Generating high-quality, detailed 3D avatars from monocular video presents a significant challenge in digital human community. Recent methods have shown the potential to recover 3D geometry from monocular video, but they typically suffer from pose variability, geometric inaccuracies and insufficient texture realism. To address these challenges, we propose ReDACT, a novel framework designed for the reconstruction of high-fidelity 3D avatars from monocular video. To improve geometric accuracy, ReDACT first proposes a cycle-deforming field ensuring consistent geometry across frames during NeRF-based 3D reconstruction for the generation of high-fidelity avatars. To further enhance texture quality, we introduce a cascaded diffusion structure that emphasizes fine-grained details, such as skin and fabric patterns, allowing for controllable texture generation. Finally, the reconstructed avatars are animated via pose-driven deformation and rendered photo-realistically using volume rendering. Qualitative and quantitative results demonstrate that our method achieves high-quality reconstruction shapes, good texture realism, and robust generalization compared to prior methods.