<p>Tensor low-rank representation (TLRR) has attracted extensive attention in the fields of image processing. However, most existing TLRR models adopt the tensor nuclear norm to impose low-rank constraints, which often leads to suboptimal approximations. Moreover, the literature still lacks effective approaches that leverage prior information from subspaces and noise to enhance low-rank representation. To address this problem, we propose a general two-stage tensor low-rank representation model (TSTLRR). In the first stage, an initial denoising model constructs a dictionary and explores the prior information of the subspace and noise. In the second stage, adaptive weighted tensor Schatten p-norm acts as a low-rank constraint to distinguish the contribution of singular values to data. The noise is split for regions with different sparsity to achieve more targeted denoising. Most existing methods are shown to be special cases of the method proposed in this paper. The optimization method for TSTLRR based on the ADMM algorithm is proposed. Compared with the state-of-the-art models, the experiments on synthetic datasets and real datasets demonstrate the effectiveness of TSTLRR.</p>