In partnership with

“Test more creative” is easy advice to give when you aren't the person who has to make it.

Somebody has to write the brief. Get the product photographed. Make the ad. Ask for another version.

Then do it all over again next week.

I've hired dozens and dozens of graphic designers over the years.

GPT image 2.5 is beating 99% of them.

One hoodie photo gave me a social ad, a catalog shot on a real looking model, and a bathroom mirror selfie that looks like a customer posted it at midnight.

MAKE EVERYTHING: campaign ads, product photos, thumbnails, and event branding

No shoot. no model. No photographer. No sample steamed on a rack at 7am.

So i gave it the kind of work my team actually ships.

All for the same hoodie. All starting with one photo off the product page. I'll show you how each format came out and how to try it on your own product.

I put together in a free ebook called MAKE EVERYTHING: reusable production prompts, plus the event branding workflow I'll show you below. Totally free. Link’s at the bottom.

Sponsored

The agentic era needs a different CRM. That’s Attio.

Teams like Parallel, Turbopuffer, and Wordsmith are already setting the pace on Attio. Get an always-on revenue engine, with agents and workflows that build pipeline, chase every buying signal, and move deals forward with your team. Whether you're working in your browser, inbox, or favorite agent, connect to your customer data in real-time through Attio's web app, MCP, API, and SDK.

one pic in, full campaign out

OpenAI dropped GPT Image 2.5 last week. i tested both versions, (Flare and Sunburst), against GPT Image 2 on over 100 generations.

I picked a Jackie Robinson pullover from Roots of Fight. I wear a lot of their stuff, and my grandpa was a huge Brooklyn Dodgers fan.

Here's the photo i started with.

Freeze the real photo so every asset descends from it. That key to this method. Sure it's faster to let the model invent a hoodie but without the real thing the model gets the details wrong.

So the product photo stays attached to every generation. Here's the bleacher ad prompt with the sleeve direction and the brand named. Copy the structure, Swap the product.

Reproduce the garment from the attached photo: the same heather royal blue vintage-washed fleece, the same white Brooklyn 'B' with the red '42' beneath it on the wearer's LEFT chest, the same white patch on the wearer's LEFT sleeve, the same white drawcords, kangaroo pocket, ribbed cuffs and hem. Use the reference photo as the source of truth. Do not move, resize, recolor or restyle any graphic.

A social ad for Roots of Fight, 1:1, in a vintage American athletic aesthetic: warm film grain, muted period colour, classic block and script sports typography. The hoodie laid flat on weathered wooden bleacher seats in low afternoon sun. Across the top in white vintage block capitals: "HE WORE 42 SO THEY COULD WEAR ANYTHING." Use the attached Roots of Fight logo at bottom right. No gradients or glow.

That's the Sunburst output. The headline printed verbatim. The product stayed recognizable, right down to the stitching.

five winning formats we ship every week

My team uses these five static formats on client accounts. Sme five, different brands. I tried every one with the hoodie. My team runs these five static formats across dozens and dozes of client accounts, Over and over. They’re proven winners. I tested the same five built from the one hoodie photo.

Us vs Them. Clean side, messy side, four lines of copy each.

→ It works because a reader picks a side before reading a word. The hoodie version puts the real Jackie Robinson pullover against a cracked sweatshirt on a plastic hanger.

The radial stat callout. The product in the middle, hand drawn arrows pointing at the proof around it.

→ It works when the facts are the pitch. The hoodie version pulls four real ones off the product page.

The UGC mirror before and after. To phone selfies of the same person in the same bathroom, weeks apart, dates stamped on.

→ It works because it reads like a person instead of a brand.

October 3 in a cracked college hoodie under bad light,

October 31 in the Roots of Fight with a coffee, the old hoodie folded on the sink. four weeks off the college hoodie.

The native iphone notes screenshot. The pitch typed into the notes app and screenshotted, product dropped in beside it.

→ It works because it reads like something you’d forward.

The kitchen whiteboard. A hand drawn before and after on a whiteboard in a real kitchen, product held up next to it.

→ It works because a marker sketch reads as a person thinking out loud.

Five formats, done in twenty minutes. Prompts are in the pack.

