Synthesizing realistic cardiac ultrasound videos is a valuable capability for medical simulation, data augmentation, and modeling physiological dynamics. However, adopting high-fidelity video diffusion models in this domain is challenging due to limited domain data, narrow distributional variation, and high computational cost. While pre-trained 2D image generation diffusion models offer strong image priors, they lack temporal coherence and cannot generate consistent video without architectural changes. In this work, we introduce EchoAdapter, a lightweight temporal adapter that enables pre-trained 2D general image diffusion models to generate echocardiography videos without changing and re-training the backbone. EchoAdapter adopts an autoregressive frame-by-frame generation strategy, conditioning each frame on previously generated frames and a cardiac temporal embedding via LoRA-augmented FiLM modules. To improve robustness, we incorporate an uncertain condition mechanism that adaptively blends reference features based on structural confidence. Trained on the EchoNet-Dynamic dataset, our method matches the performance of fully-trained video diffusion models and achieves the best results on parameter-efficiency-aware metrics.