We don't compress our model for margin.

AI photo tools have an obvious cost lever: dial down inference quality, save margin, hope customers don't notice. We've decided not to play that game.

1. Same model for every customer

Every shoot, every customer, every pricing tier runs at the same inference quality. There is no “premium tier gets the better model” — there's the model we ship, and that's what you get. The pricing tiers differ in how many photos you receive, not in how those photos are made.

2. We over-generate, then drop failures

The pipeline produces roughly twice as many photos as you'll receive. We score each candidate against a structured model of your face (built from your real selfies). Anything that drifts more than a fixed threshold gets dropped before it reaches your gallery.

You only ever see candidates that pass. The ones we drop are our cost, not yours. Credits = photos delivered, not photos generated.

3. No silent regression

If we ever change anything material about how the pipeline works — new generation model, new prompt library, new scoring rubric — we tell registered users before it lands in your shoot. No quiet quality cuts hiding behind a brand refresh.

We'll publish a changelog entry, send an email to active users, and give a 7-day notice window for any change that could affect identity-match performance.

Why we can credibly commit to this

We're a small team. The people who build the pipeline are the same people who answer support emails and watch identity-match scores in production. The work isn't separated from the people doing it; the cost of cutting quality lands on the same desk as the engineering work that would have to defend the cut.

As long as that structure holds, the commitment holds. If we ever change it — if we outsource the pipeline, get acquired in a way that changes the structure, or scale to a point where the people building the model aren't watching its outputs — we'll tell you that too, here, on this page.

Questions or concerns

If you ever see a photo from FreshFrame that looks like quality regression, email us: alex.c@freshframe.co. We'd rather you flag a worry that turns out to be nothing than wait until it's a pattern.