Character.AI has announced AvatarFX, a new video generation model that can animate both AI characters and uploaded images. According to the company, the technology can generate photorealistic videos with audio that maintain consistency in face, hand, and body movements, even for long-form content.
The system builds on diffusion model architecture with a parameter-efficient training pipeline to create realistic lip, head, and body movements based on audio. Character.AI claims AvatarFX can animate a variety of subjects beyond humans, including 2D animated characters, 3D cartoons, and even pets.
With the potential for misuse, Character.AI has implemented several safety features:
- Content filtering for dialogue that would violate policies
- Blocking video generation using photos of minors, politicians, and public figures
- AI modification of human photos to make subjects unrecognizable
- Watermarking to indicate generated content
- Terms of service prohibiting impersonation, bullying, and deepfakes
Despite these safeguards, concerns remain. TechCrunch notes that Character.AI has already faced lawsuits from parents alleging its chatbots encouraged children to harm themselves or others.
The company plans to gradually roll out AvatarFX to its platform, with CAI+ subscribers getting early access. The feature is currently in closed beta testing.