How to Prompt Pix-Vu: Get the Best AI Ad Creative Every Time

Pix-Vu generates Facebook and Instagram ad creative from a few short fields: a product description, an audience description, a daily budget and one or more product photos. The output you get back — headlines, primary text, and full ad images — is only as sharp as the input you give it. This guide walks through exactly what to type into each field so you stop wrestling with regenerations and start launching campaigns you'd actually run with your own money.

What's Actually Running Under the Hood

Before we get into the prompting itself, it helps to know which models are doing the work, because each one rewards a different kind of input.

Ad copy — OpenAI GPT-5. Every headline, primary text, and description Pix-Vu produces comes from GPT-5 with a custom system prompt trained on $100M+ in winning Facebook ad spend. It's instructed to use proven direct-response frameworks (PAS, AIDA, Before/After) and to keep headlines under 40 characters, primary text between 125–250 characters, and descriptions under 30. The model handles tone, language, and CTA logic — your job is to feed it a clear product and a clear customer.

Ad images — Google Nano Banana Pro (Gemini 3 Pro Image Preview). This is the top-of-leaderboard image-editing model in 2026. It scores 94% accuracy on text and label rendering and is specifically tuned for product preservation — meaning it can take your real product photo and drop it into a new scene without distorting the bottle, label, or branding. If Pro is unavailable on the call, Pix-Vu automatically falls back to Gemini 2.5 Flash Image (Nano Banana Flash), which is still excellent.

Why this matters for prompting: Pix-Vu intentionally has no text-to-image path. The AI must edit one of your real product photos. That single design choice is what separates "AI ads that look like AI ads" from creative your customers actually click. Garbage photo in → garbage ad out. A clean reference photo and a sharp visual hook → a Facebook ad that earns its CPM.

The Four Fields That Decide Your Output Quality

Inside the create campaign flow you'll be asked for four things that drive the AI:

1. Product description
2. Target audience
3. Daily budget
4. Product photos

Get these right and Pix-Vu's first draft is usually launch-ready. Get them wrong and you'll burn regenerations chasing your tail. Here's how to nail each one.

1. Product Description: Tell the AI What It's Selling

This is the single most leveraged field in the whole tool. GPT-5 uses it to choose the angle, the pain point, the framework, and the CTA. Vague descriptions produce vague ads.

Bad: Skincare

Better: Organic skincare

Best: Sensitive-skin face serum made with British-grown sea buckthorn. Fragrance-free, vegan, dermatologist-tested. Sold direct-to-consumer at £29 for a 30ml bottle. Customers buy it because every other "natural" serum on the market still triggers their rosacea.

Notice what the best version includes:

• What it physically is (face serum, 30ml)
• The hero ingredient or differentiator (British sea buckthorn)
• Concrete proof points (fragrance-free, vegan, derm-tested)
• The price (anchors the AI's tone — premium vs. value language)
• The pain it solves (rosacea triggers from competitor "natural" products)

You don't need a full marketing brief. Three to five sentences is the sweet spot. If you can't articulate the pain point in one sentence, write it down first — the AI cannot invent emotional truth your product doesn't have.

Things that hurt your output:

• One-word descriptions ("supplement", "course", "app")
• Pasting your About Us page (too vague, too corporate)
• Listing every feature with no story
• Talking about your company instead of your product
• Hype language like "revolutionary" or "game-changing" — GPT-5 is explicitly told to avoid these and you're just wasting tokens

2. Target Audience: Tell the AI Who It's Talking To

GPT-5 changes vocabulary, references, and emotional register based on this field. A "30-year-old triathlete in Manchester" gets a very different ad than "menopausal women in Texas dealing with hot flashes" — even if the product is the same hydration powder.

Bad: Everyone

Better: Women 25-45

Best: Women 30-50 in the UK who have just been diagnosed as perimenopausal, are skeptical of HRT, and have already tried 2-3 supplement brands that did nothing. They read Stylist, listen to The Diary of a CEO, and search "natural perimenopause symptoms" at 2am.

The pattern:

• Demographics (age, gender, location)
• Life stage or trigger event (just diagnosed, just had a baby, just signed a lease)
• Prior failed solutions ("already tried X and it didn't work")
• Cultural shorthand (publications, podcasts, search behaviour)

Cultural shorthand is the secret weapon. When you mention publications, podcasts, or behaviours, GPT-5 borrows the register of those communities — and your ad suddenly sounds like a friend, not a brand.

3. Visual Hook (in chat-driven creative): Tell the AI How to Show It

When you're using the dashboard chat to iterate creative, Pix-Vu lets you pass a "hook" — a one-line visual concept that becomes the scene around your product. Hooks are where the magic happens with Nano Banana Pro.

The model is excellent at scene transformation but it only renders what you describe. If you write nothing, you'll get a default studio shot. That's fine. But specific hooks are how you generate scroll-stopping creative.

Weak hooks:

• On a table
• Outside
• With flowers

Strong hooks:

• Held in a runner's hand at a misty London park bench at 6am, golden hour light, Strava-on-wrist composition
• Floating above a marble bathroom counter with soft morning light streaming through linen curtains, water droplets on the bottle
• Sitting on the dashboard of a beat-up VW campervan parked at a Cornish surf beach, neoprene and salt-air vibes

What makes them strong:

• A specific human moment (not just an object)
• Light direction (golden hour, soft morning, harsh midday)
• Texture and material cues (marble, linen, neoprene, salt)
• A narrative that matches your audience — the runner hook is for athletes, the bathroom hook is for self-care customers, the VW hook is for outdoor lifestyle

A good rule: if you can imagine a photographer shooting it in 30 seconds, the AI can render it. If your hook would confuse a photographer, it'll confuse the model too.

