With the rapid advancements in facial editing technologies, deepfake detection for multi-step continuous facial manipulation has emerged as a crucial research area in computer vision and visual security. Current state-of-the-art detection methods typically rely on traditional convolutional neural network (CNN) feature extractors, which struggle to capture fine-grained local manipulation traces, thereby limiting their detection performance in complex, continuous forgery scenarios. To address this challenge, this paper introduces a coordinate-based multi-scale strategy, MSCP-SeqFakeFormer. By incorporating coordinate positioning between the CNN feature extraction and Transformer encoder, the approach leverages multiple scales and receptive fields to effectively enhance local detail features associated with facial forgeries. This enhancement significantly improves the model’s ability to recognize manipulation types and their sequences. Experimental results on the publicly available Seq-DeepFake dataset demonstrate that the proposed method outperforms the baseline model in both fixed-length (Fixed-Acc) and adaptive-length (Adaptive-Acc) evaluation metrics, and also achieves superior performance in the facial image recovery task.