Introduction

Hi3D just open-sourced Hi3D v1.0, an image-conditioned 3D refiner that takes a coarse mesh plus one reference image and rebuilds it as high-resolution geometry, up to an effective 1536³ sparse extraction grid. The weights are on Hugging Face, inference code is on GitHub, and the team announced the drop on

on October 23.

This is the model that previewed as Sparc3D. It isn't a full image-to-3D generator by itself. It's the second stage you bolt on after one: feed it the blobby mesh from Hunyuan3D, TRELLIS, or your own blockout, and it carves in the panel lines, teeth, scales, and moldings that first-pass generators smear out.

What shipped

  • Model: Hi3D v1.0 (geometry refinement built on Sparc3D)
  • Vendor: Hi3D (Oceano HK Holding Limited)
  • Modality: image + coarse mesh to refined 3D mesh (geometry only)
  • Access: open weights under the Hi3D Open License Agreement 1.0; inference code included, training code and eval harness are not
  • Where to run it today: locally from the GitHub repo on an NVIDIA GPU, or the hosted Hi3D app at hi3d.ai if you don't have the hardware
  • Weights: two 3.68 GB transformers (transformer_512, transformer_1024) plus a 0.38 GB sparse VAE
  • Training data: about two million curated 3D assets, per the project report

How the refinement ladder works

Hi3D samples sparse latents on coordinates taken from your coarse mesh, decodes them with a self-pruning sparse VAE, extracts a mesh, then revoxelizes that mesh as support for the next pass. You pick how far up the ladder to climb:

Requested resolutionPassesTransformer used

512³

1

transformer_512

1024³

2

transformer_512, then transformer_1024

1536³

3

transformer_512, then transformer_1024 twice

Each pass runs 25 scheduler steps by default. The repo is upfront that these are nominal sparse grid sizes, not a dense 1536³ volume or a fixed triangle count.

Install and run

You need Linux, Python 3.10, an NVIDIA GPU with compute capability 8.0 or newer, and a CUDA 12.0–12.5 toolkit. There's no CPU path, and flash-attn has no fallback.

git clone https://github.com/hi3d-official/Hi3Dv1.0.git
cd Hi3Dv1.0
pip install torch==2.5.0+cu124 torchvision==0.20.0+cu124 \
  --extra-index-url https://download.pytorch.org/whl/cu124
pip install packaging ninja psutil
pip install --no-build-isolation flash-attn==2.7.4.post1
pip install -c docker/constraints.txt -e ".[inference]"

Then refine any coarse mesh you already have:

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

Add --sequential-cpu-offload to keep less of the model on the GPU. If you want the full image-to-mesh path, demo/image_to_mesh.py ships an optional adapter that uses Tencent's Hunyuan3D-2 mini as the coarse generator (you install that code and its weights separately under Tencent's license).

What game artists should know

  • Hardware reality check. The 1536³ ladder was only validated on a 96 GB NVIDIA H20. The team hasn't published minimum VRAM, so treat 1024³ as the sensible first try on smaller cards.
  • Geometry only. No textures, materials, or vertex colors. You still need a texturing pass (Meshy, Scenario, Hunyuan3D-Paint, or hand work), plus retopo and UVs before anything goes in an engine.
  • Garbage in, garbage out. Big structural errors in the coarse mesh survive refinement, and hidden sides are a plausible guess, not a reconstruction.
  • Scale is reset. Output is recentered and normalized to unit radius, so re-scale on import.

License, in plain terms

Commercial use is allowed. If you integrate Hi3D into a commercial product or service, you have to show "Powered by Hi3D 1.0" in the UI, About page, or docs. If a related product passes 200,000 monthly active users or USD 1,000,000 in monthly revenue, you must notify Hi3D and apply for a separate license within 60 days. Plain output meshes that don't contain Hi3D weights or code don't trigger the branding or scale terms, so shipping refined props in your game is the low-friction case.

Conclusion

Hi3D v1.0 is a focused tool: it won't invent a character from a prompt, but it gives indie teams an open, local way to push AI-generated or blocked-out meshes from "mushy" to "sculpted" before texturing. If you've been running Hunyuan3D or TRELLIS and fighting soft detail, this is the refinement pass to test this weekend, as long as you have a big-memory GPU.

—Titus