About PortraitLab — Who Builds This AI Headshot Pipeline

Our day-to-day work is the pipeline itself: checking uploaded selfies against a face-quality gate, running identity-conditioned generation, merging the original face back into each rendered frame, and delivering the files. That operational experience is the basis for the guidance on this site — we describe the system we run, not a product we resell.

We are not portrait photographers or retouchers, and we do not pretend to be. Where a studio shoot is the better choice — large-format print, art-directed brand campaigns, a shoot that needs wardrobe and styling direction on set — our pages say so. Our claim is narrower: for the places a headshot is actually displayed — LinkedIn, résumés, team pages, clinic directories — a generated set from good selfies is usually indistinguishable at display size and costs a fraction of a studio booking.

If you want the editorial policy in detail — what we cover, how we work, and the correction log — it is published on the editorial team page.

Unpaid headshot tools are funded one of two ways: advertising and upsells, or the data in the upload. The first model pushes watermarks, resolution caps and paywalls into the output; the second turns face photos into inventory. We are not willing to do the second with portraits, and we do not want the first to decide what your headshot looks like.

Charging from the first image keeps the incentives simple. The $9 Sample Pack (5 images) is a real paid order with no watermark, a commercial licence and one free redo; the 40, 60 and 100-image packages use the same pipeline at a lower per-image rate. The refund policy is written down and linked from every page. You can see the exact numbers on pricing and how the pipeline runs on how it works.

• Input gate first. A photo is checked for a single clear face, usable resolution, a head angle within roughly 30 degrees, no sunglasses or heavy shadow, and no second person in frame. Failing photos are rejected before any generation.
• Identity over aesthetics. The set is reviewed for whether the person is recognisable across many frames, not just whether the first frame looks good. A beautiful image that does not resemble the customer is a defect, not a result.
• Consistency within a set. Backdrop tone, crop and light direction should match across images so the set can be used on one page without looking like a collage.
• Redo as the corrective. When a set misses, the response is a full-set regeneration with adjusted identity guidance rather than one patched file — and the written refund policy if the redo still fails.

We do not publish “accuracy percentages”, satisfaction scores or A/B comparisons against other tools. Those numbers would either be invented or measured on a sample too small to mean anything.

Technical claims on this site come from the documentation of the components we actually run, plus the licensing terms of the model weights we use. If a provider changes its pricing or terms, the affected page is updated — the editorial date at the top of each page shows when it was last reviewed.

• fal.ai — flux-pulid endpoint — The identity-conditioned generation endpoint behind the main pipeline, documented as commercial use.
• Replicate — cdingram/face-swap — The face-merge stage; the model card states commercial use is permitted.
• Black Forest Labs — FLUX — The base image model family; licensing differs between the dev weights and hosted commercial use.
• Google MediaPipe — face detection — The documented face-detection component used by the upload quality gate.
• Replicate — pricing and hardware rate card — The per-second GPU rates behind the unit costs quoted in our pricing explanations.

One deliberate exclusion: we do not use InsightFace model weights, because their license restricts the trained models to non-commercial research. Face detection in the upload gate uses a documented Apache-licensed component instead, and identity work runs on endpoints whose model cards permit commercial use. The reasoning is written up on how it works.

• Uploads are used only to render the order they were uploaded for.
• They are never used to train, fine-tune or benchmark a model.
• They are not published, sold or shared with advertisers.
• Originals and outputs are deleted from storage 30 days after delivery, earlier on request.
• We refuse photos of public figures and anyone under 18.

The full detail — including the processors involved — is on the privacy page. Questions or deletion requests go to privacy@headshotsnap.com.

Email editorial@headshotsnap.com with the URL and the specific claim. Substantive corrections are applied to the page and noted in the update log on the editorial team page. Product support, billing and privacy requests have their own channels listed on the contact page.