Character AI, an artificial intelligence platform based in California, introduced its first video generation model on Monday. Named AvatarFX, this innovative image-to-video model is designed to create 2D and 3D animated characters. According to the company, the resulting videos will maintain temporal consistency in facial expressions, hand, and body movements. Furthermore, the model will incorporate speech generated by Character AI’s proprietary text-to-speech (TTS) technology. AvatarFX is expected to be available in the coming months, initially for paid subscribers.
Character AI’s AvatarFX Will Not Generate Realistic Human Characters
In a recent blog post, the company elaborated on the features of its new video generation model. Character AI has predominantly focused on text and image-based applications until now, but with AvatarFX, it is venturing into the realm of AI-generated videos. Currently, access to the model will be granted to Character AI+ subscribers first, although it remains unclear if the feature will eventually be extended to free users.
AvatarFX is capable of generating videos featuring 2D animated characters, 3D cartoon figures, and non-human faces. A key highlight of this model is its focus on maintaining consistency in face, hand, and body movements across frames, which is intended to reduce issues like additional limbs and distorted facial expressions. However, it is essential to note that these capabilities will only be confirmed upon the actual release of the tool.
In contrast to many video generation models, AvatarFX does not utilize text inputs for video creation; it exclusively accepts images. Character AI argues that this approach gives users more control over the final output. The videos will also integrate speech generated through the company’s native TTS technology.
Regarding its technological framework, Character AI revealed that AvatarFX is constructed on flow-based diffusion models. The foundational model is built using the Document Image Transformer (DiT) architecture, which employs a transformer encoder. The company asserts that a new inference strategy was implemented, aiming to uphold visual quality, motion consistency, and expressive diversity, even in longer video durations.
The introduction of AI-driven video generation tools often raises concerns about deepfake technology and the potential for harmful content. To address these issues, Character AI has implemented several safeguards. AvatarFX features integrated safety filters that review the dialogue users input for video creation.
The AI model prohibits the generation of videos using images of minors, prominent political figures, and other notable individuals. Other human images are altered during the generation process to obscure identity, according to the company’s claims. Additionally, every produced video will be watermarked to indicate its AI-generated nature, though details on whether this watermark is visible on the video or included in the metadata remain unclear.
Character AI has established new terms of use for AvatarFX, barring its application for impersonation, harassment, deepfake creation, and unauthorized use of protected intellectual property. Breaches of these terms will lead to a “strict” one-strike ban policy.
