StrokePlan-RNN: Predicting Ordered Drawing Procedures from Face Line Art

<title>Abstract</title> <p>A finished drawing preserves its appearance but not its construction the ordered stroke sequence by which it was made, and by which drawing is taught, is lost once the image is complete. We ask whether a plausible construction order can be recovered from the finished image alone, and formalize this as ordered drawing-plan prediction from static face line art. To study the task we build a controlled dataset of 640 face line drawings annotated with semantically grouped, temporally ordered stroke sequences, produced with a custom annotation application, and we propose StrokePlan-RNN, an image-conditioned baseline that pairs a fine-tuned ResNet-18 encoder with a two-layer LSTM decoder and separate stroke and stopping heads. Because standard geometric objectives do not penalize sequence errors, we evaluate with order-aware metrics, namely Pairwise Order Accuracy, Kendall's τ, and Semantic Group Timing, alongside geometric measures. On the held-out test set the model attains an average point error of 3.95 pixels, outperforming a random-order baseline and a decoder-only ablation, and orders strokes well above chance (POA 0.74, τ 0.48), recovering the coarse-to-fine timing of all seven facial groups to within 0.05 of their normalized ground-truth positions. Inspection of the predicted replays identifies four systematic failure modes, premature stopping, over-generation, local ordering errors, and dense-detail instability, each traceable to the architecture or training protocol. The task formulation, dataset, baseline, and evaluation framework establish that construction order is partially recoverable and directly measurable; whether the resulting replays are instructionally useful remains an open question requiring human evaluation.</p>