This paper offers yet another static analysis method aimed at classifying malware families, by disassembling the executables with Radare2 and traversing the static call graph to train CNNs on instruction-based RGB images. The instruction-based family detection should have the potential to model common behavioral patterns, thus creating a profile for various families and actors. The experiments are carried out on the BODMAS, MalImg, and IBD (internal Bitdefender dataset). Our method’s performance is compared to another static feature selection method – the EMBER features. Furthermore, we reveal proof of correlation between packers and malware families in all three datasets. Our conclusion states that the proposed model’s accuracy does not reach the EMBER feature’s performance due to the high number of packed files in these datasets. However, its stability still motivates its use since the instruction-based information cannot be altered easily as header-based features – our observations infer that while classifying malware families, ML methods which ignore unpacking the samples may overfit the data, learning packer traits instead of actual family behaviour, offering no explainability over the decision.