Neural-IMage-Assessment

Why we included this project

Teams that need to rank images by quality, not just detect what is in them, will find this a useful starting point. It is a PyTorch port of the NIMA paper, and instead of returning a single aesthetic score it predicts a distribution, so you can order images by expected quality and also see how confident the model is. That uncertainty signal is what makes it practical for deciding what to surface in a gallery or filter out of a feed. The repo includes training and inference scripts, an ImageNet-pretrained VGG-16 base, and pre-split AVA annotation files, so reproducing the pipeline or retraining on your own data takes little glue code. It is a research codebase rather than a maintained product, and the author is upfront that the learning rate needed adjusting and hyperparameters were not heavily tuned, which saves you the same trial and error.