A leading company in personalized AI chatbot services, Character.AI, has unveiled a new video generation model called AvatarFX. This model enables users to transform images into lifelike AI chatbots.
Currently, the model is accessible to a limited number of users in a closed beta testing phase. According to the company, AvatarFX can animate images, allowing them to talk, sing, and convey emotions. It maintains high consistency with photorealistic human faces and supports videos with multiple speakers, accompanied by audio, animated limbs, hands, and body movements.
"Our model performs exceptionally well across various styles and scenarios, from real humans to mythical creatures, and even inanimate objects with faces," stated Character.AI in their announcement.
To build the video generation model, the company explained that its multimodal team developed a specialized parameter-efficient training process, enabling diffusion AI models to generate realistic lip, head, and body movements based on audio sequences. Diffusion is an AI image generation method that gradually adds noise to training data and then reverses the process to create new images.
For audio generation, the company utilized its proprietary text-to-speech model.
To reduce costs, Character.AI mentioned that state-of-the-art distillation techniques were employed during training, minimizing the number of steps required for runtime diffusion. This significantly cuts down on computational requirements during deployment without compromising quality.
Character.AI emphasized that AvatarFX stands apart from other avatar generation models as it can effortlessly create high-quality 2D animated characters, 3D cartoon figures, and videos featuring non-human faces. Consistency is maintained even in longer videos, and there's no need for text-to-image generation.
The company highlighted that, even during the testing phase, they are prioritizing the safety of user-uploaded content to prevent misuse and deepfakes. User-written conversations will be filtered through content safety mechanisms, and photos of minors, high-profile political figures, or other notable individuals will be blocked. Additionally, videos will be watermarked to clearly indicate that they are AI-generated.
Recently, Character.AI has faced multiple lawsuits alleging that its personalized AI chatbot characters promote self-harm and violence against children. In one case, a mother from Florida claimed her 14-year-old son committed suicide after becoming addicted to one of the company’s ultra-realistic chatbots.
In response, the company launched a new Parental Insights feature in March to enhance safety for teenagers. The feature includes dedicated models for teen users, enhanced human behavior detection and intervention systems, improved AI model responses, and provides parents and guardians insights into minors' activities on the platform.
