Content processing method and apparatus, electronic device, and storage medium
By identifying the structural features of the weight matrix in the content processing model and adaptively allocating spectral coefficients, the problem of poor model processing performance caused by the limited number of trainable spectral coefficients is solved, and the accuracy and performance of the model are improved with limited spectral coefficients.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TENCENT TECHNOLOGY (SHENZHEN) CO LTD
- Filing Date
- 2026-05-18
- Publication Date
- 2026-07-17
AI Technical Summary
During the training of content processing models, existing techniques construct a spectrum matrix associated with the weight matrix and update some spectrum coefficients. However, when the number of trainable spectrum coefficients is limited, the trained model performs poorly and has low accuracy.
By acquiring the training content and the content processing model, the structural features of each weight matrix are determined, the number of coefficients is adaptively allocated, key weight matrices are identified and more spectral coefficients are allocated to them, and non-key weight matrices are prevented from occupying too many spectral coefficients. The updated spectral matrix is then used to update the weight matrix to obtain the trained content processing model.
Despite the limited number of trainable spectral coefficients, this study improves the processing performance and accuracy of the content processing model, ensures that the key weight matrix receives more spectral coefficients, and enhances the model's expressive power and result accuracy.
Smart Images
Figure CN122412958A_ABST