A method and system for constructing an ancient costume generation model based on a double-flow high-frequency perception model

By constructing a DSHF-Diff model and utilizing frequency domain enhancement and dual-stream attention mechanisms, the problems of high-frequency detail distortion and form errors in the generation of ancient costumes were solved, achieving high-fidelity digital generation of ancient costumes that conforms to historical forms.

CN122263971APending Publication Date: 2026-06-23NORTHWEST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
NORTHWEST UNIV
Filing Date
2026-01-30
Publication Date
2026-06-23

AI Technical Summary

Technical Problem

Existing text-to-image generation technologies struggle to maintain the authenticity of high-frequency details and adherence to historical norms when generating images of ancient Chinese clothing, resulting in texture distortion and format errors.

Method used

A dual-stream high-frequency perception-based ancient clothing generation model (DSHF-Diff) is constructed. High-frequency texture features are extracted through a frequency-domain enhanced reference encoder. Combined with pose structure flow and semantic style flow, a low-rank adaptive weight and dual-stream spatial reference attention mechanism are used to achieve consistent generation of texture, pose and semantics.

Benefits of technology

It significantly improves the high fidelity of ancient costumes, solves the problems of high-frequency detail distortion and shape errors, and realizes high-fidelity digital restoration of ancient costumes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122263971A_ABST
    Figure CN122263971A_ABST
Patent Text Reader

Abstract

This invention belongs to, but is not limited to, the field of computer vision technology, and discloses a method and system for constructing an ancient clothing generation model based on a dual-stream high-frequency perception model. The model constructs a collaborative dual-path architecture, introducing a frequency-domain enhanced reference encoder and explicitly strengthening the high-frequency components in image features using Fast Fourier Transform, significantly enhancing the model's ability to perceive fine-grained textures such as intricate embroidery and fabric warp and weft. Simultaneously, this invention designs a dual-stream spatial reference attention mechanism and combines it with a noise-perception adapter synchronization strategy to achieve precise pixel-level mapping between static high-frequency features in the reference image and the target pose space. This method outperforms existing controllable image generation frameworks in metrics such as FID, LPIPS, and CLIP-I, providing an efficient technical path for the rigorous restoration of cultural heritage.
Need to check novelty before this filing date? Find Prior Art