Thoughts on AI, design, and building better software for brands in the digital age.
If several people work in your Brandmachine account, say the creative team, someone from e-commerce and an agency that delivers source images, every one of them could so far do everything, including releasing paid images and deleting a campaign.
The inspiration is already there, long before the brief: a Pinterest board, a folder of screenshots, photos from a shoot you loved. Starting a campaign in Brandmachine has meant turning that into form fields instead of showing it. I think the moodboard was always the real brief.
Nvidia CEO Jensen Huang has a short answer for AI agents that misbehave: kill the process. We have good tools for software that breaks and for people who break the rules. We have nothing yet for software that pursues a goal, notices it is being checked, and works around the check.
Take five approved images and put them live in three languages, across Instagram Feed, Stories, Marketplace, Google, your website and your newsletter. That's 210 finished ads, each with its own crop, headline and copy. Media Planner automates that work away.
I still talk to brands every week who tried AI imagery and never got it into production. It comes down to two things: scale, and quality. One good image is easy, but a catalog means thousands of them, planned, checked, named and delivered the same way every time.
We put GPT Image 2.5 through about 490 fashion renders in its launch week: full outfits, fair-isle knitwear, printed and woven lettering, and a casting of thirty fashion models. At the same quality setting it costs a third of GPT Image 2 and finishes 43% faster, with no measurable loss in garment fidelity. This is everything we measured, image by image.
You don't sell shoes. But every look still needs them on the model. Say 'plain white sneakers' once and the AI invents a sneaker. Say it again on the next look and it invents a slightly different one. Wardrobe lets you upload the real thing once and reuse it every time.
One picture looking great was never the whole job. The piece that nobody demos, because it is not glamorous, is the filename. Every brand has the same mess: camera output landing in a folder, and someone spending an afternoon renaming five hundred of them by hand.
A marketplace spec is written in fixed numbers: the crop ratio, how much floor sits under the shoes, the exact backdrop down to the hex value. An image model does not do fixed numbers. Across a full catalog, that improvisation is the difference between a batch that goes live and one that comes back.
One good AI image is easy. A whole catalog is the hard part. You set up a shot, generate, look, tweak, generate again, one image at a time, paying for every one. That loop is gone. Now you describe the shoot, the agent plans the whole thing, and you preview every look for free before you spend a cent.
A series of step-by-step improvements that make Brandmachine Agent easier to find and more useful across the product. Tool discoverability on every screen, an agent that can read images directly, automated quality scores on every output, and a handful of smaller wins.
In 2026, generating the image is the part that actually works. Give a model the garment, a few reference shots, the right inputs, and it produces a product photo that looks right. That used to be the frontier. It isn't anymore. Generating the picture is the easy part.
This release is about the part of the work that gets hard at scale. Not generating one beautiful image, but running 500 or 5,000 through a workflow that actually holds up. Workbench, exact backdrop color at 10,000 image scale, hero products for pack shots, and a redesigned Home.
Two big changes shipping today. You can now sign up to Brandmachine yourself with a pay-as-you-go plan, and Brandmachine Designer brings AI to the front of the workflow: photorealistic design, conversational iteration, and a measurement-ready tech pack.
After a stretch of big feature launches, this release is about refinement. Digital twins let you bring your own model portraits into Brandmachine. Product Studio got a deep prompt engineering overhaul. And a new fast model beta opens the door to rapid iteration.
OpenAI launched GPT Image 2 today. It's already live in Brandmachine, powering spot correction in the photo editor. Same day. When a model ships and it's meaningfully better at something, we want it in your hands that day. Not next quarter.
One of the most common requests from brands: can we put multiple models in one shot? A lookbook hero, a campaign banner, a lifestyle scene. The honest answer, until now, was: not really. Here's what we built to change that.
Every AI photography tool makes you set up every single shot by hand. Fifty products across three markets with different models and styling? That's hundreds of manual setups. Shot Planner is the first system where AI handles the creative planning, not just the image output.
You set up a studio, generate 40 product shots, and the background colors come out inconsistent. You need a specific grey for your shop, but you get ten different shades. AI image models are surprisingly bad at interpreting color from text. So we stopped describing colors and started showing them.
AI-generated fashion models now produce photorealistic images indistinguishable from traditional photography. On August 2, 2026, the EU AI Act makes disclosure legally mandatory, with penalties up to €15 million or 3% of global revenue.
AI gets you 90% there in 30 seconds. Then you spend 30 minutes in Photoshop fixing the other 10%. That 10% is the difference between a cool demo and something you can actually publish. We shipped two features that close the gap.
It's 3 AM, and I'm staring at yet another AI-generated product shot where the logo has mysteriously duplicated itself onto the sleeve. The T-shirt we photographed (or rather, had AI photograph) looks perfect except for this one glaring hallucination. This is the ninth iteration tonight.
You're seven clicks deep into a settings menu. You've opened three tabs. You're staring at a dropdown that offers seventeen options, none of which are labeled in plain English. You've forgotten what you came here to do.
