Hi3D

Hi3D is a 3D geometry refinement project open-sourced on GitHub by the Hi3D team, released as Hi3D v1.0. The README says it refines a coarse mesh into high-resolution geometry, conditioned on a reference image of the same object, and builds on the Sparc3D representation and refinement architecture. Refinement can run in stages at 512³, 1024³ and 1536³. It fills in detail by looking at both an image and a rough model at once. It produces geometry only, without textures, and the repository contains inference code only.

Who it's for:3D developers and researchers with NVIDIA GPUs and Linux. An ordinary computer can't run it, and readers who don't write code won't get much from it.

Before you start: it needs Linux, Python 3.10, an NVIDIA GPU with compute capability 8.0 or higher and CUDA 12; there is no CPU path. The authors only verified the top tier on an H20 with 96GB of memory and don't say which tiers smaller cards can reach. The first run downloads models from Hugging Face. Full training and evaluation reproduction isn't provided. The model and code are under the Hi3D Open License Agreement 1.0, which allows commercial use; once you reach 200,000 monthly active users or US$1 million in monthly revenue you must notify them within 60 days and apply for a separate written licence, and commercial integrations must show wording such as "Powered by Hi3D1.0". You are responsible for the rights to any coarse models and third-party assets you use.

Based on the project's public documentation. Not tested by AI Inspiration World.