Joshua Xu

Joshua Xu is the co-founder and CEO of HeyGen, an AI video company that lets people create videos featuring their likeness and voice without recording every new performance. His work grows out of a problem he shared with his co-founder, Wayne Liang: building tools for creators did not make either of them comfortable appearing on camera. They wanted people to share their expertise without having to become confident performers.

From camera engineering to generated video

Joshua studied robotics and specialized in computer vision in Carnegie Mellon’s Master of Robotic Systems Development program. At Snap, he initially worked on recommendation systems before moving into AI camera work. Generative image techniques emerging in 2018 and 2019 suggested that software could produce footage itself. Meanwhile, testing camera features required repeatedly photographing and filming himself—an awkward task for someone who describes himself as an introvert.

He founded HeyGen with Wayne in December 2020. The two had met at Carnegie Mellon, where Wayne studied human-computer interaction. Joshua concentrated on technology and Wayne on design. Their founding conviction connected those backgrounds to their own experience: having something useful to say should not require repeated filming or comfort in front of a camera.

Letting customers change the product

Before a polished software product existed, Joshua manually delivered AI-generated spokesperson videos through Fiverr. The first customer paid $5. The initial listing did not disclose the AI avatars; the team subsequently added that disclosure and continued receiving positive feedback. These orders helped establish practical uses and pricing before the company’s July 2022 product launch under its former name, Movio.

Customer demand also challenged Joshua’s technical preferences. An early computer-generated avatar approach failed to find product-market fit because customers wanted authentic-looking videos. Requests to animate photographs, which the team initially refused, became TalkingPhoto. His account of the early business describes both a failed direction and a reversal: customers’ willingness to pay changed what the team built.

Joshua used his own digital likeness in onboarding, demonstrations, and marketing, making a camera-shy founder the visible face of the product without requiring a new recording for each appearance. He also remained involved in product development, including working with the team on Avatar IV updates.

Preserving the person and expanding the production tools

HeyGen’s work extends from generating a recognizable speaker to translating their message and composing a finished video. These are related capabilities, with distinct technical and product requirements:

• Identity-preserving video: HeyGen’s research team built Avatar V to condition generation on reference-video information, preserving details of appearance alongside characteristic expressions and speaking movements. Its Sparse Reference Attention reduces the computational cost of using longer references. The work gives technical substance to the goal of producing a recognizable digital person; it is a team contribution, rather than an invention attributable to Joshua alone.

• Repeatable, multilingual communication: Joshua’s approach to business video centers on producing lessons, explanations, and other communications without repeatedly arranging a shoot. HeyGen combines avatar generation with voice and translation tools so a person’s message can reach audiences in different languages. Reproducing a face and voice also requires safeguards: Joshua describes live consent and avatar-verification measures alongside these capabilities.

• Composing the finished video: HeyGen’s June 2026 expansion into agentic video includes HyperFrames, an open-source framework for agents to compose videos. This broadens the company’s work beyond generating a performance to helping assemble the finished result.