Ecommerce ops · Production

Google uses Veo generative AI to create shoppable 3D product visualizations from 2D product images for Google Shopping

The problem

Online shopping cannot replicate the tactile, hands-on experience of physical stores, and creating high-quality 3D product visualization tools at scale has been costly and time-consuming for businesses.

First attempt

The first-generation NeRF approach suffered from noisy camera pose signals and struggled with thin-structured products like sandals and heels; both early approaches required estimating precise camera poses from sparse images.

Workflow diagram · grounded in source
1
Product images submitted
trigger
“create high quality and shoppable 3D product visualizations from as few as three product images”
2
Veo generates 360° spin
ai_action
“supervised Veo to generate 360° spins conditioned on one or more images”
3
Novel views with lighting capture
ai_action
“Veo was not only able to generate novel views that adhered to the available product images, but it was also able to capture complex lighting and material interactions (i.e., shiny surfaces)”
4
Interactive 3D views published
output
“This technology is already enabling the generation of interactive 3D views for a wide range of product categories on Google Shopping”
Reported outcome

The Veo-based third-generation approach generates interactive 3D views from as few as three product images, generalizes across furniture, apparel, and electronics, and is already deployed on Google Shopping without requiring precise camera pose estimation.

Reported metrics
Minimum product images requiredthree
Training dataset sizemillions of high quality, 3D synthetic assets
Hallucination reductionreduce hallucinations
Reported stack
VeoNeural Radiance Fields (NeRF)view-conditioned diffusion model
Source
https://research.google/blog/bringing-3d-shoppable-products-online-with-generative-ai/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The Veo-based third-generation approach generates interactive 3D views from as few as three product images, generalizes across furniture, apparel, and electronics, and is already deployed on Google Shopping without re…

What tools did this team use?

Veo, Neural Radiance Fields (NeRF), view-conditioned diffusion model.

What results were reported?

Minimum product images required: three; Training dataset size: millions of high quality, 3D synthetic assets; Hallucination reduction: reduce hallucinations (source-reported, not independently verified).

What failed first in this deployment?

The first-generation NeRF approach suffered from noisy camera pose signals and struggled with thin-structured products like sandals and heels; both early approaches required estimating precise camera poses from sparse…

How is this ecommerce ops AI workflow structured?

Product images submitted → Veo generates 360° spin → Novel views with lighting capture → Interactive 3D views published.