Nekodeus/hunyuan3d-ov

Hunyuan3D 2.1 — OpenVINO IR Models

OpenVINO IR format conversion of Hunyuan3D 2.1 components for optimized inference on Intel CPUs and GPUs.

What is OpenVINO?

OpenVINO (Open Visual Inference and Neural network Optimization) is Intel's open-source toolkit for optimizing and deploying AI models. It dramatically speeds up inference on Intel hardware (CPU, iGPU, Arc GPU, NPU) compared to vanilla PyTorch.

Why convert to OpenVINO?

Benefit	Detail
CPU speedup	2-5x faster inference on Intel CPUs vs PyTorch
iGPU acceleration	Run on Intel integrated GPUs (no dedicated GPU needed)
Arc GPU support	Optimized for Intel Arc discrete GPUs (A770, A750, etc.)
Lower memory	FP16/INT8 quantization reduces VRAM usage significantly
No CUDA dependency	Deploy anywhere — no NVIDIA GPU or CUDA toolkit required
Smaller models	FP16 weights are ~50% smaller than FP32 PyTorch checkpoints

Conversion Details

Parameter	Value
Tool	openvino.convert_model() via ONNX intermediate
ONNX opset	18
Precision	FP16 (default for OpenVINO IR)
Source weights	dgrauet/hunyuan3d-2.1-mlx
Framework	PyTorch → ONNX → OpenVINO IR

Conversion workflow

1. Load safetensors weights into pure PyTorch modules
2. Export to ONNX (opset 18) via torch.onnx.export()
3. Convert ONNX to OpenVINO IR via ov.convert_model()
4. Save as .xml (graph) + .bin (weights) pair

This ONNX intermediate step is required because ov.convert_model() on large transformers can hang on CPU-only systems. GPU-accelerated ONNX export + OV conversion completes in minutes.

Components

Component	Description	Size
clip_vit	CLIP ViT-L/14 text+image encoder	~580 MB
dinov2_giant	DINOv2-Giant image encoder (painting)	~2170 MB
image_encoder	DINOv2-Large image encoder (shape)	~490 MB
paint_vae_encoder	Paint VAE encoder	~65 MB
paint_vae_decoder	Paint VAE decoder	~95 MB
shape_vae_decoder	Shape VAE decoder (16-block transformer)	~385 MB
dit	DiT shape denoiser (21 double-stream blocks)	~4050 MB
paint_unet	Paint UNet (SD-style, 12ch in / 4ch out)	~1650 MB

Usage (Python)

import openvino as ov

core = ov.Core()
model = core.read_model("dit.xml")
compiled = core.compile_model(model, "GPU")  # or "CPU"

# Run inference
infer_request = compiled.create_infer_request()
infer_request.infer({"x": latent, "t": timestep, "context": text_emb})
output = infer_request.get_output_tensor(0).data

Target Hardware

• Intel Arc A750 (8GB VRAM) — primary target
• Intel iGPU — UHD 630/Iris Xe (reduced resolution)
• Intel CPU — Core i5/i7/i9 (slower but functional)

License

Same as original Hunyuan3D 2.1 — see LICENSE.