基于推理时VLLM缩放与跨域整流流反转的训练无关风格化抽象方法及装置
By employing VLLM scaling and cross-domain rectified flow inversion during inference, the problems of insufficient identity preservation and limited stylization in stylization techniques are solved, achieving high-quality stylized abstraction and evaluation, applicable to various abstract styles and subjects, and reducing computational costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TAIYUAN UNIVERSITY OF TECHNOLOGY
- Filing Date
- 2025-11-24
- Publication Date
- 2026-07-17
AI Technical Summary
Existing stylization techniques suffer from problems such as insufficient identity preservation, limited stylization degree, and the need for additional training when dealing with abstract style transfer, and the evaluation methods are inaccurate.
We employ a training-independent stylization abstraction method based on inference-time VLLM scaling and cross-domain rectified flow inversion. Through multiple rounds of VLLM interaction, we extract detailed identity features, iteratively optimize identity representation, and use rectified flow technology to balance style and structure. Finally, we combine the StyleBench evaluation method for comprehensive quality measurement.
It achieves high-quality stylized abstraction, maintains the core identity characteristics of the subject, reduces computational costs, is applicable to various abstract styles and subjects, provides interpretability and controllability, and provides accurate evaluation results.
Smart Images
Figure CN121544452B_ABST