Single Inference Model With Source-Specific Parameters for Attribution

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

The challenges of automating media content generation and modification, ensuring accurate attribution, and addressing issues of plagiarism and copyright infringement in a digital landscape marked by overwhelming information volumes and content reuse are significant.

Innovation Solution

Systems and methods utilize a single inference model trained on different portions of a training set to generate and modify media content, attribute content to sources, and identify prompts, while providing diverse visual and audio content generation and background replacement.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If a single inference model is used to generate diverse media content, then device complexity is reduced, but manufacturing precision deteriorates because the model must handle multiple source styles simultaneously

Engineering Contradiction:

Improvemodel complexityVSAvoidcontent generation precision

Core Design Contradiction:

Device complexityVSManufacturing precision

Solution Approach 1:

The training set is segmented into multiple source-specific portions, with each portion dedicated to a particular content source. The single inference model learns to selectively attend to relevant segments based on input prompts containing source identifiers, achieving specialized precision without requiring separate models for each source.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The inference model incorporates source-specific parameters that enable it to adapt its generation characteristics locally based on the identified content source. When a prompt contains a source identifier, the model activates source-specific training examples and parameters to generate content with the appropriate style and quality characteristics for that source.

Inventive Principle:

Principle #3Local quality

2Manufacturing precision

If multiple inference models are trained on different source portions, then content generation precision improves, but device complexity increases

Engineering Contradiction:

Improvecontent generation precisionVSAvoidmodel complexity

Core Design Contradiction:

Manufacturing precisionVSDevice complexity

Solution Approach 1:

A single inference model is designed to perform multiple functions by generating content from different sources based on input prompts. The model universally handles various content sources (e.g., Wikipedia, Wikidata, Visual Genome) within one architecture, eliminating the need for separate specialized models while maintaining generation precision through source-specific parameter activation.

Inventive Principle:

Principle #6Universality (Multi-functionality)

3Measurement precision

If source-identifying keywords are included in prompts, then attribution accuracy improves, but ease of operation deteriorates due to stricter input requirements

Engineering Contradiction:

Improveattribution accuracyVSAvoidprompt input ease

Core Design Contradiction:

Measurement precisionVSEase of operation

Solution Approach 1:

The system performs self-service by automatically identifying and extracting source identifiers from input prompts without requiring manual specification. The inference model analyzes the prompt structure and automatically activates the appropriate source-specific parameters, reducing the burden on users while maintaining accurate attribution.

Inventive Principle:

Principle #25Self-service

4Productivity

If a large training set is used to improve content quality, then productivity increases, but loss of information increases due to difficulty in tracking source-specific characteristics

Engineering Contradiction:

Improvecontent generation qualityVSAvoidsource information loss

Core Design Contradiction:

ProductivityVSLoss of information

Solution Approach 1:

The large training set is segmented into source-specific portions, with each portion clearly associated with its origin. This segmentation preserves source information by organizing training examples according to their source, enabling the model to maintain distinct source characteristics even when processing diverse content from multiple sources simultaneously.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

Source-specific parameters are introduced to preserve local characteristics of each content source within the global training framework. These parameters enable the model to retain and apply source-specific stylistic and structural features when generating content, preventing information loss despite the diversity of the training set.

Inventive Principle:

Principle #3Local quality

Data Source

PatentUS12488186B2Inference based on different portions of a training set using a single inference model

Publication Date: 2025.12.02 BRIA ARTIFICIAL INTELLIGENCE LTD

AI summary

Systems, methods and non-transitory computer readable media for inference based on different portions of a training set using a single inference model are provided. Textual inputs may be received, each of which may include a source-identifying-keyword. An inference model may be a result of training a machine learning model using a plurality of training examples. Each training example may include a respective textual content and a respective media content. The training examples may be grouped based on source-identifying-keywords included in the textual contents. Different parameters of the inference model may be based on different groups, and thereby be associated with different source-identifying-keywords. When generating new media content using the inference model and a textual input, parameters associated with the source-identifying-keyword included in the textual input may be used.