You upload a handful of selfies, wait five to ten minutes, and download 30+ professional headshots. The result feels like magic, especially when you compare it to the alternative of booking a photographer, traveling to their studio, sitting under hot lights, and waiting two weeks for retouched files. But the technology is not magic. It is a stack of three well-understood machine learning ideas, each of which has been developed over the past decade. This post is an honest walk through what actually happens under the hood, and what we specifically run at AI Portrait Studio.

TL;DR

• AI headshot tools combine three things: a face encoder that captures your identity, a diffusion model that knows how to draw realistic photos, and a glue layer that keeps your face recognizable across new images
• We run PhotoMaker on Replicate (tencentarc/photomaker), which uses ID embeddings rather than per-user fine-tuning, so generation is fast and cheap
• Diffusion models are trained on huge image datasets and learn the visual patterns of professional portraiture
• Your photos are stored on Cloudflare R2 with a 48-hour automatic expiry, then deleted
• We accept HEIC files directly — modern iPhone uploads work without conversion

Step 1: Capturing Your Identity (Face Encoders)

The first thing the system does with your selfies is extract a compact mathematical representation of what makes your face yours. This is called an embedding, and it is produced by a face encoder — a neural network specifically trained to recognize and distinguish faces. The encoder doesn't store your photos; it produces a vector of numbers that captures the shape of your features, the proportions of your face, the placement of your eyes and mouth, and other identity-defining cues.

Face encoders have existed in production for over a decade — they are the same family of models that power face unlock on phones and tagging suggestions on photo apps. Modern versions can preserve identity across different angles, lighting conditions, and even modest changes in expression or hairstyle.

Step 2: The Generative Model (Diffusion)

The model that actually paints the new headshot is a diffusion model. Diffusion models learn to generate images by reversing a noising process: during training, the model sees images that have been progressively corrupted with random noise, and it learns to undo that noise step by step. After training on diffusion models trained on millions of images, the network has learned, in a deep statistical sense, what realistic photos look like — including what professional portraits look like, how studio lighting falls on faces, and how clothing textures, backgrounds, and depth-of-field tend to appear in editorial photography.

Stable Diffusion and its descendants are the most well-known examples in this family. They generate new images from text prompts — 'a professional headshot of a person in a navy blazer, soft studio lighting, neutral background' — by starting from pure noise and iteratively denoising into a photo that matches the prompt. The output is a brand-new image, not a stitched-together version of training data.

Step 3: Keeping You Recognizable (Identity Preservation)

The hard part is making sure the diffusion model's output actually looks like you. There are two main approaches in production. The older one is per-user fine-tuning: you train a small model adapter (a LoRA or DreamBooth model) on your specific photos, and that adapter biases the diffusion model toward generating your face. This produces high-quality results but takes 15-45 minutes per user and is expensive to run at scale.

The newer approach uses ID embeddings: the face encoder from step 1 produces a vector that gets injected into the diffusion model at generation time, conditioning the output to look like the person in the reference image. This is what PhotoMaker does. It eliminates the per-user training step entirely, which is why generation can complete in 5-10 minutes instead of 30-60. The tradeoff historically was a small drop in identity fidelity, but the gap has narrowed significantly with newer model versions.

• Per-user fine-tuning (LoRA, DreamBooth): higher fidelity, slower, more expensive per user
• ID embeddings (PhotoMaker, IP-Adapter, InstantID): faster, cheaper, slightly less fidelity but rapidly improving
• Hybrid approaches use ID embeddings as a base and apply lightweight adapters when needed
• Newer architectures (FLUX-based ID models) are pushing the fidelity gap close to zero

What We Actually Run

AI Portrait Studio runs PhotoMaker on Replicate. The model identifier is `tencentarc/photomaker`. PhotoMaker uses the ID-embedding approach described above, conditioned on a small set of reference photos (your selfies). On the generation side, it produces 1024x1024 outputs across multiple style prompts — corporate, business casual, creative, formal — to give you variety from a single upload.

We chose PhotoMaker over per-user fine-tuning because the speed difference matters at our price point. A 5-10 minute generation lets us deliver 30+ photos for $12.90. Per-user fine-tuning would push the cost into the $30-$50 range, which exists in the market but isn't the niche we're serving. If you want a deeper comparison of where AI headshots fit relative to traditional photography, our AI versus traditional photographer breakdown lays it out.

