Abstract: In order to synthesize large-scale plants in a virtual forest scene, this paper presents a real-time synthesis method of large-scale plants by tilling the pre-generated samples based on Poisson disk tiles. In the sample generation phase, we adopt a corner matching manner and use the relaxation dart throwing algorithm to generate the plants distribution in samples, with which we overcome the corner problem and disk crossing-border problem existed in the traditional border matching method. In the synthesis phase, a corner-based direct stochastic tilling method is used to rapidly synthesize the distribution info in the current camera view, which can achieve a real-time performance when synthesizing large-scale plant distribution. Furthermore, this paper proposes a strategy for synthesizing multi-species, non-uniform distributions of plants. With this strategy, multi-level sample sets are generated by extracting the sample subsets from the original high-density sample set. In the synthesis process, the right levels of samples will be selected for tilling according to the current species and density info on the user-designed map, and then we synthesize the multi-species, multi-density distributions of plants. Based on the method proposed in this paper, we have developed a forest scene modeling and roaming system. Experimental results show that the method is very efficient in the construction of large-scale forest scenes. Even when the number of plants reaches ten million, the synthesis can still be performed at an interactive rate.