In　the　most　of　exiting　Local　linear　embedding　(LLE)-based　image　super-resolution　methods,　a　Low　resolution　(LR)　image　can　be　represented　as　a　linear　combination　of　LR　training　samples.　In　these　methods,　the　combination　coefficients　of　the　LR　image　are　directly　used　to　estimate　the　High　resolution　(HR)　image.　However,　experimental　results　show　that　the　LR-LLE　coefficients　are　different　from　the　corresponding　HR-LLE　coefficients.　To　bridge　the　gap　between　LR　and　HR　images,　a　novel　LLE-based　face　hallucination　algorithm　is　proposed.　An　LLE　coefficients　prior　model　is　introduced　to　reduce　the　coefficient　errors.　In　this　prior　model,　the　LLE　coefficients　of　the　interpolated　LR　face　image　are　used　to　constraint　the　reconstructed　coefficients.　Experimental　results　show　that　the　proposed　method　can　provide　improved　performance　over　the　compared　methods.