Generating 3D Shoppable Product Visualizations with Veo Video Generation Model

google · E-commerce · 2025 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: 3D product visualization for online shopping

Google wanted to make online product shopping feel more tangible and interactive by generating 360-degree product visualizations from only a few product images, while reducing the cost, manual effort, and reliability problems of earlier 3D reconstruction approaches.

Constraints:

• online visualization tools can be costly and time consuming for businesses to create at scale
• the system needs to work from sparse product imagery rather than controlled full-coverage scans
• thin structures such as sandals and heels are difficult to reconstruct from limited views
• earlier methods suffered from inaccurate camera poses and ambiguity from sparse input views
• the generated views must adhere to product images while filling unseen sides
• the solution needed to generalize beyond shoes to furniture, apparel, electronics, and other categories

Architecture

Google evolved its 3D shoppable-products system from NeRF-based reconstruction, to a view-conditioned diffusion prior with score distillation sampling, to a Veo-based video-generation approach. The latest architecture fine-tunes Veo on millions of high-quality synthetic 3D assets rendered from varied camera angles and lighting conditions, producing paired image/video training data so Veo can generate consistent 360-degree product spins from one or more product images.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  synthetic_3d_asset_dataset["Synthetic 3D asset dataset"] --> rendering_pipeline["Rendering pipeline"]
  rendering_pipeline --> paired_image_and_video_dataset["Paired image and video dataset"]
  paired_image_and_video_dataset --> fine_tuned_veo_model["Fine-tuned Veo model"]
  product_images["Product images"] --> fine_tuned_veo_model
  fine_tuned_veo_model --> three_sixty_degree_product_visualization["360-degree product visualization"]
  three_sixty_degree_product_visualization --> google_shopping_surface["Google Shopping surface"]
  google_shopping_surface --> online_shoppers["Online shoppers"]

Components

Data flow

Product images enter a generation pipeline that conditions a fine-tuned Veo model. The training process starts with millions of curated 3D synthetic assets, renders them under multiple camera angles and lighting conditions, and converts them into paired image/video examples. At inference time, Veo receives one or more product images and generates a consistent 360-degree video visualization that can be surfaced as an interactive product view on Google Shopping.

Why this architecture

Google moved to fine-tuned Veo because earlier 3D reconstruction pipelines depended on difficult camera-pose estimation and struggled with sparse views, thin geometry, lighting, and material interactions. Veo's diffusion-based video generation could be trained on synthetic paired image/video data, simplify the inference problem, and generalize across product categories while improving realism from a small number of product images.

Models used: Neural Radiance Fields · view-conditioned diffusion model · DreamFusion score distillation sampling variant · Veo

Scale

• The product experience is aimed at billions of people who shop online every day.
  “Every day, billions of people shop online”article, stated
• The first-generation NeRF system needed five or more product images.
  “from five or more images of the product.”article, stated
• The Veo training setup used millions of high-quality 3D synthetic assets.
  “we first curated a dataset of millions of high quality, 3D synthetic assets.”article, stated
• The latest approach can generate from one image, with as few as three images improving quality and reducing hallucinations.
  “with one image, you can generate a realistic, 3D representation of the object.”article, stated

  “as few as three images capturing most object surfaces are sufficient to improve the quality of the 3D images and reduce hallucinations.”article, stated
• The second-generation approach scaled to many shoes viewed daily on Google Shopping.
  “generate 3D representations for many of the shoes viewed daily on Google Shopping.”article, stated

Alternatives considered (roads not taken)

NeRF-based 3D reconstruction from five or more images

Rejected because: The NeRF approach required many subproblems and suffered from inaccurate camera poses and ambiguity from sparse input views.

View-conditioned diffusion plus score distillation sampling

Rejected because: The second-generation method scaled footwear coverage, but Veo generalized better across categories and handled complex material and lighting interactions.

Estimate precise camera poses from sparse product photos

Rejected because: Fine-tuned Veo simplified the problem by avoiding precise pose estimation from sparse object-centric images.

Depend on a single input image whenever possible

Rejected because: The article says one image can work, but more input views improve fidelity and reduce hallucination of unseen object sides.

Patterns

fine-tuning-adaptation (model adaptation)

Google fine-tuned Veo on paired image/video data so a general video generation model could produce product-specific 360-degree spins from product images.

synthetic-data-generation (training data construction)

The team used synthetic 3D assets rendered from many camera angles and lighting conditions to produce supervised examples for product-spin generation.

data-quality-loop (input coverage and fidelity)

Google improved generation quality by recognizing when sparse views caused ambiguity and by using additional product-surface coverage to reduce hallucinated unseen details.

Tradeoffs

Use Veo-based video generation as the third-generation architecture.

Chose A fine-tuned Veo model for 360-degree product videos over NeRF and view-conditioned diffusion pipelines as the primary approach.

Rationale: Veo better handled lighting, materials, textures, geometry, and generalization across product categories while simplifying pose-estimation complexity.

Train with synthetic 3D assets.

Chose Curated synthetic 3D assets rendered into paired image/video examples over Relying only on real sparse product photos for supervision.

Rationale: Synthetic assets let Google render controlled angles and lighting conditions at scale for supervised Veo training.

Accept sparse-image generation while encouraging more surface coverage.

Chose Generation from one or more images, with three images preferred for quality over Requiring a large full-coverage image set for every product.

Rationale: One image can generate a plausible product view, but three images covering most surfaces reduce hallucinated unseen details.

Gotchas (what bit them)

Sparse product images made 3D reconstruction ambiguous.

Root cause: The first-generation approach depended on noisy signals such as camera poses and only a handful of object-centric views.

Resolution: Google moved from NeRF reconstruction to diffusion-based and then Veo-based novel-view generation that could infer missing views more reliably.

Thin and complex footwear geometry exposed the limits of early reconstruction.

Root cause: Sandals and heels have thin structures and complex geometry that are hard to infer from a few photos.

Resolution: The later diffusion and Veo approaches improved novel-view synthesis for footwear and then generalized beyond footwear categories.

A generative model can hallucinate unseen sides of the product.

Root cause: When only one view is available, the model has no direct evidence for surfaces such as the back of the object.

Resolution: Google uses more input images when available, finding that three images covering most surfaces improve fidelity and reduce hallucination.

Precise camera-pose estimation added brittle complexity.

Root cause: Earlier pipelines needed to estimate camera positions from sparse product photos before optimizing a 3D representation.

Resolution: Fine-tuned Veo avoided the need for precise camera poses from sparse images, simplifying the problem and improving reliability.

Evaluation

The article does not describe a formal benchmark or numeric eval harness. It reports qualitative and production evidence: first-generation weaknesses, second-generation scaling for footwear, Veo generalization across categories, better material/lighting handling, reduced pose-estimation complexity, and reduced hallucination with three surface-covering images.

Metrics: ability to generate novel views · adherence to available product images · quality and realism · lighting and material interaction quality · category generalization · hallucination reduction from additional input views

Outcomes

• Google launched interactive 360-degree shoe visualizations on Google Search as the first milestone.
• The second-generation approach scaled to many shoes viewed daily on Google Shopping.
• Fine-tuned Veo generalized across furniture, apparel, electronics, and other categories.
• Veo reduced reliance on precise camera-pose estimation from sparse images.
• As few as three input images improved product-view quality and reduced hallucinated unseen details.

The case spans a multi-year production evolution from a 2022 NeRF milestone to a 2023 diffusion-prior approach and a 2025 Veo-based approach. It targets online shopping at Google scale and uses millions of synthetic 3D assets for training data.