Hi3D v1.0 — What It Does

Hi3D v1.0 is a 3D geometry refinement tool. You give it two things:

1. A coarse 3D mesh (a rough, low-detail 3D model of an object)
2. A reference image of that same object

It produces a high-resolution, more detailed 3D mesh — adding fine surface details like layered relief, architectural moldings, curved boundaries, scales, teeth, and concentric grooves.

It is built on top of an earlier system called Sparc3D, and keeps Sparc3D's internal names (sparc3d, Sparc3DPipeline) for compatibility.

How It Works

The system works in passes. Each pass refines the mesh at a higher resolution:

Pass	Structural grid	Extraction grid
1	64³	512³
2	128³	1024³
3	192³	1536³

• A 512³ request runs one pass.
• A 1024³ request runs two passes.
• A 1536³ request runs all three.

Each pass: samples sparse data points, decodes them with a neural network, extracts a mesh, then re-voxelizes (re-grids) that mesh to feed the next pass. The reference image guides the refinement throughout.

What You Get

• Geometry only. No textures, materials, or vertex colors.
• Output is recentered and normalized to unit maximum radius — original world coordinates and scale are not preserved.
• Hidden surfaces (not visible in the image) are plausible completions, not exact reconstructions.

Requirements

• Linux, Python 3.10
• NVIDIA GPU with compute capability 8.0 or newer (e.g., H20, A100-class)
• CUDA toolkit 12.0–12.5
• No CPU inference path exists
• The 1536³ setting has only been validated on a 96 GB NVIDIA H20; memory needs on smaller GPUs are unknown

What's Included / Not Included

Included:

• Inference code (Python package + command-line tool)
• Model artifacts on Hugging Face (8 files: configs + model weights)
• An optional adapter for Hunyuan3D-2 mini (a third-party coarse-mesh generator)

Not included:

• Training code
• The paper's evaluation harness
• Benchmark splits or comparison meshes

How to Use It

Python API:

from PIL import Image
import trimesh
from sparc3d import Sparc3DPipeline

pipeline = Sparc3DPipeline.from_pretrained(
    "hi3d-official/Hi3Dv1.0",
).to("cuda")

result = pipeline(
    Image.open("input.png"),
    trimesh.load("coarse.glb", force="mesh", process=False),
    resolution=1024,
    num_inference_steps=25,
    seed=0,
)
result.mesh.export("refined.glb")

Supported resolutions: 512, 1024, 1536. Default: 25 scheduler updates per pass.

Command line:

python demo/refine_mesh.py \
  --model hi3d-official/Hi3Dv1.0 \
  --image input.png --coarse-mesh coarse.glb --output refined.glb \
  --resolution 1024 --num-inference-steps 25 --seed 0

Key Limitations & Caveats

• Coarse structure matters. Serious errors in the input coarse mesh can persist in the output.
• Seeds are not guarantees. GPU operations are not bitwise deterministic; same seed + same inputs may produce slightly different meshes.
• GPU kernels are mandatory. flash-attn has no fallback; it must compile successfully.
• Resolution is a request, not a promise. If the target resolution can't be reached, the code raises an error rather than silently returning lower quality.
• Background removal is opt-in (requires installing the demo extra).

License & Commercial Terms

Governed by the Hi3D Open License Agreement Version 1.0:

• Commercial use is allowed.
• If a product using it exceeds 200,000 monthly active users OR USD 1,000,000 monthly revenue, you must notify the licensor and apply for a separate written license within 60 days.
• Commercial integrations must display "Powered by Hi3D 1.0" linking to the official website.
• Standalone outputs (meshes that contain no Hi3D materials or model derivatives) do not trigger branding or scale-license obligations.
• Existing Apache-2.0 and MIT terms still apply to original/third-party portions.
• Third-party components (Hunyuan3D-2, DINOv2, CLIP) keep their own licenses.

Quick Facts

• Release date: October 23, 2026
• Training data: ~2 million curated 3D assets (per the technical report)
• Live demo: Available on Hugging Face Spaces
• Contact: contact@hi3d.ai