You scroll past an AI image that stops you cold. The lighting, the composition, the mood — you want to make something in that exact style. So you open Midjourney, type your best guess, and get… something that's almost right but not quite. Then you tweak for an hour and never close the gap.
The problem isn't your prompt-writing. It's that you're working forward (idea → prompt → image) when you should be working backward (image → prompt). Reverse-engineering an image is a learnable skill, and once you have it, recreating any style becomes mechanical instead of magical.
Here's how to do it.
Step 1: Decompose the image into its layers
Every image, no matter how complex, is a stack of decisions. Train yourself to read them in this order:
Subject — what is it, literally? ("a red fox")
Action / pose — what's it doing? ("curled asleep")
Environment — where? ("snowy pine forest")
Composition — how is it framed? (centered, rule-of-thirds, close-up)
Lighting — direction, hardness, time of day (soft morning backlight)
Color & mood — palette and emotional tone (cool blues, serene)
Style / medium — photo, 3D render, oil painting, anime
Camera / lens — for photoreal work (85mm, shallow depth of field)
Write one phrase per layer. Stack them and you already have 80% of the prompt:
a red fox, curled asleep, in a snowy pine forest, centered composition, soft morning backlight, cool blue palette, photorealistic, 85mm f/1.8, ultra-detailed
That single sentence will get you dramatically closer than "cute fox in snow."
Step 2: Identify the lighting — it's the biggest tell
If you only get good at reading one layer, make it lighting. It's what separates amateur output from professional, and it's the hardest thing to guess.
A glowing edge separating subject from background? That's rim / backlighting.
A soft triangle of light on a portrait subject's cheek? Rembrandt lighting.
Warm, long shadows, low sun? Golden hour.
Visible rays / shafts of light? Volumetric lighting.
Naming the lighting setup explicitly in your prompt is often the single change that takes a result from "meh" to "how did they do that."
Step 3: Figure out which model made it
Different generators have visual fingerprints, and recreating a look is much easier in the model that produced it:
Midjourney — painterly default aesthetic, dramatic lighting, slightly "too perfect" textures.
Stable Diffusion / SDXL — more literal, sometimes rougher edges, highly controllable with weights and negative prompts.
DALL-E 3 — excellent at following natural-language descriptions and rendering text.
Flux — strong at long detailed prompts and reliable text rendering.
Then add the right syntax for that model — Midjourney's --ar 16:9 --stylize 400, or an SD negative prompt like blurry, lowres, extra fingers, watermark.
Step 4: When you don't have an hour — automate the read
Doing this by eye is a great skill to build, but sometimes you just want the prompt. There are tools that analyze an image and output the structured prompt for you — I use a free one called PromptShotAI: you upload the image and it returns the subject, style, lighting, composition, and a model-ready prompt. It's a fast way to check your own read against a machine's, which is also how you get better at doing it manually.
If you'd rather keep a reference handy, this AI prompt cheat sheet lists the per-model syntax (Midjourney parameters, SD weights, DALL-E/Flux phrasing) on one page.
The mindset shift
Stop treating prompting as a creative act of writing and start treating it as an act of observation. The prompt is just a description of decisions someone already made. Get good at reading those decisions — subject, environment, composition, lighting, color, style, camera — and you can recreate (and then remix) any image you admire.
What's a style you've been trying to nail? Drop the reference in the comments and let's decompose it together.
Top comments (0)