A user has created a distilled version of the LivePortrait ONNX model that can run entirely within a web browser using WebGPU. The original implementation took 30 seconds to generate a single frame, but this new proof-of-concept reduces generation time to less than 30ms on an RTX 5090.
- The distilled model achieves approximately 25fps performance in the browser.
- It runs 100% client-side without requiring server-side processing.
- The model was trained using a small dataset for only a few hours.
- Output quality is described as "just ok" and varies depending on the input portrait.
This development demonstrates that LivePortrait can be significantly optimized for real-time, local inference in web environments.