To tackle real-time denoising of noisy laser speckle blood flow images, a novel lightweight denoising speckle contrast image generative adversarial network (LDSCI-GAN) is proposed. In the framework, a lightweight denoiser removes noise from the original image, and a discriminator compares the denoised result with the reference one, enabling efficient learning and optimization of the denoising process. With a multi-scale loss function in the log-transformed domain, the training process significantly improves accuracy and denoising by using only five frames of raw speckle images while well-preserving the overall pixel distribution and vascular contours. Animal and phantom experimental results indicate that the LDSCI-GAN can eliminate vascular artifacts while retaining the accuracy of relative blood flow velocity. In terms of peak signal-to-noise ratio (PSNR), mean structural similarity index (MSSIM), and Pearson correlation coefficient (R), the LDSCI-GAN outperforms other deep-learning methods by 3.07 dB, 0.10 (