words like 'cinematic' and 'masterpiece' feels productive, since those are the words you see everywhere. But they carry almost no information the model can act on. Vague words let it fall back on defaults, and those defaults are exactly what make AI images look like AI images.
Here's why this matters beyond one samurai. The same failure hits anyone chasing a reference photo or a named artist's look and landing on generic mush. The model has millions of 'cinematic masterpiece' images to average toward, so give it nothing specific and it hands you the average.
What struck me here: the fix costs nothing. No new tool, no subscription, no fine-tune. The author changed how they wrote, not what they ran.
Habit 2: Order the prompt like a sentence
The next lesson is about sequence. Subject first, setting second, lighting third, camera last. Word order works like a weighting system, even when nobody says so.
Flip that order and things break. Put lighting first and the model sometimes treats light as the entire point, ignoring the subject you cared about. I think that one detail explains half the 'why is my image wrong' posts floating around.
*Ad
People are asking ChatGPT and Claude the same questions they used to type into a search bar, and most brands aren't showing up yet. Are you?
Get a free audit of your paid search and AI search presence, plus a plan to improve both.
Habit 3: Lighting is half the prompt
This is the part that sold me: treat lighting as about half of what you write. 'Cinematic lighting' means nothing specific. The version that worked described it plainly.
'Late afternoon sun coming through a window on the left, the rest of the room falling into shadow.'
That hands the model a real scene to build instead of a vibe. Most failed recreations, according to the original poster, died right here, on a detailed subject wrapped in vague or missing light. Skip the light, you skip the image.
The reason is mechanical, not mystical. Light direction, hardness, and what falls into shadow define depth, mood, and where your eye lands. Describe those three things and you've already built most of the image before you name a single object.
Habit 4: Steal ten photography terms
You don't own a camera. Doesn't matter. Writing '85mm, shallow depth of field' beats 'blurry background' every time, and the contributor says this single habit paid off more than anything else on the list.
Focal length and depth of field give you direct control the vague words can't. '85mm' tells the model to compress the background and flatter a face, while 'blurry background' just begs for a guess. One is a dial, the other is a shrug.
The same logic runs through color. 'Muted teal with rust orange accents' beats 'colorful,' and the expert runs a color picker over reference images to name the exact palette. The model speaks photography.
So speak back to it.
That vocabulary is the part you cannot fake, and reading a list of terms is not the same as knowing which one to reach for. There is an ai-images track in 3 Minute AI that runs about three minutes a day and makes you write the prompt in a built-in lab instead of nodding at an example. Daily lessons and two full courses are free, which is enough to find out whether the reps stick.
AI can build faster. Can your team decide better?
AI can draft the PRD and prototype the idea. Jira Product Discovery helps teams decide whether it belongs on the roadmap. Bring feedback and ideas together, prioritize as a team, and keep your roadmap connected to delivery in Jira.
*Ad
Habit 5: Write it like a phone call
The last habit ties the rest together: write full sentences, around 100 to 150 words, the way you'd explain a scene to a friend on the phone. Keyword lists leave gaps. The model fills those gaps with its generic defaults.
Which leads to the trick worth stealing: the phone test. Read your prompt out loud before you open your image tool. If a friend could sketch the scene from your words, ship it; if they'd ask what they're even looking at, that's your gap, not the model.
Two more moves worth trying. Describe references from memory instead of keeping the image open beside you, which forces you to keep only the details that matter. And build a running list of your ten go-to photography terms, because you'll reuse them constantly.
One honest caveat: this won't fix everything. The creator admits some styles, mixed-media textures most of all, refuse to cooperate no matter how the prompt reads. Good to know before you blame your own writing.
Blu Dot surpasses 2,000% ROAS with self-serve CTV ads
Blu Dot used Roku Ads Manager to drive incredible results for its furniture sales event. Its strategy hinged on custom audiences and retargeting, where intent was strongest.
“Roku has been a top performer,” said Blu Dot’s Claire Folkestad. “We have seen…CPMs lower than any other CTV partner we've worked with.”
*Ad
Rewrite one prompt before you generate again
Open your next image prompt and rewrite it in this order: subject, setting, lighting, then camera specs, and describe the light like you actually saw it instead of leaning on 'cinematic'.
If you want every trick in one place, here is the full breakdown of the prompt order, lighting language, and the ten photography terms.
Worth 10 minutes if you keep getting that flat, generic AI look and want a repeatable way to fix it tonight.
Credits to the original creator.
Poll: Your AI image comes out generic. Whose fault is it?
Hit reply and tell us why.
