Stroke prior sketch generation method based on diffusion model score distillation

By combining a multimodal vector graphic generation method with score distillation and image segmentation, the problems of insufficient structural controllability and semantic understanding of diffusion models in vector graphic generation are solved, achieving efficient and controllable vector sketch generation, which is suitable for multiple application scenarios.

CN122368258APending Publication Date: 2026-07-10ZHEJIANG UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
ZHEJIANG UNIV OF SCI & TECH
Filing Date
2026-04-27
Publication Date
2026-07-10

AI Technical Summary

Technical Problem

Existing diffusion models suffer from poor structural controllability, insufficient semantic understanding, and difficulty in directly adapting generated results to vector formats in vector vector generation tasks. Furthermore, the generation process is time-consuming and difficult to iterate quickly.

Method used

A vector sketch generation method based on a diffusion model is adopted, which combines score distillation and image segmentation. Through a semantic encoder, a diffusion model, a stroke prior analysis module, and a multi-optimization generation module, a vector sketch generation method with controllable stroke count is achieved.

Benefits of technology

It achieves highly controllable and rapid vector image generation, improves the semantic alignment capability and visual quality of the generated results, and is suitable for scenarios such as graphic design, animation production, intelligent drawing and human-computer interaction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122368258A_ABST
    Figure CN122368258A_ABST
Patent Text Reader

Abstract

The application discloses a stroke prior sketch generation method based on diffusion model score distillation. Text prompt information input by a user is obtained, and the text prompt information is input into a semantic encoder for semantic encoding processing to obtain text semantic features; the text semantic features are input into a diffusion model to generate a reference RGB image; the reference RGB image is input into a stroke prior analysis module for image segmentation, edge detection and complexity evaluation processing in sequence to obtain stroke number prior information; and the stroke number prior information, the reference RGB image and the text semantic features are input into a multiple optimization generation module to generate a stroke number controlled vector sketch result. The application automatically obtains stroke number estimation, and compared with a method of manually setting stroke number or using a fixed rendering strategy, can adaptively adjust according to image region complexity, so that the generated vector sketch is more in line with structure density and visual level requirements.
Need to check novelty before this filing date? Find Prior Art