AI image generation for ads

The model is not the skill. The brief is. Today you learn to direct an image generator the way a creative director runs a shoot — so the output is on-brand, ad-ready, and built to feed the loop, not AI slop.

1The model is a commodity; the brief is the craft

Yesterday (Day 12) you turned on Meta's free, in-platform enhancements — background generation, image expansion, touch-ups — that reformat and remix the assets you upload. Today you make the assets themselves from a blank canvas. That is a different job and a different risk profile, so it gets its own day.

Here is the trap to disarm first. Founders read about "the best AI image model" and treat tool selection as the decision. It barely matters. You mapped the stack on Day 11 — the one-line refresher for images is a clear division of labour, and you'll use several in one pipeline:

• Midjourney — the highest-aesthetic hero shot.
• The current FLUX model — photoreal product and lifestyle.
• Ideogram or Recraft's latest release — when there's heavy text or a logo.
• ByteDance Seedream — cheap bulk variants.
• Google's current Gemini image model — the mid-2026 default for edits, variants, and locking brand consistency across a series.

Pick by the job in front of you, not by the leaderboard. Names will change — the division of labour (aesthetic hero / photoreal / text rendering / cheap bulk / consistent edits) is the durable map.

The reason this is liberating: it means the durable skill is not "which model." It's how you brief. Two operators with identical access to the same model produce wildly different output — one ships generic stock-looking sludge, the other ships an ad that converts — because one wrote a brief and one wrote a wish. Every generic image you've ever scrolled past was a one-line prompt. Every on-brand one was a structured brief.

And remember why we're here. Back on Day 1 we established that creative is the last lever you own and that it fatigues — every winner decays as it scales. The point of generating images with AI is not to make one pretty picture; it's to feed the volume the loop needs (Day 5's explore/exploit, Day 10's matrix) without a studio budget per asset. Cheap volume only compounds if it's on-brand and tagged, so the brief is where quality and learnability both get protected.

2The six inputs that make an image on-brand

A wish is "a serum bottle on a marble counter, sunlight." A brief specifies six inputs the model cannot guess. Miss any one and the model fills the gap with its training-data average — which is exactly what "AI slop" is: the statistical mean of every image like yours that ever existed. Bland by construction.

The six inputs:

• Subject & product fidelity — what exactly is in frame, and is it the real product. This is the non-negotiable one (Section 3).
• Style reference — the look: a brand palette, a mood, 2–4 reference images of past on-brand creative. Midjourney's --sref and Style Weight, or feeding current Gemini image model your logo, brand colours and prior visuals, exist precisely to pin this down.
• Composition & format — framing, where the product sits, negative space for copy, and the aspect ratio you'll ship into (9:16 Reels, 4:5 Feed, 1:1 — your Day 10 formats).
• Lighting & treatment — and here you make a deliberate Day 9 choice: hi-fi (studio, polished, controlled light) or lo-fi (phone-shot, on-a-kitchen-counter, "a friend posted this"). AI does both; the brief decides which.
• Text-in-image — if there's a headline or badge baked into the pixels, you specify the exact words. In 2026 the text-rendering specialists (Ideogram, Recraft) and current Gemini image model render legible, correct text; older models still produce garbled glyphs. Route text-heavy work to a model that can spell.
• Persona & angle cues — the human the image implies (Day 7) and the message it carries (Day 8). A "budget-conscious parent, before/after relief" frame looks nothing like a "status-seeking professional, aspirational" frame, even for the same product.

A worked example. Say a skincare brand wants Feed creative for one concept across two personas. The wish — "serum on marble, sunlight" — gives you a forgettable stock photo on the first try and three near-identical re-rolls after. The brief gives you a generation matrix. Hold concept, product and palette fixed; vary persona, treatment and format: 2 personas × 2 treatments (hi-fi / lo-fi) × 2 ratios (4:5 / 9:16) = 8 tagged frames from one structured brief in an afternoon. At a studio shoot that's a half-day and four figures; here it's a few euros of credits. That is the leverage — but only because every frame is a deliberate, labelled point in the matrix, not eight rolls of the same dice.

3Product fidelity: composite, don't hallucinate

Here is the single most expensive mistake in AI ad imagery, and the one rule that separates amateur output from professional: never let the model invent your product.

If you describe your product in words and let the generator draw it, you will get a bottle, a sneaker, a dashboard — plausible, beautiful, and wrong. The label is gibberish, the logo is a melted approximation, the cap is the wrong shape, the device has six buttons instead of four. To you it's obviously not your product. To a customer who clicks through and sees the real thing, it's a bait-and-switch — and trust, the thing your whole funnel runs on, takes the hit. For regulated or premium categories it's worse: a hallucinated product can imply a feature you don't sell.

The fix is compositing. Generate the scene with AI — the marble, the light, the lifestyle context, the model's hands — and place your actual product photograph into it. Modern editors make this a single step: you inpaint the real product onto an AI-generated background (inpainting = regenerating only a selected area, so the rest of the image is untouched; the mask is that selection), or mask it in and let current Gemini image model colour-match the lighting so it sits naturally. The AI does what it's genuinely good at (cheap, infinite, plausible environments); the camera does what it must (an honest record of the thing you'll ship). Same logic protects faces and any text the customer must trust — generate the world, anchor the truth.

This same discipline scales hi-fi and lo-fi (Day 9) equally. A hi-fi composite drops the real bottle into a studio-lit AI set. A lo-fi composite drops the same real bottle into an AI-generated "messy bathroom shelf, phone flash, slightly off-centre" — and now you have a native-looking UGC still that didn't need a creator, a kitchen, or a half-day. Both are on-brand because in both the product is real.