How Our AI Headshot Generator Works
Most “AI headshot” products are a single diffusion call with a prompt like “professional headshot, studio lighting”. That works for a demo and falls apart for a forty-image set, because a single call has no way to stay faithful to a specific face across forty different poses. Our pipeline is deliberately more than one call.
The order is: quality gate, preset selection, identity-conditioned generation, face merge pass, normalisation, delivery and deletion. Six stages, two model calls per image, and a fixed price that does not change based on what you upload.
- Input required
- 3–8 selfies
- Model calls
- 2 per image
- Delivery
- 1024–2048px JPG
- Retention
- 30 days
Pipeline
The six stages, in order
Each stage exists because of a specific failure mode observed in single-call tools.
- 1
Upload and quality gate
You upload three to eight selfies. An on-server check rejects photos with no face, more than one face, heavy blur, extreme head angles, sunglasses or too little resolution, and tells you exactly why so you can retake it.
Nothing is generated for a photo that fails the gate, so bad inputs never cost you a redo.
- 2
Style preset selection
The preset fixes backdrop tone, wardrobe level, lighting direction and crop. Five presets cover corporate, creative, actor, medical and real-estate use.
The preset can be changed on the free redo without paying again.
- 3
Identity-conditioned generation
The pipeline renders new portraits from your identity and the preset. Each image is a new frame — new pose, new light, new expression — not an edit of a single photo.
Typical cost is a fraction of a cent to a few cents per image depending on resolution.
- 1
Face merge pass
A second pass compares the generated face with your original photos and pulls your facial features back into the output. This is what keeps a forty-image set recognisable instead of drifting into a generic person.
This stage is why the set holds up at 20 images, not just the first one.
- 2
Normalisation and export
Backgrounds are flattened to the preset tone, crops aligned, and files exported as 1024px JPG — or the enhanced resolution set on the top tier.
Per-image cost and generation time are fixed, so package price does not change with what you upload.
- 3
Delivery and deletion
You get a download page with individual files and a ZIP. Order files stay available for 30 days, and the original uploads are deleted on the same schedule — earlier on request.
Redo requests are accepted for 48 hours after delivery.
Deep dive
Why two model calls instead of one
Identity preservation and photographic quality are different problems, and one model is rarely best at both.
Call one: identity-conditioned generation
The first call takes your source photos, builds an identity embedding and generates a new portrait for the chosen preset. It is good at producing a plausible photograph — correct light, correct pose, correct background — but it has a known weakness: the further you get from the source photos, the more the face drifts. By image thirty, the person in the frame may be a cousin rather than you.
Call two: face merge
The second call detects the generated face, compares it with your original photos and merges the identity region back into the generated frame. This is a targeted correction, not a wholesale replacement — it preserves the generated pose, expression and lighting while restoring the specific features that make the face yours: eye spacing, nose shape, jawline, skin tone.
What still goes wrong
The merge pass can be too strong (the face looks pasted on) or too weak (the drift remains). Both are tunable, and both show up as a set-wide problem rather than a single bad image, which is why the redo regenerates the entire set with adjusted guidance instead of replacing one file. It is also why the free redo exists: no automated identity check is perfect, and a second attempt with better source photos fixes most misses.
Why delivery is a set, not a single image
A set lets you pick. Forty images will contain a few frames that flatter you more than the others — different expressions, slightly different light, a better angle. Those are the ones you publish. Single-image tools give you one chance to get it right, which is why their customers so often ask for a refund.
Privacy
What happens to your photos
Three commitments, written into the privacy policy rather than left to interpretation.
Used only for your order
Uploads condition the generation for the order they belong to. They are not used for model training, not published, and not added to any product demo.
Deleted after 30 days
Originals and generated files are removed from storage after 30 days. Earlier deletion on request, from your order email, at any time.
No public figures, no minors
You must be the person in the photos, or have that person's permission. Photos of public figures and of anyone under 18 are prohibited and will be refused.
Full detail — including processors, retention and your rights — is on the privacy page.
Presets
The five style presets
Every preset is available in every package; the preset sets backdrop, wardrobe level, light direction and crop.
FAQ
Process questions
How long does generation actually take?
Roughly two minutes for 40 images, three minutes for 60 and four minutes for 100 at normal queue depth. The page shows live percentage progress while the set renders.
What happens if a photo fails the quality gate?
You see the specific reason — no face detected, more than one face, too much motion blur, head angled too far, sunglasses, resolution too low — and can re-upload before any generation runs. There is no charge for rejected uploads.
Does the pipeline train on my face?
No. Your uploads are used to condition the generation for your order and nothing else. They are not added to a training set, are not published, and are deleted after 30 days.
Which providers process the images?
Generation and face merging run on third-party inference platforms (fal and Replicate) that are contracted for commercial use; payments are processed by our merchant of record; files are stored in object storage with a 30-day lifecycle rule. The privacy page lists the categories in detail.
Why do I need three photos minimum?
One photo gives the model a single angle and a single light, which makes it hard to separate your face from the pose. Three to five photos across slightly different angles and lighting give the identity pass enough signal to keep the set consistent.
What if the model produces an image that looks nothing like me?
That is what the free redo is for. Redo regenerates the set with adjusted guidance and, if needed, a different preset. If the redo still fails, the refund policy applies.
Keep exploring
Related pages
See it work on your own photos
Every set is generated after payment: the pipeline runs on paid orders only, starting with the $9 Sample Pack (5 headshots).
Start with the $9 Sample Pack