A free studio,
because $300 for a profile photo is absurd.

For most people, a great headshot meant calling a photographer, blocking out half a Saturday, paying upwards of $300, then waiting two weeks for retouched files. For the other half of the world, a headshot meant cropping a vacation photo and hoping the lighting wasn't too brutal.

AI changed the math in 2024. Suddenly the same image, sharp eyes, soft loop lighting, a blazer you don't own, could be generated from a phone selfie in under a minute. Then the AI tool industry promptly turned it into a $35 to $75 paywall, with 12-photo upload requirements, 30-photo email gates, and subscriptions that auto-renew until you call your bank.

FreeHeadshot is the version that didn't do that. One selfie. One minute. Three free results, watermarked but actually usable. $9 to $49 once, no recurring charge, if you want the watermark-free pack at Full HD with all 100+ styles.

We get a lot of "is this a wrapper around ChatGPT" questions, so here is the honest list of what runs every time you press Generate. No mystery proprietary models, no "we trained our own AI". A small team can only credibly stand behind a thin layer on top of best-in-class vendors, so that's what this is.

• Image generation: Google Gemini 2.5 Flash Image API, the model Google shipped in October 2025. It is currently the best general-purpose image model for instruction-following at the price point. We do not train it. We send it your photo, a structured photographic prompt, and parameters; we get back generated images.
• Post-processing: the open-source Sharp library for Node. We auto-orient based on EXIF, run a saliency-guided crop so your face stays centered, apply light tone grading, add a small amount of micro-grain to dodge the plastic-AI look, and a soft vignette. All of this is deterministic image math, not AI.
• Quality scoring: a small bank of heuristics we wrote ourselves. We measure cheek-region high-frequency variance (proxy for skin texture vs over-smoothing), Laplacian edge density (sharpness), and color-cast standard deviation. We aggregate these into a 0-to-1 quality score per image and rank the output so the strongest result lands first. The bottom of the batch gets quietly dropped before you see it.
• Hosting: Vercel for the application and edge functions, Cloudflare in front as a CDN and DDoS shield. The site is statically generated where possible to keep latency under 100ms on the first byte for the homepage.
• Authentication: Supabase magic-link. No password storage at all. You enter your email, you get a one-time link, you sign in. Sessions are cookie-based and expire after 30 days of inactivity.
• Payments: Creem. We never see card numbers. We see only the last four digits and a Creem transaction reference, which we keep for tax and refund purposes.
• Database: Supabase Postgres for the small amount of account metadata we need (email, plan, generation history). Row-level security policies enforce that no user can read another user's records, even if the application logic is compromised.
• Analytics: Vercel Web Analytics, cookieless. We count page views. We do not track individuals. There is no Google AdSense, no Facebook Pixel, no third-party tracker on the site.

This is what happens when you press Generate. We walk through it in detail because the "where does my face go?" question deserves a real answer.

1. Your browser uploads the photo over HTTPS (TLS 1.3) to a Vercel-hosted endpoint. The endpoint is geographically close to you, automatically picked by Cloudflare based on your location.
2. We strip EXIF metadata from the upload (GPS coordinates, camera model, timestamps). The model does not need that data, and you should not be exporting it to anyone.
3. We forward the stripped photo to Gemini 2.5 Flash Image over Google's encrypted API endpoint. Along with the photo, we send a structured prompt: focal length (85mm), aperture (f/3.2), lighting (loop or Rembrandt depending on style), framing (chest-up to avoid hand artifacts), and grade reference (Kodak Portra for color, Tri-X for black-and-white).
4. Gemini returns generated image bytes. Typically three to five candidate images per request depending on the style and tier.
5. Sharp post-processes each candidate. Auto-orient, saliency-aware crop to 1024x1024 master, light L-channel sharpening for catchlight and lash detail, tone grade based on the selected style, soft vignette, and a touch of micro-grain to break up the "pristine AI" pixel signature.
6. We score each candidate using the cheek-variance, edge-density, and color-cast heuristics described above. The aggregate score sorts the output so the strongest image displays first.
7. Free-tier outputs get a small bottom-right watermark baked in by Sharp. Paid-tier outputs do not. All outputs are exported as JPEG at quality 92 for fast delivery, with a sRGB color profile.
8. The finished images come back to your browser inline as base64 data URIs in the JSON response. There is no Content-Disposition attachment header, so nothing auto-downloads. The original photo and any temporary intermediate files are deleted from our infrastructure within 24 hours.

Negative claims build trust faster than positive ones. Here is the explicit list of things we have decided not to build, and will refuse to build, regardless of what someone offers us.

• We will not train any AI model on user-submitted photos. Not Gemini (which we do not control, but which Google contractually does not train on customer content under its API terms). Not any downstream model. Not any future model.
• We will not build face-swap features. The whole point of FreeHeadshot is that the output looks like you. Face-swap inverts that and is the foundation of most deepfake abuse.
• We will not build "undress" or de-clothing filters. These exist on other tools. We refuse the entire category.
• We will not generate images of people who did not consent. The pipeline is wired to require an uploaded photo as conditioning. We do not allow text-only generation of named people, including celebrities, politicians, or any private individual.
• We will not retain face embeddings or biometric vectors beyond the few seconds of a single generation request. There is no "user face database" sitting on our servers.
• We will not sell, rent, or share user photos with any third party. The only company that ever sees the bytes is Google's Gemini API for the single purpose of performing the generation you requested.
• We will not run advertising on the site. No display ads, no sponsored content, no affiliate offers in the studio. The way the site looks today is the way it stays.
• We will not auto-renew any paid tier. Every paid purchase is one-time, full stop. The infrastructure to silently bill you again does not exist.
• We will not move users from free to paid via dark patterns. No countdown timers pretending we will delete your free results, no "your trial ends in 24 hours" emails, no buried cancel buttons.

