Multi-view　subspace　clustering　is　extensively　investigated　for　its　ability　to　extract　essential　information　from　multiple　data.　However,　tensor-based　methods　often　encounter　several　limitations:　1)　They　suffer　from　high　computational　complexity　due　to　the　construction　of　a　global　affinity　matrix;　2)　The　sophisticated　semantic　information　among　samples　remains　under-explored.　To　address　these　issues,　we　propose　a　comprehensive　framework　called　tensor-derived　large-scale　multi-view　subspace　clustering　with　faithful　semantics,　which　replaces　the　original　graph　with　a　trustworthy　anchor　graph.　In　particular,　a　graph-optimization-based　anchor　selection　strategy　is　designed　to　obtain　salient　points,　and　thus　the　anchor　graph　is　computed　to　decrease　the　computational　complexity　of　constructing　the　representation　matrix.　Subsequently,　a　refinement　approach　is　designed　to　flexibly　extract　essential　semantics　between　nodes　by　dividing　the　graph　into　significant　components　and　undesired　connections.　These　matrices　preserving　important　information　are　fused　into　a　tensor　that　is　constrained　by　a　nuclear　norm　to　retain　its　low-rank　property.　Meanwhile,　the　undesired　links　should　be　eliminated　to　avoid　confusing　the　clustering　results.　Finally,　the　spectral　embedding　is　employed　to　directly　guide　the　learning　of　anchors　and　graphs.　The　proposed　model　achieves　a　remarkable　improvement　of　3.3%　and　13.1%　of　ACC　on　the　NoisyMNIST　and　Prokaryotic　datasets　while　reducing　high　computational　complexity　compared　to　other　subspace-based　clustering　approaches.