The Versatile Video Coding (VVC) standard significantly improves compression efficiency over its predecessors but introduces substantially increased computational complexity, mainly due to its flexible Quad-Tree plus Multi-Type Tree (QTMT) partition structure. To reduce the partition-search complexity in VVC intra coding, this paper proposes a targeted fast partition decision framework for 32 \(\times \) 32 Coding Units (CUs). The proposed framework uses a Transformer-CNN predictor to estimate the probability distribution of six 32 \(\times \) 32 CU partition modes, and then applies a confidence-aware candidate-retention strategy to reduce unnecessary rate-distortion optimization (RDO) evaluations while retaining selected candidate modes for the final encoder-side decision. Experimental results on VTM 7.0 under the All-Intra configuration show that the proposed method reduces the total encoding time by 44.5%–56.1% with a 1.32%–2.95% Bjøntegaard delta bitrate (BDBR) increase on the tested sequences. These results indicate that the overall framework can provide a practical complexity–RD trade-off for the targeted 32 \(\times \) 32 CU level. However, because CNN-only/Transformer-only ablation, exact model complexity, and encoder-side overhead decomposition are not available in the current study, the independent contribution of the Transformer component and the detailed source of the runtime gain should be interpreted cautiously rather than as isolated evidence of architectural superiority.
