WebGPU support for Vision Tasks

[pgn-dev]

MediaPipe Solution (you are using)

FaceLandmarker, HandLandmarker, PoseLandmarker

Programming language

typescript

Are you willing to contribute it

None

Describe the feature and the current behaviour/state

Use WebGPU for running vision tasks. Currently the TaskRunners appears to be running WebGL with OffScreenCanvas

Will this change the current API? How?

No response

Who will benefit with this feature?

Everyone

Please specify the use cases for this feature

Should make prediction on browser-based solutions faster.

Any Other info

No response

[schmidt-sebastian]

Our Vision tasks do use WebGL at the moment, as our WebGPU support is currently focused on our LLM Inference API. Please do watch our announcement over the coming weeks and months in this space.

[pgn-dev]

Thanks for confirming @schmidt-sebastian

Do you expect a significant improvement in latency over WebGL with WebGPU?

TFLite already seems to support WebGPU. Would running the TFLite models from the task files work out of the box?