Control of experimental impact in cognitive and neuroscience studies is complicated by the mixed cognitive and affective effects of stimuli from different emotional image databases. One effective solution to this problem is to use semantically neutral images. We created a new database of colored abstract images, the AFIS, with annotations for emotional valence, complexity, average brightness, pixel brightness variance, and contrast. To validate these annotations, we employed several methods: a pre-trained random forest model to estimate emotional valence; ResNet50 and VGG19 convolutional neural networks to estimate the probability of semantic classification; and calculations of two-dimensional and three-dimensional fractal dimensions, as well as Lempel-Ziv complexity, to assess overall image complexity. Our results indicate that different valence classes are distinguished by their color palette composition. Furthermore, the brightness and contrast characteristics of the images contribute to the valence of the AFIS stimulus set. The fractal images included in the database are sufficiently abstract and semantically independent. The next stage of this work involves collecting expert ratings of the stimuli's emotional characteristics (valence and activation parameters) and assessments of their complexity. The Affective Fractal Image System addresses a gap in stimulus databases by providing affective norms for color palettes alongside varied complexity parameters. The AFIS database can be used to monitor subjects' levels of activation or emotional states and is applicable in studies of perception, memory, and thinking.