ROI Prompt Processing for Large Multimodal Model Efficiency

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Large multimodal models (LMMs) inefficiently process entire media files to generate summary passages, leading to unnecessary computing costs and resource usage on regions not necessary for providing the desired summary.

Innovation Solution

A system and method that allows users to specify a region of interest (ROI) within media files, segmenting and tokenizing only those regions, and generating prompts based on these segments to be processed by LMMs, reducing unnecessary computation and enabling higher throughput.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If the LMM processes the entirety of the media file to generate summary passages, then the completeness of information processing is improved, but the computing cost and resource usage increase unnecessarily

Engineering Contradiction:
Improvecompleteness of information processingVSAvoidcomputing cost
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent applies segmentation by dividing the media file into multiple regions and identifying specific regions of interest (ROIs) that are relevant to the query. Instead of processing the entire media file, the system segments it and only processes the identified ROIs, thereby reducing computing cost while maintaining the reliability of information processing for the specific query context.

Inventive Principle:
Principle #1Segmentation

2Loss of information

If the LMM processes the entirety of the media file, then all possible information is analyzed, but the processing time and computational resources are wasted on unnecessary regions

Engineering Contradiction:
Improveinformation coverageVSAvoidprocessing time
Core Design Contradiction:
Loss of informationVSLoss of time

Solution Approach 1:

The patent applies the extraction principle by identifying and extracting only the relevant regions of interest from the media file that are necessary to answer the specific query. The system uses the query context to determine which regions contain the needed information, extracts those regions, and processes only them, thereby preventing information loss for the specific query while significantly reducing processing time.

Inventive Principle:
Principle #2Taking out (Extraction)

3Reliability

If the entire media file is processed, then comprehensive analysis is achieved, but the computational efficiency and throughput decrease

Engineering Contradiction:
Improvecomprehensive analysisVSAvoidcomputational efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent applies local quality by applying different processing strategies to different regions of the media file. Regions identified as relevant to the query receive full processing attention, while other regions are excluded or given minimal processing. This localized approach maintains comprehensive analysis for the specific query context while significantly improving computational efficiency and throughput.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP4672026A1Region of interest prompt processing for large multimodal models
Publication Date: 2025.12.31 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP4672026A1 patent drawingFigure 1
  • EP4672026A1 patent drawingFigure 2
  • EP4672026A1 patent drawingFigure 3~5

AI summary

A method for processing a multimodal prompt. The method includes receiving a multimodal prompt including a media file and information related to a region of interest (ROI) of the media file. The method further includes determining a ROI of the media file based on the information related to the media file and generating a plurality of media tiles of interest associated with the ROI. The method further includes encoding the plurality of media tiles of interest and using a large multimodal model (LMM) to process the encoded plurality of media tiles of interest according to a natural-language input of the prompt to generate a response.