Synthetic Data Generation for Visual Language Model Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing visual language models (VLMs) face challenges in accurately performing multi-modal tasks, particularly in tasks that require counting objects or describing positional relationships in images, due to limited and biased training datasets.

Innovation Solution

The method involves creating a larger and more diverse training database for VLMs by generating training examples using a text-to-image model. This is done by obtaining text descriptions of possible images, generating images from these descriptions, and forming tuples with the images and corresponding text descriptions. These tuples are then used to train the VLM, focusing on difficult multi-modal tasks to improve accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If conventional training datasets are used for VLM training, then the training process is simple and data collection is easy, but the model accuracy on difficult multi-modal tasks (counting objects, describing positional relationships) is insufficient

Engineering Contradiction:
Improvemodel accuracy on difficult multi-modal tasksVSAvoidtraining database size
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent uses text-to-image models to generate synthetic images that copy and replicate complex visual patterns, object arrangements, and spatial relationships from existing training data. These generated images are then paired with their corresponding text descriptions to create augmented training examples, effectively copying the complexity of real-world scenes without requiring additional real image collection

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent performs preliminary actions by using text-to-image models to pre-generate training examples before the main VLM training process. The text descriptions are generated first, then converted to images, creating ready-to-use training tuples that can be immediately incorporated into the VLM training database, avoiding the need for complex post-processing of real images

Inventive Principle:
Principle #10Preliminary action

2Reliability

If larger training databases are created to improve VLM accuracy, then model performance on difficult tasks improves, but computational cost and data generation complexity increase

Engineering Contradiction:
Improvemodel accuracy on counting and positional tasksVSAvoidtraining database formation process
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent introduces text-to-image models as intermediary components that bridge the gap between text descriptions and image data. These intermediaries automatically convert text prompts into corresponding images, simplifying the training database formation process by eliminating the need for manual image annotation or complex image processing pipelines

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The system uses itself to generate training data - the text-to-image models are trained on existing data and then automatically generate new training examples without human intervention. This self-service approach scales efficiently, as the models can generate large numbers of training tuples autonomously, reducing the need for complex manual curation processes

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4546213A1Synthetic data generation for training visual language models
Publication Date: 2025.04.30 DEEPMIND TECH LTD
  • EP4546213A1 patent drawingFigure 1
  • EP4546213A1 patent drawingFigure 2
  • EP4546213A1 patent drawingFigure 3

AI summary

A method is proposed for generating a visual language model (VLM) neural network trained to perform a multi-modal task on an input dataset comprising an input image to generate a token output which is a text response to the input dataset. The VLM is trained using a training database comprising tuples of sample input datasets and corresponding sample token outputs. The sample input dataset of some of the tuples comprises an image generated from a text description by a text-to-image model, and the corresponding sample token output comprises at least part of the text description.