3D Diffusion View Synthesis With Stochastic Conditioning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for novel view synthesis of three-dimensional scenes face challenges in achieving 3D consistency and efficiency, particularly in few-shot settings, often requiring hyper-networks or test-time optimization.

Innovation Solution

The proposed 3D diffusion models employ a machine-learned diffusion model with stochastic conditioning and an X-UNet architecture, allowing for improved 3D consistency by randomly sampling conditioning frames from an image set during denoising steps, and sharing weights between input frames.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If existing methods for novel view synthesis are used, then 3D consistency can be achieved, but computational resources and complexity increase due to requirements for hyper-networks or test-time optimization

Engineering Contradiction:
Improve3D consistencyVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent changes the parameter of conditioning by using stochastic sampling of conditioning frames from the image set during denoising steps, rather than using fixed or optimized conditioning parameters. This allows the model to achieve 3D consistency through parameter variation during inference without requiring complex hyper-networks or test-time optimization procedures

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The diffusion model is designed to handle multiple scenes with a single model instance by using weight sharing between input frames and stochastic conditioning. This universal approach eliminates the need for scene-specific hyper-networks while maintaining 3D consistency across different scenes, reducing overall system complexity

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Reliability

If existing methods for novel view synthesis are used, then 3D consistency can be achieved, but training time and resources increase due to hyper-networks or test-time optimization

Engineering Contradiction:
Improve3D consistencyVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent performs preliminary training of a single diffusion model on diverse scene data with stochastic conditioning during the training phase. This preliminary action equips the model with generalizable 3D consistency capabilities that transfer to new scenes without requiring additional test-time optimization or hyper-network training, thus reducing training time and resources

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

Instead of training separate hyper-networks or performing test-time optimization for each scene, the patent uses a single pre-trained diffusion model that can be applied to multiple scenes through stochastic conditioning. This copying approach reuses the same model instance across different scenes, eliminating redundant training procedures

Inventive Principle:
Principle #26Copying

3Productivity

If a single model is used to handle multiple scenes, then computational efficiency improves, but achieving 3D consistency becomes more difficult

Engineering Contradiction:
Improvecomputational efficiencyVSAvoid3D consistency
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent introduces dynamic stochastic conditioning during the denoising steps, where conditioning frames are randomly sampled from the image set at each denoising step. This dynamic approach allows a single model to adapt to different scenes and maintain 3D consistency without requiring scene-specific customization, achieving both computational efficiency and reliability

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The model uses parameter changes through stochastic sampling of conditioning frames during inference. By varying the conditioning parameters dynamically rather than using fixed parameters, a single model can maintain 3D consistency across multiple scenes while improving computational efficiency through weight sharing and eliminating the need for complex hyper-networks

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4548254B1Three-dimensional diffusion models
Publication Date: 2026.03.04 GOOGLE LLC
  • EP4548254B1 patent drawingFigure 1A
  • EP4548254B1 patent drawingFigure 1B~1C
  • EP4548254B1 patent drawingFigure 2

AI summary

Provided are systems and methods to perform novel view synthesis of a three-dimensional (3D) scene with a machine-learned diffusion model. Example implementations of the proposed models may be referred to as "3D Diffusion Models" or 3DiM. The models described herein can be or include an image-to-image diffusion model that takes one or more (e.g., a single) reference views and one or more (e.g., a single) relative poses as input and generates the target view. Thus, the machine-learned diffusion models described herein can perform novel view synthesis from as few as a single image.