When I first ran paid ads for my agency back in 2021, every still image creative had to either be built by me in Canva or sent to a graphic designer at $20 to $50 a pop. Testing a batch of, say, 22 to 24 creatives could cost you close to $1,500 before you drop a penny on ads. Right now I'm spending $900 a day on my low ticket funnel and some of my best performing creatives are still images that came out of one workflow. A few agency owners we've shown it to have pulled their cheapest cost per leads ever from the same process. The best part is that the ads were created entirely by AI and cost maybe 10 to 20 cents in credits each per creative. The actual tools are simple. Claude writes the image prompt, Higgsfield produces the image. The part that makes it work is what you do before you prompt anything. Most people open an AI tool and type "give me 12 ad concepts for a roofing lead gen offer" and let it rip. What comes back is generic because the strategy was generic. The image is only ever as good as the thinking you feed it. So instead I steal formats that are already proven. I go into the ads library of the biggest B2B advertisers I can find, sort by impressions, and look for still image formats they keep running. If a format shows up over and over near the top, it's working. From doing this for a client last week across four niches (roofing, tree services, siding, tint and PPF) I landed on four formats: A phone mockup where the notifications on the screen are the client's dream outcome. For a roofer that's a $40,000 deposit hitting the bank and a text saying "new roof restoration quote." A fake news article layout with the client avatar in their work environment and a headline that matches the hooks from the video scripts. A whiteboard sketch comparing the old way against the new way, with a little diorama of a finished job sitting in front of it. Bold text on a red background. Stupidly simple and it's one of Cole Gordon's top performers. Four formats, four niches, sixteen ads. Each format is deliberately nothing like the others, and that matters. If your first batch is four colour variations of the same layout, you're not testing anything.