hallo2

Why we included this project

Creators and researchers who need a talking-head video from a single portrait and an audio track will find Hallo2 unusually capable: it animates a still face to match the speech, keeping the identity recognizable while pose and expression stay coherent, and it is built to hold that animation over long stretches rather than a few seconds. The showcase includes a 23-minute 4K clip of a speech, real engineering work on long-duration stability and high-resolution output that most one-off demo models never attempt. The repo ships full inference scripts and pretrained weights on Hugging Face, so you can go from a square-cropped frontal portrait and an English WAV file to a finished video without training your own model. That makes it a practical starting point for dubbing or digital avatars, and it fits lecture-style content as well, though the audio has to be English and the pipeline expects an A100-class GPU. Because the code and checkpoints are open, it also works as a baseline for teams building their own portrait-animation pipelines.