The precise detection of grape clusters is a prerequisite for achieving yield estimation，picking and other operations， but existing methods are still difficult to achieve lightweight and accurate detection of multi.variety grape clusters. To enhance the accuracy，robustness，and generalization of multi.variety grape cluster detection in complex natural scenes，a model named ESIC-YOLOv8n is proposed based on the improved YOLOv8n model. In this model，EMA and SA attention modules are respectively added to the Backbone and Neck networks of YOLOv8n to strengthen the network′s feature extraction and multi.scale feature fusion capabilities，meanwhile，to reduce the interference from occlusion or overlap in grape cluster detection and to improve the detection accuracy and recall. In addition，by replacing CIoU with Inner CIoU in the head and using auxiliary boxes to improve the accuracy of overlapping object detection，the overall detection accuracy and generalization of the model was enhanced. As a result，the ESIC-YOLOv8n model achieves a detection accuracy of 87. 00%，a recall rate of 81. 60%，mAP of 88. 90%，and F1 score of 84. 21%，representing improvements of 1. 05%，2. 90%，1. 48% and 2. 00%，respectively，compared to the original YOLOv8n model. The results indicate that the ESIC-YOLOv8n model possesses high accuracy，good generalization，and lightweight characteristics，providing technical support for research on grape yield estimation and harvesting.