What it does
This KPI predicts the emotional tone of an image (from unpleasant to pleasant). It is based on a deep learning model trained to predict the degree to which an image has a positive, neutral, or negative sentiment.
Why it matters
The emotional tone of imagery greatly influences consumer reactions. Positive imagery can trigger mirror neuron responses, leading to heightened engagement and positive brand associations, while negative imagery can have the opposite effect.
How it works
The model evaluates user-provided images, assigning a sentiment score from 0 (negative) to 100 (very positive). This score categorizes the image's sentiment, with the context of the asset type used to refine the rating into low, medium, or high sentiment.
How to achieve great results
The illustration shows sample images with model predictions. Top left image has lowest, bottom right image has highest image sentiment.
Images with strong positive sentiment often include joyful or affectionate moments. Conversely, images with strong negative sentiment tend to depict distressing or unpleasant situations.
Examples with assets:
AI models used
Sentiment models at aimpower are based on the latest deep learning frameworks used in scientific literature for the purpose of emotion modelling, including convolutional deep learning networks and vision transformers. The images used to collect human ground truth data originate from various sources, including social media platforms, large scale webscrapes (Common Crawl) and marketing materials (e.g. ads).
In total we leverage 500 Mio+ images/text pairs for pretraining and 15K+ images for finetuning our models with human ground truth data. Human ground truth is collected via online / web-based tasks. Evaluation of our models follows the scientific best practices. Our model predicts sentiment ratings with r = 0.92 [P < 0.0001, permutation test; root mean square error (RMSE), 0.4809]
Science background
Many scientific theories describe the emotional or affective states of humans (see Moors, 2022). Among them is the emotional model proposed by Russell (1980), the emotional diagram model developed by Plutchik (2001), the Pleasure, Arousal, Dominance (PAD) emotional state model proposed by Mehrabian (1996) and finally, the three-dimensional cube model developed by Lövheim (2012). At aimpower, we use Russell’s circumplex model of affect as a point of reference, from which emotions in the valence (“sentiment”) and arousal (“activation”) axes are distributed. This valence-arousal model has been widely used in the field of psychology. Valence (“sentiment”) represents the pleasantness of an emotional stimulus, ranging from unpleasant to pleasant. Arousal (“activation”) is the intensity of emotion provoked by a stimulus, ranging from low to high.