一种图像生成模型训练方法、图像生成方法、装置及设备
By fine-tuning the diffusion model and optimizing its dissimilarity, the problem of insufficient image generation quality for rare categories was solved, and high-quality image generation was achieved.
CN118552633BActive Publication Date: 2026-07-17LANGCHAO ELECTRONIC INFORMATION IND CO LTD
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- LANGCHAO ELECTRONIC INFORMATION IND CO LTD
- Filing Date
- 2024-04-12
- Publication Date
- 2026-07-17
AI Technical Summary
Technical Problem
Existing diffusion models struggle to generate high-quality images when training on rare classes with limited data.
Method used
By fine-tuning the pre-trained diffusion model using the training set, adjusting the category information and number of images in the training data, and optimizing the model using dissimilarity calculation, high-quality rare category images are generated.
Benefits of technology
The training performance of the diffusion model on rare categories has been improved, ensuring the generation of high-quality rare category images.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN118552633B_ABST
Abstract
本发明提供一种图像生成模型训练方法、图像生成方法、装置及设备，涉及机器学习领域，方法包括：利用训练集对预训练扩散模型进行微调训练，得到第一扩散模型；其中，训练集包含多种类别的训练数据，训练数据包含类别信息和训练图像，扩散模型用于根据输入信息生成图像；将各类别的类别信息输入第一扩散模型，得到第一扩散模型输出的各类别的生成式图像；确定各生成式图像与同一类别的训练图像间的差异度，得到各类别对应的差异度；根据各类别的差异度对训练集中各类别的训练图像数量进行调整，并利用调整后的训练集对第一扩散模型进行微调训练，得到完成训练的第二扩散模型；可提升训练数据稀少的稀有类别在扩散模型中的训练效果。
Need to check novelty before this filing date? Find Prior Art