likakandel.comLocation
Augsburg, Germany
Project
Prospects see themselves in the studio before they call
An AI rates video ads from the Meta Ad Library. Where a video has room to improve, it places the person in a studio set that fits, with the face kept pixelated.
The starting point
Studio time is hard to sell while people cannot picture it. Many companies shoot their video ads on a phone in the office and have no idea how much better they could look. A price list does not show them. And going through every ad by hand to pick the right set takes time, even for ads that already look good.
Why I built this
I was given the task of showing companies with weak video ads how they would look in a studio. The Meta Ad Library is full of ads shot against a bare wall in poor light. Telling a company that their video could look better convinces no one. A picture of themselves sitting in a studio set does. Building that picture by hand every time was not an option. So I built a system that spots weak ads, pixelates the face and places the person in a set that fits.
The solution
The prospect does not get a quote but a picture of themselves in a better studio. The system works in two ways. On a routine basis, it takes video ads from the Meta Ad Library, rates one frame and filters out everything that already looks professional. Only when the score is weak does it suggest matching sets. In the internal tool, you upload a photo instead, mark the people and pick a set. The preview then switches between sets for single shots and group shots. Either way, the face is pixelated first, then Gemini places the person into the set. Most of the work went into the prompt: the person has to look as if they were really sitting there, and the face still has to stay pixelated.
How it works
Drag or use the arrow keys to move the steps
- Step 1: Screenshot from the ad leads to Image analysis
- Step 2: Image analysis leads to Continue only if needed
- Step 3: Continue only if needed leads to Studio match
- Step 4: Studio match leads to Blur the face
- Step 5: Blur the face leads to Place the person in the studio
- Step 6: Continue only if needed leads to Next prospect (good enough)
- Step 7: Internal tool leads to Blur the face (own photo)
Steps as text
Step 1: Screenshot from the ad
A frame from a video ad in the Meta Ad Library is captured as a screenshot. A single picture is enough to judge the video.
Step 2: Image analysis
An AI rates the quality of the video: shooting style, lighting and how the background is set up. It gives a score and names exactly what a studio would do better.
Step 3: Continue only if needed
If the score is weak, the video has room to improve and the process continues. With a good score, it already looks professional. This prospect does not need a studio, so the system moves on to the next one.
Step 4: Studio match
The result of the analysis is matched against the studio sets: bright or dark, sofa or table, colour accents such as blue, green or orange. The outcome is a short list of possible setups in the style of the ad, each with its reasoning.
Step 5: Blur the face
A program detects the face and pixelates it before the image goes to an AI model. That keeps the person unrecognisable in the final picture too. In the internal tool this is switched on by default.
Step 6: Place the person in the studio
An image model from Google (Gemini) places the person into the chosen set. Clothing, gaze and gesture stay as they were, the lighting adapts to the set, and the face stays pixelated. With several people, you get a group shot plus a single portrait of each person.
Tools
- OpenAI
- Gemini
- Python
- OpenCV
A similar workflow for your business?
In a short call, we'll work out which of your steps can be automated.