How do you put your face on an AI character?
Three methods, from easiest to best: one-shot face swap apps (one photo, seconds), reference-image tools built into major image models (a few photos, no training), or LoRA fine-tuning (15-30 photos, 10-30 minutes of GPU time, best likeness). Use only your own face — putting someone else's face on intimate imagery is a federal crime under the TAKE IT DOWN Act.
Why — the first-principles explanation
The reason this is three methods and not one is that "your face" is not a single kind of information, and each method captures a different amount of it.
A face swap does the least. It detects the face region in a target image, extracts a compact numeric summary of your face — an embedding, essentially a few hundred numbers describing the geometry — and re-renders the target's face using those numbers, then blends the edges. This is fast because it edits pixels in one region and touches nothing else. It's also why swaps look uncanny from the side or in odd lighting: your identity got compressed into a face-shaped patch, and the model has no idea what the back of your head looks like or how your jaw catches light from below.
Reference-image conditioning does more. Modern image models can take a few photos of you as an extra input and steer generation toward that identity while building the whole image from scratch. Because the face isn't pasted on afterward, the lighting and pose are consistent with the scene. The trade-off is that the model is approximating you from a handful of examples, so it drifts — you'll get someone who looks like your cousin.
LoRA fine-tuning does the most, and understanding why explains the photo requirements. You're not showing the model pictures; you're actually changing the model's weights so that a trigger word points at your face. A LoRA (Low-Rank Adaptation) trains a small add-on layer rather than the whole model, which is why it takes minutes instead of weeks. The model learns identity by finding what stays constant across your photos while everything else varies. So if all 20 photos are selfies from the same arm's length in the same kitchen, the model concludes the kitchen and the arm's length are part of your face — and it will faithfully reproduce them. Variety is the training signal. Different angles, lighting, backgrounds, and expressions are what let it isolate you from your circumstances.
One last thing, and it isn't a footnote: the legal exposure here is real and recent. The TAKE IT DOWN Act, signed May 19, 2025, makes knowingly publishing non-consensual intimate imagery — including AI-generated deepfakes — a federal crime, with up to two years' imprisonment for depictions of adults and up to three for minors. Covered platforms had until May 19, 2026 to build 48-hour notice-and-removal systems, and the FTC began enforcement in 2026. Your own face, your own choice. Anyone else's face is a different category entirely.
An example that makes it click
Think about teaching someone to recognize your friend Sam. Method one: you show them a passport photo and say "cut this out and tape it on." Fast, and it looks exactly like a photo taped onto a body — because it is. Method two: you show them four photos and say "draw someone like this." Better integrated, but they'll draw Sam's cousin.
Method three: you show them thirty photos — Sam laughing, Sam in shadow, Sam from below, Sam outdoors, Sam at a party. Now they actually learn Sam. But here's the catch: if all thirty photos are Sam in the same red hoodie, they'll draw the red hoodie every time, because as far as they can tell, the hoodie is Sam. That's why variety matters more than photo count.
How to do it
- Decide which method fits. One photo and you want it now → face swap app. A few photos and you want scene-consistent images → reference-image conditioning in a major image model. Best likeness and repeatable across many images → LoRA fine-tuning.
- For a face swap: upload a clear, front-facing, well-lit photo of yourself plus the target image. Results degrade sharply with profile angles, glasses, heavy shadow, or a large age or face-shape gap between you and the target.
- For reference-image conditioning: supply 3-5 varied photos as reference alongside your prompt. No training step, no wait — the model conditions on them at generation time.
- For a LoRA, gather 15-30 photos with deliberate variety: multiple angles, different lighting, different backgrounds, different expressions, some full-body. Variety is what lets the model separate you from your surroundings.
- Crop and clean the training set. Remove other people entirely, remove heavily filtered shots, remove anything blurry, and avoid photos where you wear the same distinctive item — the model will bake it into your identity.
