Emotional images are often remembered better than neutral images, but this advantage may depend on how they are encoded. This study examined whether emotional valence and encoding condition shape image-level memorability, whether ResMem-predicted memorability corresponds to human derived memorability scores, and whether predicted visual attention  (DeepGaze IIE) overlaps with memorability-relevant image regions (AMNet). Participants viewed negative and neutral images during either passive viewing or active notetaking and completed a delayed old/new recognition task after 14 days. Human memorability was compared with ResMem predictions, and spatial overlap between DeepGaze IIE saliency maps and AMNet memorability maps was assessed. Negative images were remembered better than neutral images during passive viewing, but this advantage disappeared during active encoding. Active note taking increased memorability for both image types, especially neutral images. ResMem predictions correlated with human memorability only for negative images in the passive condition. Attention–memorability overlap did not differ between negative and neutral images.  These findings suggest that image memorability is shaped by both emotional content and encoding context, and that the correspondence between model-predicted and human memorability may be context- and valence-dependent.