Multimodal Diffusion Model Training for Detailed Text-to-Image Control

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing Text to Image (T2I) diffusion models struggle to capture complex details and nuances in image generation due to the inherent simplicity of text descriptions, often failing to generate images that fully represent the intended scenario.

Innovation Solution

A method involving the use of universal visual language, such as product images and sketches, to enrich descriptive information, combined with a Multimodal Diffusion Transformer (MMDiT) architecture, to train a target model through N diffusion networks, enhancing the model's ability to process multimodal information and improve controllability.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of information

If text descriptions are used as the sole input for image generation, then the model operation is simple, but the ability to capture complex details and nuances is insufficient

Engineering Contradiction:
Improvedescriptive informationVSAvoidmodel structure
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The patent combines multiple input modalities (text descriptions, reference images, sketches) into a unified multimodal input framework. The MMDiT architecture integrates text encoders and image encoders to process different modalities simultaneously, merging their features into a comprehensive representation that captures both semantic information from text and visual details from images, thereby reducing information loss in the generation process.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The MMDiT architecture is designed as a universal model that can handle multiple input types (pure text, text with reference images, text with sketches) through a single unified framework. The model's diffusion networks can process different modality combinations, making the system multi-functional while maintaining a consistent architectural structure, thus avoiding the need for separate models for each input type.

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

2Manufacturing precision

If multiple diffusion networks are used to process multimodal information, then the controllability of image generation is improved, but the training complexity increases

Engineering Contradiction:
Improveimage generation controlVSAvoidtraining process
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent segments the image generation process into multiple sequential diffusion networks (N diffusion networks in total). Each diffusion network processes the generation at a different stage, with early networks handling coarse structures and later networks refining details. This segmentation allows progressive control over the generation process, improving controllability while distributing the training complexity across multiple manageable stages rather than requiring one extremely complex model.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The model performs preliminary processing of input modalities through dedicated encoders (text encoder, image encoder) before feeding them into the diffusion networks. Reference images and sketches are pre-processed to extract features, and text is encoded to obtain semantic representations. This preliminary action prepares the data in a standardized format, reducing the complexity of the main diffusion process and enabling better control over the generation stages.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If reference images and sketches are integrated into the input, then the detail representation is enhanced, but the processing difficulty increases

Engineering Contradiction:
Improvedetail captureVSAvoidmultimodal processing
Core Design Contradiction:
Measurement precisionVSDifficulty of detecting and measuring

Solution Approach 1:

The patent introduces encoder networks as intermediary components that bridge different modalities. The text encoder transforms text into semantic features, the image encoder processes reference images into visual features, and sketch encoders convert sketches into structural features. These encoders act as mediators that convert diverse input formats into a unified feature space, making it easier for the diffusion networks to process multimodal information without directly handling the complexity of raw multimodal data.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The model transforms different input modalities into standardized feature representations through encoders, changing the parameter space of each modality to a common format. Text embeddings, image features, and sketch representations are all converted to compatible feature vectors with consistent dimensions and distributions. This parameter transformation enables the diffusion networks to process diverse inputs uniformly, reducing processing difficulty while preserving detail information.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4708140A1Target model training method, multimodal data processing method, and apparatuses and devices therefor
Publication Date: 2026.03.11 BEIJING BAIDU NETCOM SCI & TECH CO LTD
  • EP4708140A1 patent drawingFigure 1~2
  • EP4708140A1 patent drawingFigure 3(a)
  • EP4708140A1 patent drawingFigure 3(b)

AI summary

The present disclosure provides a target model training method, a multimodal data processing method, and apparatuses and devices therefor, relating to the field of artificial intelligence technology, and in particular to the fields of computer vision, deep learning, large model and other technologies. The target model training method includes: inputting (S101) sample data into a preset model to obtain initial multimodal features of the sample data; and using (S102) the initial multimodal features and a preset noise feature to perform model training on N diffusion networks in the preset model to obtain a target model when parameters of an image-text encoder of the preset model are fixed.