How to Art-Direct AI Images So They Stop Looking Like AI.

Here's the end result: a product shot or hero image that a client scrolls past without clocking it as AI-made. Not because you hid the tool, but because you directed it like a photographer instead of describing it like a wishlist. Nine steps. Step 4 (killing your junk tokens) is the one that fixes 80% of the "why does this look fake" problem on its own.

This matters more than it did six months ago. Creative Bloq reported this week on a wave of restaurant menus and shop signage getting replaced by uncanny AI-generated "folk art" — smoothed-out, over-lit, structurally wrong in ways nobody can quite name but everybody clocks. That's the AI look. Solopreneurs generating their own marketing images are one lazy prompt away from the same thing. You don't need a bigger budget to avoid it. You need better direction.

What you need

Any modern image model with a text prompt field (Midjourney, Nano Banana, Firefly, whatever you've already got a subscription to), a CFG/guidance scale slider if your tool exposes one, and 20 minutes with a product or scene you actually need an image for.

Step 1: Write the prompt like a photo brief, not a wishlist

Most AI-slop prompts read like a list of nouns someone wants in the frame: "coffee shop, cozy, warm, aesthetic, product photography." That's a mood board with no photographer attached — there's nobody making decisions about angle, light, or focus. Structure it instead the way you'd brief an actual photographer: subject and action first, then lens, then light, then mood — in the first 10-15 words. "A ceramic mug on a wooden counter, shot with an 85mm lens, afternoon light through a side window, steam rising." Notice there's exactly one mood word in there, and it isn't "cozy" — the light and steam do that work instead. Common mistake: burying the actual subject in the middle of a paragraph of adjectives. The model weights early tokens more heavily. Put the thing you care about first.

Step 2: Name a real light source, not a vibe

This is the single biggest lever you have. AI images look fake most often because the lighting has no source — it's just uniformly bright, like the whole scene is lit from inside a softbox the size of the sky. Real photos have light coming from somewhere. Say where. "Afternoon light through a window, slight underexposure on the left side," or "rim light from a doorway behind the subject," or "practical lights visible in the background, warm and slightly uneven." Directional, imperfect light is the fastest way to make a flat render feel like it was actually shot.

Step 3: Add one or two imperfections, no more than three

Real cameras, real film, and real skin all have texture. AI defaults to smoothness because smooth is the statistical average of "polished" in the training data. Fight it directly: add film grain, natural skin texture with visible pores, slightly messy hair, or visible fabric weave to your prompt. Pick one or two, not five. Overload the prompt with imperfection words and you get a different kind of fake — deliberately gritty in a way that reads as try-hard. One texture cue plus one light cue usually does more than either alone.

Step 4: Delete every junk token in your prompt

Go back through whatever prompt you'd normally type and look for "masterpiece," "highly detailed," "8k," "professional photography," "award winning." These used to nudge older models toward quality. In 2026 models they're increasingly dead weight — noise the model has learned to mostly ignore, or worse, weight toward the over-processed "AI showcase" look those phrases got trained on. Replace each junk token with an actual detail. Instead of "highly detailed," describe what's detailed: "individual thread texture visible on the sleeve" beats "highly detailed clothing" every time. This one swap is the highest-leverage fix in this whole tutorial — if you only do one step, do this one.

Step 5: Lower your guidance/CFG scale if your tool exposes it

If your model has a CFG or guidance scale slider (Midjourney's --stylize, Stable-Diffusion-based tools' CFG, etc.), a high value forces the model to match your prompt aggressively — and aggressive matching is what produces that deep-fried, over-saturated, over-sharpened "AI" texture. Drop it from a default around 7 down to 3.5-4. You'll lose some literal prompt-adherence, but you gain naturalism — softer contrast, less waxy skin, less oversaturated color. If your tool hides this setting, look for "creativity" or "variation" sliders; they usually control the same thing.

Step 6: Use negative prompting as a real tool, not an afterthought

If your model supports a negative prompt field, use it deliberately: blurry, low quality, watermark, extra fingers, plastic skin, oversaturated, deep fried. This isn't a copy-paste catch-all — tailor it to what your specific model tends to get wrong. Run five test generations of the same prompt with no negative prompt, note the recurring artifact, then add that specific thing. Pro tip: keep a running negative-prompt list per model. Every model has its own tells, and yours will start repeating after ten or so generations.

Step 7: Generate in batches of four and compare, don't accept the first result

Every serious model gives you multiple variations per generation. Treat this like a contact sheet, not a vending machine. Generate four, lay them side by side, and look specifically for: does the shadow direction match the light source you described? Does the surface material look plausible for what's sitting on it? Is there one variation where a background detail has gone slightly wrong? Pick the most physically consistent one, not the prettiest one. A gorgeous image with light coming from two directions at once will read as fake the moment someone glances at it twice.

Step 8: Composite in one deliberate flaw by hand

Even a well-directed AI image can feel too clean once it's finished. Bring it into Photoshop, Affinity, or whatever you've got, and add one manual imperfection: a slight vignette that isn't perfectly centered, a tiny dust speck, a barely-there chromatic aberration at a high-contrast edge. This is a five-second move that breaks the last bit of algorithmic perfection your eye doesn't consciously register but somehow always catches.

Step 9: Zoom to 200% and check the tells before you ship it

Before anything goes out under your name, zoom into three places: hands, text, and repeating patterns (tile grout, brick, fabric weave). These are still where models slip. A sixth finger, melted signage text, or a repeating pattern that doesn't actually repeat correctly will undo eight good steps in one screenshot. Fix it with a targeted inpaint on just that region — don't regenerate the whole image and lose the composition you already nailed.

If you get stuck

If your images still read as "AI" after all nine steps, the usual culprit is stacking too many style adjectives back into the prompt out of habit. Re-read your prompt and count the mood words versus the physical/technical words. If mood words outnumber technical ones, you've drifted back into wishlist territory. Cut in half and try again. And if it's specifically hands or signage that keep breaking, that's not bad luck, that's the model. Both are still weak points across most 2026 releases because both require the model to track many small interdependent shapes at once (finger count, letterforms) instead of one blended texture. Don't fight the model on those two; inpaint them by hand every time and move on.

Next Monday: how to batch-export product variants for different social formats using Photoshop Actions — one file, twelve outputs, zero repetitive clicking.