This study develops and evaluates deep learning methods for automated creativity assessment in nonverbal tests, focusing on Form B of K. Urban’s Test of Creative Thinking — Drawing Production (TCT-DP). Seven pretrained architectures — MobileNet V2, AlexNet, ResNet-18, ViT, EfficientNet V2, ResNeXt-101, and DenseNet-121 — were fine-tuned to predict total creativity scores based on drawings from 1,138 participants aged 6–20 years. The dataset was split into training (70%), validation (15%), and test (15%) sets, with images resized to 1240 × 1600 pixels and subjected to random augmentations (flips, rotations, contrast, and saturation adjustments), followed by normalization using ImageNet statistics. ResNet-18 achieved the highest prediction accuracy (R² = .69) when excluding the Speed criterion, comparable to results for Form A, despite a smaller dataset and a narrower score range. The mean absolute error was 0.52 points (SD = 4.94), with prediction errors not exceeding 5 points in 70% of cases. To interpret model decisions, the explainable AI method Grad-CAM was applied, generating heatmaps based on gradients from the final convolutional layer. Qualitative analysis revealed that the model highlighted completed shapes, connecting lines, elements extending beyond the frame, and title annotations, generally aligning with expert TCT-DP criteria. However, visual biases were identified: repetitive elements and heavy shading inflated scores, while some boundary-breaking elements were overlooked, leading to underestimations. These findings underscore the potential of deep learning and explainable AI in creativity diagnostics, while highlighting limitations. Future research should explore specialized test designs and advanced XAI methods to further refine automated evaluation.