Visual saliency models seek image regions that attract human attention and are central to applications where human perception matters. We present UniRare ( https://github.com/numediart/UniRare ), a universal plug-in framework that reconciles top-down (learned, task-driven) and bottom-up (stimulus-driven) attention within deep saliency models that use convolutional encoders. UniRare derives a rarity-based bottom-up map directly from encoder features of any existing saliency model, highlighting statistically uncommon structures in the input. This map can be inspected on its own or fused with the existing model’s original, predominantly top-down saliency to expose complementary cues. UniRare requires no additional training, minimal integration effort, and operates as a drop-in component. We demonstrate seamless integration with three diverse saliency models (TranSalNet, UniSal, and TempSal) and show that the rarity maps complement their outputs by emphasizing low-level, stimulus-driven evidence that top-down predictors may overlook. UniRare thus offers a simple, training-free path to jointly analyze and combine bottom-up and top-down mechanisms in visual saliency.