This paper presents a deep multi-scale video prediction method that improves upon traditional Mean Squared Error (MSE) loss by introducing three complementary strategies: a multi-scale architecture, adversarial training, and an image gradient difference loss. The goal is to generate sharp, accurate future video frames without requiring explicit tracking of pixel trajectories. The method is evaluated on the UCF101 and Sports1m datasets, and compared to previous approaches such as recurrent neural networks and LSTM-based models. The proposed approach uses a convolutional network to predict future frames from a sequence of input frames. To address the inherent blurriness of MSE-based predictions, the authors introduce a multi-scale architecture that allows the model to capture both short- and long-range dependencies. They also employ adversarial training, where a discriminative network is trained to distinguish between real and generated frames, forcing the generator to produce more realistic predictions. Additionally, a new loss function based on image gradient differences is introduced to preserve the sharpness of the predicted frames. The results show that the combination of these strategies significantly improves the quality of video predictions compared to traditional MSE-based methods. The multi-scale model, adversarial training, and gradient difference loss all contribute to sharper and more accurate predictions. The best results are achieved by combining the multi-scale architecture, $ \ell_1 $ norm, gradient difference loss, and adversarial training. The paper also compares the proposed method to previous approaches, including those based on recurrent neural networks and LSTM models. The results demonstrate that the proposed method outperforms these approaches in terms of PSNR, SSIM, and sharpness measures. The method is particularly effective in capturing motion in dynamic scenes, as it focuses on the moving areas of the images rather than the static background. The authors conclude that their approach provides a more accurate and sharper prediction of future video frames compared to traditional methods. The model is fully differentiable, allowing it to be fine-tuned for other tasks if needed. Future work will focus on evaluating the classification performance of the learned representations in a weakly supervised context, such as on the UCF101 dataset. The method could also be extended to combine with optical flow predictions or used in applications where next frame prediction is required without explicit optical flow.This paper presents a deep multi-scale video prediction method that improves upon traditional Mean Squared Error (MSE) loss by introducing three complementary strategies: a multi-scale architecture, adversarial training, and an image gradient difference loss. The goal is to generate sharp, accurate future video frames without requiring explicit tracking of pixel trajectories. The method is evaluated on the UCF101 and Sports1m datasets, and compared to previous approaches such as recurrent neural networks and LSTM-based models. The proposed approach uses a convolutional network to predict future frames from a sequence of input frames. To address the inherent blurriness of MSE-based predictions, the authors introduce a multi-scale architecture that allows the model to capture both short- and long-range dependencies. They also employ adversarial training, where a discriminative network is trained to distinguish between real and generated frames, forcing the generator to produce more realistic predictions. Additionally, a new loss function based on image gradient differences is introduced to preserve the sharpness of the predicted frames. The results show that the combination of these strategies significantly improves the quality of video predictions compared to traditional MSE-based methods. The multi-scale model, adversarial training, and gradient difference loss all contribute to sharper and more accurate predictions. The best results are achieved by combining the multi-scale architecture, $ \ell_1 $ norm, gradient difference loss, and adversarial training. The paper also compares the proposed method to previous approaches, including those based on recurrent neural networks and LSTM models. The results demonstrate that the proposed method outperforms these approaches in terms of PSNR, SSIM, and sharpness measures. The method is particularly effective in capturing motion in dynamic scenes, as it focuses on the moving areas of the images rather than the static background. The authors conclude that their approach provides a more accurate and sharper prediction of future video frames compared to traditional methods. The model is fully differentiable, allowing it to be fine-tuned for other tasks if needed. Future work will focus on evaluating the classification performance of the learned representations in a weakly supervised context, such as on the UCF101 dataset. The method could also be extended to combine with optical flow predictions or used in applications where next frame prediction is required without explicit optical flow.