Storage, Privacy, and HEIC Support

Once you upload, your selfies go to Cloudflare R2 with an automatic 48-hour expiry. After that window, the files are deleted. We hold the originals only long enough to run the generation and let you re-download your results if needed. We don't train on your photos, sell them, or share them with anyone outside the generation pipeline.

We also accept HEIC files directly — the format that modern iPhones use by default — so you don't have to convert your photos to JPG before uploading. Most older AI headshot services would silently fail on HEIC uploads or require manual conversion. We added native HEIC support specifically to remove that friction for iPhone users.

Why It Looks So Realistic

A few things compound to produce results that look indistinguishable from studio photography at a casual glance. First, diffusion models have learned the actual physics of how light interacts with skin, hair, and fabric — not by being told the rules, but by absorbing the patterns from millions of real photos. Second, professional portraiture is a fairly constrained genre with predictable lighting, framing, and background conventions, which makes it easier to model than, say, generating photorealistic medical illustrations. Third, the face encoder anchors the generation to your real identity, so the output is not just 'a generic professional headshot' but 'a professional headshot that looks like you.'

Where the technology still falls short: hands (notoriously hard for diffusion models), complex group shots, specific brand logos on clothing, and unusual props. For a shoulders-up portrait — which is what 99 percent of professional headshots actually are — the limitations rarely show up in the final output.

Should You Be Honest That It's AI?

Yes, when it matters. For LinkedIn, freelance platforms, your email signature, and most general professional uses, no one expects a disclosure that your headshot was AI-generated, just like no one expects a disclosure that your traditional headshot was retouched. Both are professional polish. For acting headshots, modeling portfolios, journalist bylines, government IDs, and any context where the photo is meant to certify your literal current appearance, use a real photo. AI headshots are a tool for professional presentation, not for impersonation.

Common Mistakes

• Uploading too few selfies (under 3) — the model can't capture your identity well from one or two photos
• Uploading selfies that all look the same — variety in angle, lighting, and expression improves output
• Uploading photos with sunglasses, hats, or heavy makeup — the model has less to work with
• Expecting AI headshots to look like a different person than you actually are
• Using AI headshots for contexts that require certified real-time photos (passports, IDs)
• Picking the most flattering output instead of the most accurate one — your real-life self should match the photo

FAQ

**Are AI headshots ethical?**

For professional presentation, yes. They sit on the same continuum as makeup, retouching, lighting, and wardrobe choices that have been part of headshot photography for decades. The line is the same one that has always existed: present your best self honestly, but don't impersonate someone you're not. AI headshots that look like you on a good day are fine; AI headshots that erase your actual identity cross the line.

**Will AI headshots get rejected by hiring managers?**

Almost never, because they look like normal professional headshots. Most hiring managers cannot tell the difference between a $300 photographer-shot headshot and a $13 AI-generated one for shoulders-up LinkedIn use. The exceptions are creative industries that explicitly value craft authorship (some photography, design, and editorial roles), where some recruiters check provenance more carefully.

**How are my photos protected once I upload them?**

Your originals are stored on Cloudflare R2 with a 48-hour expiry, then deleted. They are not used to train any model, not shared with third parties, and only accessible to the generation pipeline that produces your results. The face embedding generated during processing exists only in memory during generation and is discarded afterward.

**Why do I need to upload multiple photos instead of one?**

A single photo gives the face encoder one angle and one lighting condition to work from. Multiple photos give the encoder a more complete picture of your face — different angles, different expressions, different lighting — which improves the fidelity of the generated headshots. We recommend 3-10 selfies, with variety in angle and expression. Our selfie guide for AI headshots covers exactly what kinds of source photos produce the best output.

**Can the model generate me in clothes I don't actually own?**

Yes. The diffusion model generates wardrobe based on the style prompts, not based on what you uploaded. If your selfies are in t-shirts, you can still get outputs in suits, blazers, or business casual. The model treats clothing as a separate generation choice from identity, which is why one selfie batch can produce headshots across formal, business casual, and creative styles.

**How is this different from filters or face-swap apps?**

Filters apply transformations on top of an existing photo. Face-swap apps paste your face onto a different image. AI headshots generate brand-new photos from scratch, conditioned on your identity. The output isn't a modified version of your selfies — it's a new image that didn't exist before, painted by a diffusion model that learned what professional portraits look like by training on millions of real photos.