Point-E

Point-E is useful for learning how a generated shape becomes a game asset, especially where that process goes wrong. OpenAI's December 2022 release produces coloured point clouds from rendered images or simple text prompts, with a separate model for mesh conversion. The direct text model struggles with complex prompts and unusual objects; noisy points, cracks and incorrect geometry also limit the image models. The supplied mesh notebook exports a vertex-coloured PLY. Materials, rigging, collision and a careful topology check still need work before that becomes a finished asset. Use it as an educational baseline or a contained experiment, and count conversion failures and cleanup time alongside generation. For a new game's production pipeline, the dated release and OpenAI's own advice against commercial use make it a weak starting point. We have not benchmarked output quality, speed or hardware needs.

Commercial use & license

The repository's MIT licence permits commercial software use with copyright and permission notices retained. Verify the exact checkpoints and dependencies separately; reference images, likenesses and third-party assets need their own rights checks.

Adult content rules

The repository's MIT licence has no adult-content restriction, but does not clear reference images, likenesses or third-party assets. Point-E's model card discourages commercial use, and CLIP's card places deployed use outside its intended scope; these are statements of intended use rather than extra MIT conditions. Audit the exact checkpoints and dependencies. Hosted services add separate terms; no blanket adult-content permission is established.

Read the policy ↗

What you need

Python, PyTorch, OpenAI CLIP, checkpoint downloads and a notebook environment for the examples. Use English text prompts: CLIP's model card limits use to English. Examples select CUDA when available and CPU otherwise; we have not tested installation, memory needs or practical CPU performance. Plan mesh conversion, materials and asset checks separately.

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 →