Pixel-Based Vision-Language Models Without Separate Text Encoders

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional machine-learned models for multimodal vision-language tasks require separate image and text encoders, leading to high computing resource costs and inaccuracies due to language-specific text encoders and extensive pre-processing, such as tokenization and translation.

Innovation Solution

A pixel-based machine-learned encoding model processes textual content as an image, eliminating the need for separate encoders and pre-processing, using a single model to generate accurate embeddings through contrastive learning.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If separate image and text encoders are used for multimodal vision-language tasks, then the model can process different modalities, but the computing resource cost increases and accuracy decreases due to language-specific text encoders and extensive pre-processing

Engineering Contradiction:
Improvemultimodal processing capabilityVSAvoidcomputing resource cost
Core Design Contradiction:
Adaptability or versatilityVSUse of energy by moving object

Solution Approach 1:

The patent merges image encoding and text encoding into a single unified encoder that processes both modalities through the same computational graph. The encoder takes either image inputs or text inputs and produces corresponding embeddings through shared layers, eliminating the need for separate image and text encoders and reducing computing resource requirements while maintaining multimodal processing capability

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The encoder is designed as a universal model that can process multiple modalities (images and text) through the same architectural framework. The model uses modality-specific projection layers that map different input types into a common embedding space, allowing a single model to perform vision-language tasks without requiring language-specific text encoders or extensive pre-processing

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

2Adaptability or versatility

If separate image and text encoders are used for multimodal vision-language tasks, then the model can process different modalities, but the model accuracy decreases due to language-specific text encoders and extensive pre-processing

Engineering Contradiction:
Improvemultimodal processing capabilityVSAvoidembedding accuracy
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent merges image encoding and text encoding into a single unified encoder that processes both modalities through the same computational graph. The encoder takes either image inputs or text inputs and produces corresponding embeddings through shared layers, eliminating the need for separate image and text encoders and reducing computing resource requirements while maintaining multimodal processing capability

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The model uses learnable projection parameters that adaptively map different modalities into a common embedding space. The projection layers adjust the dimensional transformation for each modality, allowing the single encoder to accurately represent both images and text in the same vector space without requiring language-specific encoders or extensive pre-processing

Inventive Principle:
Principle #35Parameter changes

3Adaptability or versatility

If conventional encoders are used with tokenization and translation, then the model can handle different languages, but the pre-processing complexity and time consumption increase

Engineering Contradiction:
Improvelanguage handling capabilityVSAvoidpre-processing time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent extracts and removes the tokenization and translation pre-processing steps from the conventional encoder pipeline. By using a single encoder that directly processes text inputs without requiring tokenization or language translation, the model eliminates these time-consuming operations while maintaining the ability to handle multiple languages through the shared embedding space

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The model creates a simplified version of text processing that bypasses complex pre-processing operations. Instead of tokenizing and translating text through multiple stages, the encoder directly processes text inputs and projects them into the embedding space, effectively copying the essential function of language handling while removing the time-consuming pre-processing steps

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4375950B1Pixel-based machine-learned models for multimodal vision-language tasks
Publication Date: 2025.09.03 GOOGLE LLC
  • EP4375950B1 patent drawingFigure 1A
  • EP4375950B1 patent drawingFigure 1B
  • EP4375950B1 patent drawingFigure 1C

AI summary

A first image and textual content associated with the first image is obtained. A second image that depicts the textual content associated with the first image is rendered. The first image and the second image are processed with a machine-learned encoding model to respectively obtain a first image embedding and a second image embedding for an image embedding space including a plurality of image embeddings. The machine-learned encoding model is trained based on a difference between the first image embedding and the second image embedding.