4. Product Photos: Feed the AI Real Pixels

Pix-Vu requires at least one real product photo. The image model edits this photo into new scenes — it does not invent your product from thin air. This is deliberate. Generic AI products don't match real packaging, and your customers will notice.

Photo quality directly determines creative quality. Spend 10 minutes getting this right and every campaign downstream improves.

The minimum bar:

• Sharp focus — phone cameras are fine, blurry is not
• Even lighting — soft daylight from a window beats any ring light
• Plain background — white wall, white sheet, white paper. The cleaner the cutout, the easier for the model to drop it into new scenes
• Product fills 60-80% of the frame — too small and you lose detail, too large and the model has nothing to work with at the edges
• Label-forward angle — the model can render text well but it can only render text it can see in the source

Pro tips:

• Upload 3-5 angles of the same product. Pix-Vu uses different shots for different ad formats (square, story, landscape) and the variety lets the model pick the best base for each scene.
• If you have lifestyle shots and plain product shots, upload both. Plain shots are better for hero scenes; lifestyle shots are better when the model just needs to recolour or relight.
• Avoid heavy filters and Instagram presets on the source photo. Filters bake in colour shifts the model then has to fight against.
• Don't upload screenshots of your website or product pages. Upload the original image files.

A Worked Example: From Empty Form to Launch-Ready Ad

Let's run a real product through the whole flow so you can see how the fields combine.

The product: A new sleep-aid tea blend from a small UK herbalist.

Lazy version of the form (what most people type):

• Product description: Sleep tea
• Target audience: Adults
• Hook: On a bedside table
• Photo: A blurry phone shot of the box

GPT-5 produces something like: "Try our sleep tea tonight. Made with natural ingredients. Order now." Forgettable. The image model produces a generic tea box on a generic nightstand.

Sharp version of the same form:

• Product description: Loose-leaf sleep blend made with Suffolk-grown valerian, chamomile and a hint of lavender. Caffeine-free, organic, brewed loose in a teapot for 4 minutes. Sold at £18 for 50 cups. Customers buy it when over-the-counter sleep aids stop working but they don't want to take prescription pills.
• Target audience: Knowledge workers age 32-50 who fall asleep fine but wake up at 3am with their brain racing about work. They've already tried magnesium, melatonin, and a meditation app. They read The Cut and listen to Huberman.
• Hook: Steam rising from a hand-thrown ceramic mug on a linen-covered nightstand, lit by a single bedside lamp with a soft amber glow, paperback book and reading glasses just visible in the foreground
• Photo: 4 sharp daylight shots of the tin from front, side, top, and one with a few loose leaves spilled next to it

GPT-5 now writes: "Wake up at 3am with your brain on fire? This Suffolk valerian blend is the one Huberman fans pour when melatonin stops working." The image model returns a moody, premium, scroll-stopping ad that looks like it was shot for Kinfolk magazine.

Same product. Same five minutes of work. Wildly different output.

Five Mistakes That Kill Pix-Vu Output

1. Fragmenting your description across regenerations. Don't add detail after each AI draft. Write the full description once, then iterate the output, not the input. GPT-5 is best on the first pass when it has all the context up front.

1. Asking for tone you can't define. "Make it more fun" tells the AI nothing. "Make it sound like a friend texting from the gym" gives it a register. Be specific about voice or leave it alone.

1. Uploading product mockups instead of real photos. Render mockups and 3D renders look uncanny when re-edited. The AI will warp the label or invent fake reflections. Use real product photos.

1. Targeting "everyone interested in health". Broad audience descriptions produce broad ads, which produce poor CTRs, which Meta then punishes with higher CPMs. Narrow audiences are cheaper to advertise to and easier for the AI to write for.

1. Skipping the photo upload because "AI will make it". Pix-Vu will refuse to generate without a reference photo, by design. Generic AI product images don't match real packaging and customers notice. The 10 minutes you spend taking 4 clean shots is the highest-leverage 10 minutes in the whole workflow.

Iterating Inside the Dashboard Chat

Once your campaign is created, the dashboard chat is where you refine. You can ask for variations, tighter copy, different scenes, or new hooks — and because the model has the campaign context, you don't need to re-paste everything.

Useful chat prompts:

• "Generate three more headline variants in a more sceptical, deadpan tone."
• "Re-do the hero image with the same product but in a Brooklyn rooftop scene at sunset."
• "Write a primary text version that opens with a stat instead of a question."
• "Give me a Story-format (9:16) version of the current creative for Reels placement."
• "Rewrite for an audience that's already heard of us — assume warm traffic."

Treat the chat like an art director, not a search box. The more direction you give, the better it performs.

The Short Version

If you only remember five things from this guide, make it these:

1. Write a 3-5 sentence product description that includes what it physically is, the hero differentiator, the price, and the pain it solves.
2. Target a specific tribe, not a demographic. Mention publications, behaviours, prior failed solutions.
3. Write photographic visual hooks with light, texture and a human moment.
4. Upload 3-5 sharp, real product photos on plain backgrounds. Phones are fine. Mockups are not.
5. Iterate in chat, not in the create form. First-pass GPT-5 output is best when you give it everything upfront.

Do those five things and Pix-Vu will hand you ads that look hand-crafted by a £200/hr direct-response copywriter and a £400/day product photographer — except you got them in 90 seconds for the cost of a single AI call.

Ready to Try It?

If you've read this far, you're already in the top 10% of users in terms of how seriously you take prompting — which means your campaigns are about to outperform almost everyone else's on the platform. Open the create campaign flow, paste in a sharp product description and a real photo, and watch the difference.

Launch your first campaign with Pix-Vu →