How many photos you need to train an AI model of a person

The number depends on which kind of AI model you are building. On PixelPig, a reference persona needs one clear face photo looking straight at the camera, and a trained LoRA needs 30 to 40 photos of the same person. If you are still at the stage of picking who your model is, start with the full walkthrough on how to create an AI influencer, and once the photos are trained, the guide on making AI photos look real covers the prompting side.

Most advice on this question quotes numbers from someone's personal experiment. This page does something narrower: it reports the limits our own trainer enforces, quoted from the code and the training screen, and explains what the trainer does with each photo so the limits make sense.

Proof

These are the rules PixelPig applies when you upload a training set, taken from the training screen and the edge function that starts the run. Nothing here comes from a test we ran for this page.

The code comment next to the minimum explains the floor directly: a hard floor of 30 "keeps a too-thin set from producing a weak LoRA." The code gives no reason for the ceiling of 40. It is the cap, and we will not invent a reason.

The step count explains why variety beats volume. With 3,000 steps and one photo per step, a 30-photo set shows each photo to the model about 100 times, and a 40-photo set about 75 times. Ten near-identical selfies in a set of 30 means a third of all training time goes to one pose. The model learns that pose as part of the face.

Two routes, two photo counts

PixelPig gives you two ways to build a model, and they need very different uploads.

Reference persona: one required photo

Go to Personas, then Create Persona, then "Use my own photos". The upload screen has three slots. The face photo is the only required one. The two full-body slots are optional, and the screen says adding them "improves face-swap variety and LoRA training quality." The image and video models that accept a persona then use its photos as identity references, so nothing is trained. The AI persona page shows this flow, and Academy chapter 3 explains why the face photo should be a chest-up shot looking straight at the camera.

The tips on that screen are specific. For the face slot: clear face, neutral expression, facing forward, head and chest in frame. Not: eyes closed, sunglasses, or a sharp high or low angle. For the full-body slots: clear face, arms and hands in frame, standing rather than sitting or lying down.

Trained LoRA: 30 to 40 photos

A LoRA is a small add-on file trained on your photos so the image model learns one specific person. On PixelPig you pick the persona, upload 30 to 40 photos, and the studio uploads them, captions each one and trains the LoRA for you. Academy chapter 4 covers when a creator needs this step. Training is an extra step, not the starting point: a reference persona already works on six image models and seven video models.

How to build a 30 to 40 photo set

Start from one person and one look

Every photo shows the same person. If your persona has a tattoo or a beauty spot, it sits on the same side in every photo. Because the trainer never mirrors photos, a mirror selfie that flips a tattoo to the other arm teaches the model two versions of the body.

Spread the set across framing

The extended tips ask for the face clearly visible in most photos and a mix of close-ups and full-body shots. A workable split for a set of 30 is roughly half head-and-shoulders shots, a third waist-up, and the rest full body. That split is our suggestion, not a rule the trainer checks.

Change everything except the person

Different rooms, streets, outfits, times of day and light sources. Window light, overcast daylight, warm indoor lamps, an on-camera flash. If every photo has the same white wall behind her, the wall becomes part of what the LoRA learns.

Use originals at full size

Upload the original files, not screenshots or social media downloads that have been compressed twice. The 10MB cap per photo leaves room for full-resolution phone photos.

Check the count before you start

The counter under the tips reads "X of 30 minimum added". Remove weak photos first, then fill back up to 30 or more with stronger ones, rather than padding with near-duplicates to reach the floor.

What to keep and what to drop

Common mistakes

Padding to the minimum with duplicates. The counter only checks the number. The quality of the LoRA depends on what those 30 photos contain.

Mixing two people. A set that includes a sister, a friend or an earlier version of the persona blends faces.

Leaving filters on. Beauty filters smooth pores and reshape features. The LoRA then reproduces the filter in every image, which works against the realism you want.

Training before the face is settled. If you are still changing the face, finish the persona first. Designing the face is the chapter for that decision, and our list of AI influencer niche ideas helps you settle who the model is before you spend a training run on her.

Uploading photos of a real person without permission. See the next section.

Consent before the upload

PixelPig's Acceptable Use Policy requires informed consent from every real person in the photos you upload, including you, and the consent has to cover the actual use. Permission to appear in a holiday photo is not permission to become a persona. You must be able to show that consent if asked, and if the person withdraws it, you delete the persona. Fully synthetic personas not based on any real person are permitted.

Bottom line

For a reference persona, one sharp, front-facing face photo is enough, and two full-body photos make it better. For a trained LoRA on PixelPig, the answer is 30 to 40 photos of one person, because the trainer refuses anything else. Inside that range, spend your effort on variety: framing, outfits, light and backgrounds, with no filters, no sunglasses and no flipped tattoos. LoRA training sits on the Pro and Scale plans, listed on the pricing page.