A personalized LoRA model construction method and system, and a storage medium

By preprocessing and encoding a set of images of personal style works using a variational autoencoder, and combining text embedding vector alignment and LoRA adapter training, a collaborative generation architecture is constructed. This solves the problem of inaccurate style extraction in existing personalized LoRA models, and achieves efficient and lightweight generation of personalized artistic style images.

CN121458529BActive Publication Date: 2026-06-02XIANGJIANG LAB
View PDF 3 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
XIANGJIANG LAB
Filing Date
2026-01-06
Publication Date
2026-06-02

AI Technical Summary

Technical Problem

Existing technologies struggle to accurately capture and enhance individual artistic styles in the construction of personalized LoRA models, resulting in low personal style fidelity and blurred personalized features in generated images, failing to meet users' needs for accurate and stable personalized artistic style generation.

Method used

By acquiring a set of images of works with a personal style, preprocessing and encoding with a variational autoencoder, and outputting latent space features; combining text embedding vector alignment, freezing the backbone network of the image generation model, iteratively training the LoRA adapter, constructing a collaborative generation architecture, and outputting a LoRA model adapted to the personalized artistic style.

Benefits of technology

It significantly improves the purity of personalized artistic style representation of latent space features, accurately preserves the core stylistic elements such as unique brushstrokes, colors, and compositions, and achieves lightweight model deployment and efficient generation, ensuring accurate matching between generated images and raw image requirements.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121458529B_ABST
    Figure CN121458529B_ABST
Patent Text Reader

Abstract

The application relates to the technical field of image generation, and discloses a personalized LoRA model construction method and system and a storage medium. The method comprises the following steps: outputting a hidden space feature representing a personalized artistic style based on a variational autoencoder; outputting LoRA adjustment parameters adapted to the personalized artistic style based on the hidden space feature and a LoRA adapter; constructing a collaborative generation architecture based on the LoRA adjustment parameters, the variational autoencoder and an image generation model; encapsulating a personalized LoRA model to output an image meeting the image generation requirement and carrying the personalized artistic style. The application improves the style representation purity of the hidden space feature, retains core style elements, realizes model lightweight, reduces the parameter quantity and guarantees the adaptation accuracy, balances the image generation requirement and the style, promotes the high fusion of content and style, filters non-leading redundant features, improves the adaptation stability and supports full-link optimization through customized adaptation.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Image generation method based on single-image geometric texture joint control and terminal

    CN120259479A

  • Diffusion model optimization method and system based on U-Net parameter enhancement

    CN120580557A

  • AI diffusion model pattern generation platform and generation method

    CN121120818A