EchoMimic

EchoMimic is Ant Group's lifelike portrait-animation series. V1 ('Lifelike Audio-Driven Portrait Animations through Editable Landmark Conditioning', July 2024) introduced a then-novel idea: train on audio and facial landmarks together, so a portrait can be driven by either signal or a blend of both, with editable landmarks for extra control. V2 moved to striking semi-body human animation, and V3 (a 2025 paper accepted to AAAI 2026) unified multi-modal, multi-task human animation in a compact 1.3B-parameter model built on Wan2.1-Fun. The whole series is Apache-2.0 with code and weights public, and the repos are actively maintained with 4.3k stars on V1.

Because it is fully open, EchoMimic has become one of the most re-hosted talking-head models around: community Hugging Face Spaces, ComfyUI nodes, and inference services (including a hosted echomimic-v3 skill endpoint) all build on it. The authors claim superiority over prior methods on public and collected datasets, but those are author-reported numbers with no independent benchmark. For anyone building an avatar feature without paying per-minute API fees, it remains one of the most practical open options.

Author-reported superiority over prior methods on public datasets and a collected dataset (arXiv 2407.08136); no independent published benchmarks