Hi3DGen

Hi3DGen estimates surface directions in a normal map, then uses that map to generate a mesh. Consider it for source geometry that will pass through sculpting, material work and game-ready cleanup. Fine detail is the attraction, but generated details can disagree with the input. Check thin structures and hidden surfaces against your reference: a sharper ridge only helps if it belongs there. For mobile assets, plan to simplify geometry or bake useful detail into textures. Materials and rigging remain separate work, and the cited local demo has a checkpoint-path mismatch to resolve. Generation speed, memory needs and output fidelity have not been independently tested here.

Commercial use & license

ByteDance's project code is MIT, with copyright and permission notices required in redistributed copies. The pinned NiRNE submodule has no top-level license; its licensing scope and the exact checkpoint terms need clarification before production use.

Adult content rules

The MIT project license does not grant rights to reference images, people's likenesses or third-party assets. Adult-content permission is not established for the full dependency chain: the pinned NiRNE submodule lacks a top-level license, and exact checkpoint terms still need review. Any hosted deployment adds its own terms.

Read the policy ↗

What you need

The cited setup uses CUDA, PyTorch 2.4.0, torchvision 0.19.0, spconv 2.3.6 and xformers 0.0.27.post2, plus a recursive clone for NiRNE. Resolve the demo's normal-checkpoint path mismatch and reconcile dependency versions before running it. No universal VRAM minimum is verified.

Ranks are recommendations for the task described, based on reviewed documentation. Model ranks cover non-explicit development work. We have not run practical benchmarks. Check content policies and commercial terms separately. How the ranks work →