Pixel: prettier is not always truer

A deliberately fast beginning

The original ask was broad: experiment with a new image model and see whether marketplace car photos could be improved. The constraint was equally important—move quickly, learn like a hackathon, and avoid building a custom generative model before knowing what the product needed.

I began with two parallel audits: what was wrong with real listing photos, and what edits the model could reliably perform. Removing people or background clutter was often useful. Completing a heavily cropped car, changing damage, flipping orientation or inventing a cinematic background could make the image attractive and untruthful.

This is a public-safe account of the work. Internal identifiers, proprietary thresholds, prompts, infrastructure details and operational commands are intentionally omitted.

Defining the allowed lie

Stakeholder sessions turned ‘enhance the image’ into a narrower contract: improve presentation, remove obvious distractions and correct mild alignment while preserving the actual car. Tiny completion near an image boundary could be acceptable; reconstructing half a vehicle was not.

Prompt iteration taught a practical lesson: longer instructions did not necessarily create better compliance. Short, photographic language plus an explicit identity constraint worked better than a catalogue of every possible failure. A newer, more expensive model variant did not justify its cost in this use case.

Manual review before automation

For the first city experiment, the team reviewed roughly two thousand candidates across about five hundred cars. The work was slow, but it produced the failure taxonomy the automated evaluator needed: missing cars, synthetic-looking backgrounds, geometry changes, erased interiors, blurred components and logos, orientation flips and poor blending.

The first experiment improved CTR for the enhanced cohort from 5.25 to 5.93, while the comparison cohort moved from 6.92 to 5.35 over the measured windows. Because cars rather than users were assigned, I treat this as directional marketplace evidence rather than a perfectly isolated user-level experiment.

Turning taste into gates

Manual review could not scale. The next phase compared object detections, vehicle area, position and confidence before and after enhancement. These checks caught edits that lost parts, changed the crop too aggressively or moved the vehicle away from the intended composition.

A CLIP-based aesthetic predictor supplied a separate presentation score. Object consistency asked ‘is this still the same car?’ while aesthetic scoring asked ‘is this image better composed?’ Neither score was trusted to answer both questions.

A second city and a service boundary

In a second city, the enhanced cohort's CTR moved from 4.75 to 4.87 while average rank improved from 4.94 to 4.37; comparison cars moved differently over the same period. The smaller effect reinforced the need for multi-market evidence rather than celebrating one large number.

The production DS component became a focused API that scores aesthetics and stores an audit record. Generation, image delivery and other operational responsibilities remained with the owning product system. That boundary kept the data-science service narrow and made model ownership explicit.