一种图像快速生成方法及装置
By explicitly aligning the noise variance and performing amplitude recalibration during the diffusion model inversion process, the problem of noise underestimation is solved, enabling high-fidelity image reconstruction and editing, improving computational efficiency, and making it suitable for image enhancement and editing.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- FUDAN UNIVERSITY
- Filing Date
- 2026-05-09
- Publication Date
- 2026-07-17
AI Technical Summary
Existing diffusion model inversion methods suffer from a systematic underestimation of noise amplitude at high-noise time steps, leading to overly smoothed inversion trajectories, reduced reconstruction quality, and semantic drift. Furthermore, iterative inversion methods are computationally expensive, making them unsuitable for efficient applications.
By explicitly aligning the variances of the predicted noise and the forward reference noise during the inversion process, noise amplitude is recalibrated. The noise variance scaling factor is used to correct the noise underestimation problem, and an attention-guided spatial masking mechanism is introduced to reduce the number of U-Net calls.
Achieving high-fidelity reconstruction with fewer steps significantly reduces the number of U-Net calls, minimizes semantic drift, and improves computational efficiency, making it suitable for applications such as image enhancement and editing.
Smart Images
Figure CN122176109B_ABST