If you want to see some more ads for these winning formats…

the catalog shot nobody booked

Same hoodie. This time I asked for a model in a studio.

I didn’t worry about locations, studio time, photographers, casting, no shipping product across the country. In and out. A catalog shoot runs $3,500 to $5,000 minimum by the time the photographer, the studio, the model and the retoucher are all paid.

Real hoodie, real looking person, even commercial lighting. The front view is clean enough for a Shopify product page this afternoon.

A different look becomes another generation. You run these concurrently. At scale. Product and accuracy check still belongs to you. I’d recommend setting up an eval, basically an automated check, that flags any problems before you see them. It’ll catch about 85%.

THE OLD WAY                         NOW
───────────────────────────────   ─────────────────────────
photographer + studio + model       one product photo you already own
retouching pass                     one named prompt
book it, shoot it, wait on edits    15 to 32 seconds per generation
$3,500 to $5,000, one look          same anchor, every look you want

reply CAMPAIGN with your best product photo and my team will run one page's worth of shots off it. free for the first few.

UGC content you never need to source

Then the format that usually needs an actual human with an actual phone.

A bathroom mirror selfie. Bad overhead lighting. Real phone grain.

On TikTok or Instagram most people would go straight past it without even thinking "AI".

And that caption landed closer to my own life than I planned. The model wrote it as ad copy. But that line reads like something I'd actually post.

One photo off a store page. an ad, five weekly formats, a catalog shot, and a customer-style post. that's a campaign's worth of assets

Worth pointing out. Look at the B and the 42. If you took those photo from an iPhone pointing at a real mirror it would look backwards. Easy enough to flip. But fr realistic images you want to mimic the natural world as much as possible.

Otherwise, the image looks real enough that you can miss the thing that's wrong with it.

Hot tip: Use this as a seed image in Gemini Omni 1.1. In ten seconds you’ll have a ready to go UGC video.

run this before anything ships

Realism is done. GPT image 2.5 can make images that pass for photographs. What's left to check is whether the picture is right: the product, the physics, and the words printed on it.

the anchor check. did every angle of the product come from a frozen real photo?

Look at what happens if you don’t tell the model what’s on the BACK of the product. It makes it up. You don’t want that.

the blank check. is every logo, badge, headline and URL named in the prompt? If the model doesn’t know something (like a logo) it might make it up. Again, you don’t want that.

In my testing, without specifying a logo the model can invent one, or worse, pick a logo from a popular brand. You’ll catch this before shipping (yes you will) but it’s a pain and you don’t want that.

See below it made up a a logo for Russel Athletics.

the physical check. reflections, light sources, and the sides of the product nobody photographed.

bonus: it'll build the brand too

One more, because the run didn't stop at ads.

My wife owns a gallery in New Orleans. she's throwing a party with an absinthe brand, and it needed a name, a theme and a look before anyone could say yes to it.

GPT image 2 built three identity concepts. Each came back as a full board. posters, shirts, menus, tickets, even what the room would look like.

Then I handed those boards to gpt image 2.5 and asked for something we could actually present. It came back with a five slide pitch deck.

I packaged everything I used in this run.

you get:

→ the campaign-from-one-photo prompt, the exact bleacher structure above
→ the five weekly format prompts: us vs them split, radial stat callout, ugc mirror before and after, iphone notes screenshot, kitchen whiteboard
→ the catalog shot prompt, studio lighting, real garment, page-ready
→ the casual phone photo prompt for native social
→ three thumbnail prompts, the loud one, the clean tech one, the dramatic one
→ the three-step explainer prompt for product diagrams
→ the same-character-new-camera prompt for consistency across a set
→ the change-one-thing revision prompts that survive a client round
→ the full event brand workflow, brief to three directions to six designs to a pitch deck
→ the negative prompt rules and the before-you-post checks, one page

If you'd rather my team made the campaign, reply campaign with your product URL. we'll pick the reference photo, write the prompts and send back the assets. a handful of slots a week.

The machine can make everything.

Taste decides what gets money behind it.

Go big,

Matt

P.S. the other two pieces are free too: Winning Creative Tracker for the angle and format research, Meta Ads Kit for putting approved creative into the account. research, make, review, run.