On February 13, 2026, the AI team at Soul App proudly declared the open-sourcing of their cutting-edge real-time digital human generation model, SoulX-FlashHead. This remarkable model boasts 1.3 billion parameters and is capable of achieving an impressive inference frame rate of 96 FPS on a single, consumer-grade graphics card, such as the RTX 4090. What's more, it operates efficiently with just 6.4GB of VRAM and can support up to three concurrent streams, making it a versatile and powerful tool. For those seeking even higher image quality, SoulX-FlashHead also offers a Pro version optimized for RTX 5090 graphics cards, where its FID (Fréchet Inception Distance) and Lip-sync metrics reach SOTA (State-of-the-Art) standards. At present, both the model weights and the code have been made fully available to the public.
