Face　recognition　under　varying　illumination　conditions　is　an　unsolved　problem.　Illumination　normalization　method　is　a　common　preprocessing　method　for　this　problem.　One　of　the　most　popular　illumination　normalization　methods　is　Total　Variation　based　Quotient　Image　(TVQI)　model.　However,　only　the　illumination　invariant　information　in　the　small-scale　part　of　image　is　used　in　TVQI　model;　therefore,　the　information　is　very　limited.　In　this　paper,　a　Multi-scale　Fusion　TV-based　Illumination　Normalized　(MFTVIN)　model　is　proposed　to　resolve　the　problem.　Firstly,　it　uses　Multi-scale　Splice　TVQI　model　(MSTVQI)　to　generate　the　small-scale　part.　This　model　is　based　on　TVQI　model,　and　it　can　generate　the　illumination　invariant　small-scale　part　which　contains　more　detailed　information　than　TVQI　model.　Secondly,　TV-L2　model　is　used　to　get　the　noiseless　large-scale　part　of　human　face.　The　large-scale　part　contains　the　contour　of　human　face　and　the　shade　information.　Illumination　effects　in　the　large-scale　part　are　removed　by　region-based　histogram　equalization　and　homomorphic　filtering.　Lastly,　two　parts　are　fused　to　generate　the　illumination　invariant　face　sample.　MFTVIN　model　do　not　need　the　information　about　the　lighting　source　and　training　set.　Experimental　results　on　some　famous　face　databases　prove　that　the　processed　image　by　our　model　could　largely　improve　the　recognition　performances　under　low-level　lighting　conditions.