How AIGC Is Transforming 3D Model Generation for E‑commerce

Background

AIGC (Artificial Intelligence in Graphics and Computing) combines AI and computer graphics to generate 3D assets more efficiently, accurately, and flexibly. It addresses the low efficiency, limited precision, and poor personalization of traditional 3D modeling pipelines, which is critical for large‑scale e‑commerce platforms that need massive, personalized 3D models.

Key AIGC 3D Generation Models

NeRF (Neural Radiance Fields)

NeRF, introduced by Google in 2020, uses a small fully‑connected network (~5 MB) that maps 5‑D inputs (spatial position + view direction) to color and density. Volume rendering then synthesizes novel views without explicit geometry. Industrial adaptations focus on improving cloth detail and text clarity for e‑commerce use cases.

Point‑E

OpenAI’s Point‑E (2022) generates 3D point clouds in 1–2 minutes on a single GPU. It first creates a synthetic view with a text‑to‑image diffusion model, then conditions a diffusion model on that view to produce the point cloud. Speed is 10–100× faster than prior methods, though sample quality is lower.

https://github.com/openai/point-e

Shap‑E

Shap‑E extends Point‑E by directly generating implicit‑function parameters that can be rendered as textured meshes or NeRFs. The pipeline trains an encoder to map 3D assets to latent codes, then a conditional diffusion model on those codes. On large paired 3D‑text datasets it produces diverse assets in seconds with higher fidelity.

https://github.com/openai/shap-e/tree/main

DreamFusion

DreamFusion (Google) iteratively refines a NeRF using a 2D text‑to‑image model (e.g., Imagen). Multi‑view images are generated from the text prompt, then the NeRF is updated to improve view consistency. A single model requires ~15,000 iterations, consuming about 1.5 hours on four TPUv4 chips.

https://github.com/ashawkey/stable-dreamfusion

Magic3D

Magic3D (Nvidia) builds on DreamFusion with a two‑stage pipeline: a diffusion model first creates a low‑resolution hash‑grid 3D representation, then traditional graphics techniques up‑sample and render the model. This yields higher resolution, better visual quality, and faster generation suitable for industrial deployment.

Practical Implementation

Experiments were run on a workstation equipped with:

GPU: Nvidia RTX 3060 12 GB

CPU: Intel i9‑13900KF

RAM: 64 GB

Using a Jupyter Notebook and CUDA, the Shap‑E model generated a single 3D asset in roughly 5 minutes. The resulting meshes showed limited detail, indicating a trade‑off between speed and quality.

https://github.com/openai/shap-e/tree/main

Challenges in AIGC‑Driven 3D Generation

Data quality: Inaccurate or incomplete training data leads to defects and misalignments in generated models.

Computational resources: High‑resolution synthesis demands substantial GPU memory and compute, limiting accessibility.

Texture mapping: Obtaining perfectly matching textures for real‑world objects remains difficult, often causing seams or unnatural appearance.

Model interpretability: Generated assets can be ambiguous or hard to understand, hindering downstream use.

References

https://github.com/awesome-NeRF/awesome-NeRF

https://zhuanlan.zhihu.com/p/613679756

(Taichi NeRF Part 2) https://zhuanlan.zhihu.com/p/612102573 (Taichi NeRF Part 1)

https://www.zhihu.com/search?type=content&q=DreamFusion