The Fifth Elephant 2025 Annual Conference (18 & 19 July)

We present a production-ready lip-sync model serving millions of creators, built on a novel Latent-GAN architecture that achieves superior identity preservation and audio-visual alignment compared to other approaches.

Our system is trained on 10,000+ hours of diverse audio-visual data using custom preprocessing pipelines which include audio diarization, vocal separation, and AV synchronization etc. We demonstrate how GAN architectures with transformer attention mechanisms and VAEs can match diffusion model quality while offering faster inference speeds.

Key technical contributions include:

• Latent-GAN architecture leveraging transformer blocks and VAE improvements for high-resolution output
• Custom loss functions for facial feature consistency and identity preservation across diverse demographics
• Scalable preprocessing pipeline handling 10TB+ of heterogeneous audio-visual data.
• Production inference system built in Elixir, achieving sub-second response times at scale
• Integration framework with existing B-roll generation and AI director systems
• Solutions for common lip-sync challenges: temporal coherence, cross-identity generalization, and multi-speaker scenarios

Target audience: ML engineers, AI researchers, and developers building content creation tools, video processing systems, or scaling AI models for consumer applications.