Step1X-3D

Step1X-3D takes an image through separate geometry and texture stages, using geometric conditioning and synchronization to improve consistency between views. That makes it a useful local candidate for prop prototypes, with a mesh you can inspect before texturing. The demanding CUDA setup is only part of the commitment: its default texture workflow includes a noncommercial rendering dependency. Resolve that restriction before planning production around it. A watertight shape still needs topology that bends properly under a character rig; polished textures cannot fix that. Start with a prop whose silhouette and materials you can judge, then check the exported GLB in your engine. Character anatomy and cleanup effort remain untested here.

Commercial use & license

Project code is Apache-2.0, and the checkpoint card declares the same license. The texture stack also loads SDXL under Open RAIL++-M. Its nvdiffrast library permits users outside NVIDIA and its affiliates to use it only for noncommercial research or evaluation. Copied Hunyuan renderer files retain conflicting noncommercial/community notices; Tencent's current 2.0 license excludes the EU, UK and South Korea. Resolve these component terms before commercial use or distribution.

Adult content rules

SDXL prohibits unlawful use, exploiting or harming minors, and harassment. Copied Hunyuan components have separate territorial and social-ethics restrictions. Apache-2.0 grants no rights to reference images or people's likenesses, and this mixed stack has no blanket adult-use clearance. Hosted policies are unverified.

Read the policy ↗

What you need

The documented environment uses Python 3.10, CUDA 12.4 and PyTorch 2.5.1, plus compiled rendering extensions. The repository reports 27 GB GPU memory for standard geometry plus texture, and 29 GB for label geometry plus texture; these are example usages, not verified minimums.

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 →