MotionBricks SOMA, float16 and 8-bit ONNX (for the browser)

mmm_v6_soma_q8h.onnx (243 MB): MatMul weights in 8-bit blocks (onnxruntime MatMulNBits, block 32, symmetric), the rest float16. On the reference generation its output differs from float32 by 0.0018 on average (float16: 0.0004; outputs average 0.41), and it runs a little faster in the browser (115-125 ms on an M1 Mac). The live demo uses this.
mmm_v6_soma_fp16.onnx (380 MB): NVIDIA's float16 export.

The SOMA model of NVIDIA MotionBricks, exported to ONNX and converted to float16 so that it runs in a web browser with onnxruntime-web 1.30 on WebGPU (about 145 ms per generation on an M1 Mac). It is used by the live demo in the Space linked below.

Inputs and outputs are the same as the original model (context and target poses in, 64 frames of motion out).
This is a derivative of the original model; it is not made or endorsed by NVIDIA.

Licensed by NVIDIA Corporation under the NVIDIA Open Model Agreement. See NOTICE.txt and LICENSE-NVIDIA-Open-Model-Agreement.txt.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for monoradio/motionbricks-soma-fp16-onnx

Base model

nvidia/MotionBricks

Quantized

(2)

this model