OmniCustom AI Generates Custom Videos With Matching Audio in One Click

Researchers have unveiled OmniCustom, a new AI model that synchronously customizes both video and audio from single references. Given one reference image and one audio clip, it generates a video where the subject maintains the reference's visual identity and vocal timbre, while speaking any text prompt. Built on a Diffusion Transformer framework with specialized LoRA modules and contrastive learning, it was trained on a large-scale human dataset and outperforms existing methods in fidelity.