The invention relates to a large-scale virtual crowd real-time rendering method, which comprises the following steps of: 1, introducing the conventional grid model and extracting geometric information and animation information of the model; 2, performing octree space subdivision on the model, wherein approximate description of a part of the model related to geometric size of each node is stored in the node; 3, performing point sampling on the surface of the part of the model included by each node; 4, processing and modeling a sample point, wherein the step comprises the sub-steps of calculating sample point information by interpolation, selecting sample point animation information, oversampling and removing redundancy and the like; 5, establishing model sampling data of three-layer LOD (Levels of Detail) according to specified parameters; 6, performing view frustum culling accelerated by a GPU (Graphic Processing Unit) on a virtual crowd in a large-scale scene during the real-time rendering; 7, performing an LOD strategy accelerated by the GPU on the cuffing result, wherein the step comprises the sub-steps of selecting role LOD and ordering LOD; and 8, sequentially performing GPU skin-skeleton animation based instancing rendering on the role of each layer of LOD level. By adopting the method, quick real-time rendering of a large-scale virtual crowd can be realized.