EchoMimicV2

EchoMimicV2, from Ant Group's Alipay Terminal Technology department, goes beyond talking heads: given a reference image, an audio clip, and a hand-pose sequence, it generates half-body human animation with synchronized speech, gestures, and facial expression. Its Audio-Pose Dynamic Harmonization strategy (pose sampling plus audio diffusion) reduces redundant control conditions, and it was accepted at CVPR 2025.

Inference code and pretrained weights (English and Mandarin) shipped November 21, 2024 via GitHub antgroup/echomimic_v2 and Hugging Face. Limitations: it needs a driving pose sequence on top of image and audio, the ~12GB model weights want a 24GB GPU, and the project is explicitly for academic research with a responsible-use disclaimer — not a production avatar SDK.