Domain-Specific Video Generation via Dual Diffusion Models

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Adapting large video generative models for domain-specific tasks, such as generating animations or robotic videos, poses significant computational challenges due to the need for fine-tuning pre-trained models, which requires substantial resources and can lead to performance degradation from model overfitting.

Innovation Solution

The method involves using a combination of pre-trained denoising diffusion models and smaller domain-specific models to generate high-quality domain-specific videos by iteratively updating an intermediate representation with noise outputs from both models, eliminating the need for fine-tuning with domain-specific data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If fine-tuning pre-trained models with domain-specific data is performed, then domain-specific video generation capability is improved, but computational resources required increase significantly

Engineering Contradiction:
Improvedomain-specific video generation capabilityVSAvoidcomputational resources
Core Design Contradiction:
Adaptability or versatilityVSUse of energy by moving object

Solution Approach 1:

The model is segmented into two distinct components: a pre-trained base model that handles general video generation tasks, and a domain-specific adapter module that handles domain-specific adaptations. This segmentation allows the computationally expensive pre-trained model to be used without fine-tuning, while only the small adapter module requires domain-specific training, significantly reducing computational resource requirements.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The domain-specific knowledge is extracted from the pre-trained model and isolated into a separate adapter module. This extraction allows the main pre-trained model to remain frozen and computationally efficient, while the extracted adapter component captures domain-specific characteristics with minimal training resources.

Inventive Principle:
Principle #2Taking out (Extraction)

2Adaptability or versatility

If fine-tuning pre-trained models with domain-specific data is performed, then domain-specific video generation capability is improved, but model overfitting occurs leading to performance degradation

Engineering Contradiction:
Improvedomain-specific video generation capabilityVSAvoidmodel performance stability
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

By segmenting the model into a frozen pre-trained base and a small trainable adapter, the invention prevents overfitting because the adapter has limited capacity and cannot memorize the entire domain-specific dataset, while the robust pre-trained base provides stable general video generation capabilities.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The adapter module applies local quality adjustments to specific parts of the model architecture that are most relevant to domain-specific tasks, while leaving the rest of the pre-trained model unchanged. This localized adaptation maintains the reliability of the pre-trained components while adding domain-specific capability where needed.

Inventive Principle:
Principle #3Local quality

3Use of energy by moving object

If smaller domain-specific models are used, then computational resources are reduced, but video generation quality decreases

Engineering Contradiction:
Improvecomputational resourcesVSAvoidvideo generation quality
Core Design Contradiction:
Use of energy by moving objectVSManufacturing precision

Solution Approach 1:

The invention merges the strengths of two different model approaches: the high-quality general video generation capabilities of large pre-trained models and the efficiency of smaller domain-specific models. The pre-trained base model provides high-quality generation, while the small adapter module adds domain-specific efficiency, achieving both quality and resource efficiency simultaneously.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The model architecture functions as a composite structure combining the pre-trained base model (providing general video generation quality) and the domain-specific adapter (providing domain adaptation with minimal resources). This composite approach allows the system to achieve video generation quality comparable to large models while using computational resources similar to small models.

Inventive Principle:
Principle #40Composite materials

Data Source

PatentEP4481673A1Generating domain-specific videos using diffusion models
Publication Date: 2024.12.25 GOOGLE LLC
  • EP4481673A1 patent drawingFigure 1
  • EP4481673A1 patent drawingFigure 2
  • EP4481673A1 patent drawingFigure 3

AI summary

A computer-implemented method for generating an output video conditioned on an input comprises: receiving a conditioning input; initializing a current intermediate representation of the output video; and at each of a plurality of iterations, updating the current intermediate representation. The updating comprises: generating a first noise output by processing a first input comprising the current intermediate representation using a first denoising diffusion model conditioned on the conditioning input, wherein the first denoising diffusion model has been trained on first training data comprising a first plurality of training videos; generating a second noise output conditioned on the conditioning input by processing a second input comprising the current intermediate representation using a second denoising diffusion model conditioned on the conditioning input, wherein the second denoising diffusion model (i) is different from the first denoising diffusion model and (ii) has been trained on second training data comprising a second plurality of training videos; generating a combined noise output by combining at least (i) the first noise output and (ii) the second noise output, and updating the current intermediate representation using the combined noise output generated for the iteration.