AI image models have a documented history of producing worse results for non-Western faces, darker skin tones, certain hair textures, and faces that wear religious garments or visible disability aids. The 2023 wave of headshot tools made this problem visible at scale: a Bloomberg investigation that year highlighted how dramatically results dropped for darker-skinned subjects on early diffusion-based tools.

Gemini 2.5 Flash Image has measurably narrowed that gap compared to older Dreambooth and LoRA pipelines, but no AI image generator is bias-free. Here is how we approach the problem in practice:

• We test every new style across Fitzpatrick skin tones I through VI before we ship it.
• We test with and without prescription glasses, with multiple hair textures (straight, curly, coily, locs, braids), with religious head coverings (hijab, kippah, turban), and across a range of face shapes.
• The quality scoring described above is style-agnostic. A weak result is filtered regardless of who the subject is.
• When a style fails consistently for a demographic, we either tune the prompt or pull the style. We have pulled styles for this reason and we will again.
• If the tool fails for you and the failure looks demographic, please email us with the original photo and a description. We treat that as a high-priority bug, not a complaint.

We also published an honest audit at the skin-tone bias post, with examples and the data we collected.

Most AI headshot tools prompt models with marketing adjectives, “professional,” “flawless,” “beautiful.” That produces the waxy, plastic-skin look every reviewer complains about. FreeHeadshot prompts in the language of working photographers: 85mm at f/3.2, soft loop lighting from camera-left, single iris catchlight at 10 o'clock, Kodak Portra 400 grade. The result reads as a real studio shot, not a beauty filter.

Every output passes through a post-process pipeline (eye-region sharpen, color grade, vignette, micro-grain) before it lands on your screen. We also score every result on internal quality metrics, cheek-region texture variance, edge density, color cast, and rank the strongest as the hero result.

Read the full technical breakdown →

Honest question, honest answer. The economics are not complicated.

• Per-user cost on the free tier: roughly $0.02 in API and infrastructure cost across the three generations. Most free users do not return. Most do not pay.
• Per-user cost on a paid tier: $0.20 to $0.80 depending on the tier and number of generations consumed. Margin is solid but not extraordinary.
• Conversion rate from free to paid: we built the budget assuming 2% conversion. If we hit 2%, the math works. Today the actual rate is in that range.
• Marketing budget: zero. We do not run ads. The only acquisition channels are organic search, AI search citations (which is what most of the recent content work has been about), and word of mouth.
• Team size: small enough that one founder reads every support email personally. Salaries and infrastructure are the only ongoing costs.

The reason the free tier is real and the paid tier is cheap is that the cost structure allows it. We do not have a venture-funded growth target forcing the price up. We have a sustainable break-even, and we built the pricing around that.

Honest limits beat marketing claims. If you hit one of these and we did not warn you, you would assume the tool was broken or that we lied. Both are worse than just telling you what does not work.

• Very low-resolution source photos. If the upload is under about 500 pixels on the long side, results degrade fast. The model has nothing to anchor identity on.
• Harsh backlighting. Window-behind-the-subject photos where the face is in shadow give the model too little signal. We do our best but the failure rate is higher.
• Extreme angles. The model is trained primarily on front-facing and slight three-quarter photos. Profile shots and looking-up shots produce less convincing results.
• Group photos. The model will pick one face but the result is unreliable. Crop to a single subject before uploading.
• Heavy filters on the source. If the source photo already has aggressive smoothing, beauty filter, or extreme color grade applied, the model has trouble identifying the underlying face. Use an unedited photo.
• Government identification. AI headshots are not photographs. They are not accepted for passports, drivers' licenses, or most government processes. They are appropriate for LinkedIn, websites, sales decks, conference badges, and press kits, not for ID cards.
• Hands and fingers in the frame. We crop to chest-up to avoid this, but if your source photo includes hands, occasional artifacts may slip through.

Where the tool fails for you, the workaround is almost always: better source photo, different style, regenerate. If that does not solve it, email us and we will look at the specific case.

Composite of what shows up in our support inbox and in the dashboard. Not a marketing persona list, an actual pattern we see week to week.

• Job seekers updating LinkedIn after a layoff. The single biggest segment. Often the user got the news that morning and wants the new photo live before the recruiter call that afternoon.
• Founders fundraising. AngelList page, deck founder slide, press kit, all of which need the same shot in different aspect ratios. The bulk paid tier covers it for the cost of a sandwich.
• Sales and BD reps. The team that lives or dies by their LinkedIn presence. Multiple-times-per-year refresh as territory changes or quota cycle resets.
• Conference speakers. The press-kit photo, the badge, the website thumbnail, the Twitter handle photo. All from a single selfie.
• Podcast hosts. Square crop for the show art, rectangular for Spotify and Apple, black-and-white for the press release.
• Remote teams. When you have to assemble a team page and nobody works in the same city, AI headshots give a consistent look across the page.
• Students applying for internships. The students who would otherwise put a vacation crop on their resume and lose interviews to it.