Face swap vs deepfake — what's the difference?

The short version

Face swap and deepfake refer to the same family of AI model — a system that detects a face in a target and replaces it with another face, preserving expression, lighting, and motion. The difference is the language, and the language tracks who is talking about it.

• Face swap is the everyday consumer term. The result is a video of you in a different scene. The tone is playful, casual, and consumer-product.
• Deepfake is the AI-research / synthetic-media / news-coverage term. It entered the public vocabulary through academic papers and news coverage of misuse. The tone is technical, often critical, often associated with the harm cases.

Both terms describe the same technology. The ethics and legality are downstream of the use (consent, intent, target), not the model.

Technically, how a face swap works

Under the hood, a modern face-swap model has three steps:

1. Detect the face in the target video frame. The model finds the bounding box, the landmarks (eyes, nose, mouth), and the head pose.
2. Generate the new face from the source image. The model takes your selfie and renders it as a new face that matches the target's expression, lighting, and head pose.
3. Blend the new face back into the frame. The model composites the new face over the target, smoothing the boundary so the swap is invisible at a glance.

Most modern models do this on a per-frame basis, with a temporal consistency pass to keep the face from flickering across frames. The result is a video that reads as a single continuous shot, not a composite.

Where the ethical line actually is

The line is consent + intent, not the technology. Three checks separate a legitimate use from a harmful one:

• Consent from the subject. The face in the output has agreed to be there. If it's your own face, you can use it however you want. If it's someone else's, they need to have agreed to this specific use.
• Adult subject. No minors in the output. This is a hard rule. The safety stack (facial-age estimation + NSFW classification + CSAM hash matching) is the enforcement layer.
• Not used to deceive in a way that causes harm. Satire, parody, and commentary have legal protection. Impersonation for fraud, defamation, or non-consensual intimate imagery does not. The watermark on a free render exists to break the "reasonable viewer would believe" part of the deception test.

Any one of those failing turns a legitimate tool use into an unethical one. The model is the same. The use is what matters.

How to tell a face swap from a real video

You often can't, on a single frame. The signs are subtle:

• Edge artifacts around the jaw or hairline. The blend pass isn't perfect; under high-contrast lighting or rapid head motion, you can see a faint line.
• Slight lighting mismatch. The face's lighting tracks the target scene, but if the source photo was shot under different light, the match isn't perfect.
• Occasional flicker on motion. The temporal consistency pass keeps most clips stable, but fast head turns can show a brief mismatch.
• Provenance is the easier tell. If the clip was posted by a face-swap tool with a watermark, it's clearly a face-swap. Without that, the answer is "look at the source, not the output" — most viral "deepfake" claims turn out to be either real or a low-effort composite, not a model output.

For a more serious test, there are detection tools (e.g., the Hive, Sensity) that score the probability a clip is a model output. They're not perfect, but they're useful as a second opinion.

How the law is catching up

The legal framework is downstream of the use, not the technology. Three things are settled in most US jurisdictions:

• Non-consensual intimate imagery is illegal. The face-swap-vs-NCII distinction is whether the subject consented to the result, not whether the model exists.
• Impersonation for fraud is illegal. Using a face-swap to pose as a real person for financial gain is fraud regardless of the technology used.
• Defamation via synthetic media is illegal in many states. Putting words or actions in a real public figure's mouth that they didn't say, in a context where viewers could reasonably believe the clip is real, can be defamatory.

The unsettled areas are around political deepfakes, non-consensual intimate imagery where the source is a publicly available photo, and platform liability. The category is moving fast — the legal answer that applies today may not be the one that applies in 12 months.

What responsible tools do

Three signals separate a responsible face-swap tool from an irresponsible one:

• Publishes a Trust & Safety page with a real takedown SLA (24-hour acknowledgement, 72-hour removal). Not "we take safety seriously" — an actual address, a real response-time commitment, a real removal process.
• Watermarks the free tier so the output is clearly identifiable as a face-swap. This is a trust feature, not a paywall trick.
• Runs automated checks for minor faces and CSAM hash matching. A legitimate tool is running at least these two checks on every input. If a tool doesn't mention either, it isn't running them.

Standin does all three. The full Trust & Safety page is at /safety.

Related pages

Common questions

What is the difference between face swap and deepfake?

Technically, the same. The terms are used differently — 'face swap' for consumer products, 'deepfake' for AI-research / news-coverage. The ethics and legality are downstream of the use, not the technology.

Is a face swap the same as a deepfake?

Yes, technically. The model does the same thing. The ethics are downstream of the use.

Are face swaps legal?

Depends on the jurisdiction and the use. NCII, fraud, and defamation are illegal. Satire and parody have stronger protection but aren't unlimited.

How do you tell a face swap from a real video?

You often can't, on a single frame. The easier tell is provenance — look at the source, not the output. Detection tools exist but aren't perfect.

What's the ethical line?

Consent + adult + not used to deceive in a way that causes harm. Any one of those failing turns a legitimate use into an unethical one.