A Keras image-quality judge that trains big models on small GPUs

What it does

NIMA assigns a mean and standard deviation score to images, effectively rating their aesthetic quality. This repo provides a Keras and TensorFlow implementation with pre-trained weights for MobileNet, NASNet Mobile, and Inception ResNet v2, all trained on the AVA dataset. You can use it to automatically inspect image quality or drop it in as a loss function to nudge generative models toward prettier outputs.

The interesting bit

The standout is a three-stage pre-training pipeline for models too large to fit on a single GPU: extract features to TFRecords, train a small classifier on top, and optionally fine-tune the full stack if memory allows. It is an admission that not everyone has a V100, and the README is upfront that this shortcut sacrifices performance unless you finetune afterward.

Key highlights

• Pre-trained weights provided for NASNet Mobile (0.067 EMD), Inception ResNet v2 (~0.07 EMD), and MobileNet (0.0804 EMD) on the AVA dataset.
• Scores images via mean and standard deviation, useful both for standalone evaluation and as a differentiable loss function.
• Includes a feature-extraction workaround to train oversized models when direct single-GPU training is impossible.
• Supports batch evaluation of directories or individual image paths.

Caveats

• NASNet models do not support runtime resizing; images must be pre-resized to 224×224 before scoring.
• The pre-training workaround for large models yields weaker performance unless you complete the optional fine-tuning step.
• Corrupted images in the training set drastically slow down TensorFlow Dataset buffering, so dataset cleaning is mandatory.

Verdict

Useful for developers building generative pipelines or automated curation tools who need a ready-made aesthetic metric in Keras. Skip it if you need a turnkey, paper-exact reproduction or if your workflow depends on resizing NASNet inputs on the fly.