InstructAvatar is an innovative text-guided method for generating 2D avatars with rich emotional expression. This model controls avatar emotions and facial expressions via a natural language interface, offering fine-grained control, improved interactivity, and generalization ability for generated videos. It utilizes an automated annotation process to construct a training dataset of instruction-video pairs and incorporates a novel dual-branch diffusion base generator capable of predicting avatars simultaneously based on audio and textual instructions. Experimental results demonstrate that InstructAvatar outperforms existing methods in fine-grained emotional control, lip-sync quality, and naturalness.