- Train the LoRA on a hosted service or locally on a GPU. Typical run is 10-30 minutes. Pick a rare trigger word (like a nonsense token) so it doesn't collide with concepts the model already knows.
- Test with the trigger word in varied prompts. If your kitchen keeps appearing in every image, your training set lacked background variety — fix the data, not the prompt.
- Only ever do this with your own face, or with a face whose owner gave you explicit, informed permission. Under the TAKE IT DOWN Act, knowingly publishing non-consensual intimate deepfakes is a federal crime carrying up to two years' imprisonment for adult depictions and three for minors.
Key facts
- The TAKE IT DOWN Act was passed by Congress on April 28, 2025 and signed into law on May 19, 2025; its criminal prohibitions took effect immediately.
- Penalties under the Act: fines and up to two years' imprisonment for sharing intimate deepfakes or authentic intimate images of an adult without consent; up to three years where a minor is depicted.
- Covered platforms had until May 19, 2026 to implement a notice-and-removal process, and must remove flagged depictions within 48 hours of notice; the FTC enforces this.
- LoRA (Low-Rank Adaptation) training typically needs 15-30 varied photos and 10-30 minutes of GPU time — it modifies a small add-on layer rather than retraining the full model.
- Face-swap methods compress identity into a face embedding of a few hundred numbers and re-render only the face region, which is why results fail at profile angles and unusual lighting.
- Training-set variety matters more than photo count: a model learns identity from what stays constant while everything else changes, so repeated backgrounds or clothing get absorbed into the learned identity.
▶ The 60-second explainer (script)
How do you put your face on an AI character? Three methods, and the difference is how much of "you" each one actually captures. Method one, face swap. You give one photo. The tool crunches your face into an embedding — a few hundred numbers describing your geometry — then re-renders the target's face using those numbers and blends the edges. Takes seconds. And it looks pasted on, because it is. Your identity got squeezed into a face-shaped patch. The model has no idea what the back of your head looks like, so profiles and weird lighting fall apart. Method two: reference images. Modern image models take a few photos of you as input and build the whole image around that identity. Lighting matches the scene because nothing was pasted. But it's approximating you from a handful of examples, so it drifts — you get your cousin. Method three: LoRA fine-tuning. This one actually changes the model's weights so a trigger word points at your face. Fifteen to thirty photos, ten to thirty minutes of GPU time. Here's the key insight: the model learns your identity by finding what stays constant while everything else changes. So if all twenty photos are selfies in your kitchen at arm's length — congratulations, the kitchen is now part of your face. Variety is the training signal. Different angles, lighting, backgrounds, expressions. And this isn't a footnote: the TAKE IT DOWN Act, signed May 2025, makes publishing non-consensual intimate deepfakes a federal crime — up to two years for adults, three where a minor's involved. Your face, your call. Anyone else's, get explicit permission.
What authoritative sources say
People also ask
How many photos do I need for a good likeness?
For a LoRA, 15-30 is the usual range — but variety beats quantity. Twenty photos from twenty different angles and settings will beat sixty selfies taken in the same room.
Why does my AI face look like me but not quite?
Usually a training-data problem. If your photos share a background, a lighting setup, or a piece of clothing, the model treats that as part of your identity. Diversify the set and retrain.
Is it legal to put someone else's face on an AI character?
It depends entirely on the content and your jurisdiction. Non-consensual intimate imagery is now a federal crime under the TAKE IT DOWN Act. Beyond that, state right-of-publicity laws, defamation, and platform rules all apply. Get explicit permission.
Do I need an expensive GPU?
No. Hosted LoRA training services run the job on their hardware for a few dollars. Locally, roughly 12GB of VRAM is a common practical floor for training at reasonable speed.
Can I remove a deepfake of myself from a platform?
Yes. Under the TAKE IT DOWN Act, covered platforms were required to have a notice-and-removal process by May 19, 2026, and must remove flagged non-consensual intimate depictions within 48 hours of a valid request. The FTC enforces this.