The history of face swap.

Phase 1: The darkroom era (1800s–1990s)

Putting one face in another image is older than photography itself — painters had been doing it for centuries. Photography made it mechanical. Working photographers and commercial retouchers developed a bag of tricks:

• Cut-and-paste negatives. Cut the head from one negative, paste it onto another, rephotograph the result. The boundary was visible, the technique was labor-intensive, and the output was a single image — not a video.
• Double exposure. Expose the same frame twice, with two different subjects. Produced ghosting and registration problems; often visible as such.
• Airbrushing and retouching. Used to soften boundaries and match lighting across composites. The standard commercial-retouching toolkit through the 20th century.
• Photo montages in political imagery. Soviet-era retouchers were particularly known for removing people from official photos. Stalin's purges led to a body of work that was technically competent, politically devastating, and historically well-documented.

Throughout this era, the ethical line was the same as today: consent from the subject, no impersonation for fraud, no non-consensual intimate imagery. The technology changed; the line didn't.

Phase 2: Computer-vision face swap (1990s–2016)

Academic computer-vision research in the 1990s and 2000s developed the foundations of automated face-swap. Major contributions came from graphics and vision research groups — face detection (the Viola–Jones framework, 2001), face landmark estimation, and the first learned face-swap models.

Through the 2000s and early 2010s, the research was mostly in academic papers. The models worked on still images, took significant compute to run, and were not consumer-accessible. The output was a single image, not a video, and the use cases were mostly in film VFX and academic research.

Phase 3: The deepfake era (2017–2019)

2017 was the watershed year. Three things happened in quick succession:

• First open-source face-swap models. Researchers released the first open-source implementations of automated face-swap models. The technology moved from academic papers to anyone with a consumer GPU.
• The 'deepfake' term enters public vocabulary. A Reddit user named the practice, and the name stuck. The term tracks the misuse cases more than the technology, which is one reason "face swap" is the preferred term in consumer contexts.
• First major misuse cases. Non-consensual intimate imagery, public-figure impersonation, and political manipulation got mainstream news coverage. The legislative response (US state laws on NCII, EU AI Act provisions, China's deepfake regulations) dates from 2018 onward.

This is also when the safety tooling started: facial-age estimation, NSFW classification, CSAM hash matching, and watermarks. The category learned that misuse was the first-order problem and built the defenses in response.

Phase 4: The consumer era (2020–present)

2020 onward: consumer face-swap tools become a real product category. Reface, FacePlay, Wombo, and a long tail of mobile apps brought the technology to a mass audience. The shape:

• Mobile-first UX with a polished template-browsing experience.
• Subscription pricing for the polished apps ($30/month for Reface Pro was the standard).
• Web-based credit models for tools aimed at marketers and content teams.
• Telegram-native face-swap (Standin, launched 2025) for users who don't want to install an app.
• Safety tooling now table-stakes — facial-age estimation, CSAM hash matching, watermarks, published takedown SLAs.

The current wave is the consolidation phase. The technology is mature, the safety stack is established, and the differentiators are now form factor (chat-native vs app vs web) and use case (cinematic vs meme vs marketing), not the model quality.

What the history teaches

Three things, in order of importance:

• The technology is new, the impulse isn't. Putting one face in another image goes back to the 19th century. The AI model makes it fast and accessible; the underlying impulse is the same.
• The ethical line has always been consent. Not the technology. Not the medium. Not the era. The line is the same in the darkroom era, the computer-vision era, and the consumer era: consent from the subject, no fraud, no NCII.
• Safety tooling emerged in response to misuse. The facial-age estimation, NSFW classification, CSAM hash matching, and watermarks that are now standard were not in the first wave of consumer tools. They were added because the category learned that misuse was the first-order problem.

The next phase of the history is going to be about provenance — the technical standards for marking model output as model output, so the "reasonable viewer would believe" test is settled by the file itself, not by a watermark. The C2PA standard is one direction. The category is moving.

Where Standin fits in the history

Standin is the chat-native, pay-per-clip wave of the consumer era. Telegram-native (no install), pay-per-clip ($9+), one free render (watermarked), and the most publicly documented safety stack in the category. The bot is a 2025 product; the underlying model is 2017; the impulse to put your face in a different scene is older than photography.

Related pages

Common questions

When was face swap invented?

The impulse is 19th-century. The computer-vision work is 1990s. The consumer era is 2017 onward.

Who invented face swap?

No single inventor. The darkroom era was working photographers and retouchers. The computer-vision era was academic research groups. The consumer era was open-source releases in 2017.

How did deepfakes become mainstream?

Three waves: 2017 (open-source release + the 'deepfake' term), 2018–2020 (misuse cases + legislative response), 2020 onward (consumer tools + safety tooling).

Is face swap a new technology?

The AI model is new (2017). The impulse isn't. 19th-century photo composites did the same thing with cut-and-paste darkroom work.

What changed in 2017?

Open-source face-swap models, the 'deepfake' term, and the first major misuse cases. The legislative response dates from 2018 onward.