Character AI, a California-based artificial intelligence platform, announced the launch of its inaugural video generation model on Monday. Named AvatarFX, this innovative model is designed to transform images into videos, enabling the creation of both 2D and 3D animated characters. The company asserts that the resulting videos will maintain temporal consistency in facial, hand, and body movements, while also incorporating speech generated by its own text-to-speech (TTS) systems. AvatarFX is expected to debut in the coming months, with early access granted to paid subscribers.
AvatarFX: No Realistic Human Characters
In a detailed blog entry, Character AI elaborated on the capabilities of its new video generation model. Historically, the firm has concentrated on creating text and image-based models. With AvatarFX, it is venturing into the realm of AI-generated video content. Currently, access is limited to Character AI+ subscribers, and it remains uncertain if the model will become available to free users in the future.
AvatarFX is capable of producing videos featuring 2D animated characters, 3D cartoon personas, and non-human faces. A key feature highlighted by the company is the model’s temporal consistency, which is designed to ensure that characters appear stable across frames, reducing the likelihood of glitches such as extra limbs or awkward facial expressions. Nevertheless, these claims will only be verifiable upon the model’s official release.
In a departure from typical video generation models, AvatarFX limits input to images rather than text prompts. This design choice is said to enhance user control over the generated video outputs. Furthermore, the accompanying speech will utilize Character AI’s proprietary TTS technology.
Focusing on the underlying architecture, Character AI disclosed that AvatarFX utilizes flow-based diffusion models. The foundational model is constructed from the Document Image Transformer (DiT) architecture, a transformer encoder model optimized for visual quality and motion consistency, which is particularly effective in longer videos.
The emergence of AI-driven video generation technologies often raises questions about the potential for deepfakes and harmful content. To address these concerns, Character AI has implemented a series of safety filters within AvatarFX to monitor dialogue input for videos. Additionally, the model prohibits generating videos from images of minors, notable political figures, or celebrities, while also anonymizing other human images during the generation process. Each video produced will include a watermark, indicating its AI-generated nature, although specifics regarding watermark placement—whether visible or in metadata—remain unclear.
To further safeguard against misuse, Character AI has introduced new user agreements governing the use of AvatarFX. These terms strictly forbid impersonation, harassment, deepfake creation, and unauthorized use of protected intellectual property, with violations subject to a stringent one-strike policy leading to a ban.
