Nekodeus/triposr-ov

TripoSR — OpenVINO IR Models

OpenVINO IR conversion of TripoSR for fast 3D reconstruction on Intel CPUs and GPUs.

Conversion Details

Parameter	Value
Tool	openvino.convert_model() via ONNX intermediate
ONNX opset	18
Precision	FP16
Source	TripoSR ONNX checkpoints (nerf.onnx, triplane.onnx)

Why OpenVINO?

• 2-5x faster inference on Intel CPUs vs PyTorch/ONNX Runtime
• Intel Arc GPU + iGPU acceleration, no CUDA/NVIDIA required
• Lower memory with FP16 IR weights
• Deploy anywhere with openvino runtime only

Components

File	Description
triposr_triplane	Triplane decoder (main 3D reconstruction)
triposr_nerf	NeRF renderer

Usage

import openvino as ov
core = ov.Core()
model = core.read_model("triposr_triplane.xml")
compiled = core.compile_model(model, "GPU")  # or "CPU"
out = compiled({"input": image_tensor})[0]