DDPM-Inspired Single Image Super Resolution
This team course project explored single-image super resolution using DIV2K. The models mapped a low-resolution image directly to a larger image. Diffusion models were an architectural reference; the implementation had no noise-conditioned reverse-diffusion sampling chain.
The retained report includes notebook code and training output. Reviewing that code exposed evaluation errors that invalidate the earlier interpretation of its performance table.
Models explored
The simple model used convolutional layers, transposed-convolution upsampling, and a tanh output. A second model used six residual blocks with channel attention, pixel-shuffle upsampling, and a global connection from the bilinearly upsampled input. Later configurations increased the number of residual blocks and considered VGG-19 feature loss alongside a pixel loss.
These configurations changed several architectural and training choices together. The record does not isolate the contribution of residual learning, attention, or perceptual loss.
Errors reproduced in the evaluation
The function named train_and_evaluate_model passed val_dataset to model.fit and then scored the same dataset. The calling cell loaded DIV2K's validation split. The resulting numbers therefore do not establish held-out performance.
The simple model also had incompatible preprocessing in the final comparison. It was trained on inputs and targets in , then given raw pixel inputs during evaluation. Its signed output was clipped to and cast directly to an 8-bit image, without inverse normalization.
The range error is reproducible without model training: a perfect signed-range prediction of a synthetic RGB gradient receives 4.77054 dB under the old cast. Applying the inverse transform reconstructs the reference to numerical precision. The reported 4.80 dB consequently cannot support the earlier claim that direct image prediction inherently failed.
Corrected evaluation
The replacement code requires separate training and evaluation image IDs, applies explicit input and output conventions, and computes metrics in RGB float space with a fixed range. It scores the model and bicubic baseline on every supplied evaluation image and retains per-image results. The checks verify range handling and reject overlapping splits before training.
The original checkpoints are not in the available source bundle, so these repairs do not produce a replacement held-out score. The previously quoted PSNR/SSIM table is withdrawn as evidence of generalization. The report now begins with the correction and retains the original notebook afterward.