We investigate the use of wavelet-space feature decomposition in neural super-resolution for rendering pipelines. Building on recent neural upscaling frameworks, we introduce a formulation that predicts stationary wavelet coefficients rather than directly regressing RGB values. This frequency-aware decomposition separates low- and high-frequency components, enabling sharper texture recovery and reducing blur in challenging regions. Unlike conventional wavelet transforms, our use of the stationary wavelet transform (SWT) preserves spatial alignment across subbands, allowing the network to integrate G-buffer attributes and temporally warped history frames in a shift-invariant manner. The predicted coefficients are recombined through inverse wavelet synthesis, producing resolution-consistent reconstructions across arbitrary scale factors. We conduct extensive evaluations and ablations, showing that incorporating SWT yields superior perceptual quality compared to industry baselines, while maintaining real-time performance on modern hardware. Taken together, our results suggest that wavelet-domain neural super-resolution provides a principled and efficient path toward higher-quality real-time rendering, with broader implications for neural rendering and graphics applications

Hands-on solutions and prototypes from our department teams.

Nepali Barnamala consists 36 consonants, 12 vowels and 10 Nepali digits. Among them, this paper uses the 36 consonants and 10 Nepali digits for the recognition using machine learning based algorithm mainly: Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Random Forest (RF) and several architectures of neural networks. In this paper, different kernel tricks of SVM with different regularization parameters has been used to train model and has compared their accuracy and F1-score. In KNN, accuracy and F1-score are compared with different values of K and distance matric. In Neural Networks, training accuracy, training loss, validation accuracy, and validation loss are compared with different number of hidden layers regularization parameters and learning rate. Different hyperparameter of random forest are changed and compared to their corresponding result. This paper uses the Kaggle dataset of school students’ Nepali handwritten characters. The dataset is CSV format with 78,200 rows for forty-six different classes with 1024 (32*32 image size) columns plus one column for label of characters for training and 13,800 rows for testing. For handwritten Nepali Barnamala recognition, the best average accuracy is 93.51% of neural networks with four hidden layers.