Abstract
<title>Abstract</title> <p>Recovering an ordered drawing procedure from a completed image that is, inferring not only what strokes compose a drawing but in what sequence they should be executed remains a largely unexplored problem in structured visual prediction. We introduce StrokePlan-RNN, an image-conditioned autoregressive model that predicts ordered stroke sequences from static face line art. The model takes a completed line drawing as input and produces a replayable vector drawing plan describing a plausible construction procedure. We construct a dataset of 640 face line art images annotated with ordered stroke sequences organized by semantic facial groups, and train an encoder–decoder architecture (ResNet-18 encoder, two-layer LSTM decoder) end-to-end with Smooth L1 reconstruction and binary cross-entropy stopping losses. On a held-out test set, StrokePlan-RNN achieves a mean point error of 3.95 pixels and a sequence length error of 3.1 strokes, substantially outperforming both a random-order baseline and a decoder-only ablation. Qualitative replay analysis suggests that the model captures aspects of coarse-to-fine procedural organization, though the current evaluation primarily measures geometric reconstruction accuracy rather than ordering quality directly. We discuss the gap between reconstruction metrics and procedural evaluation, identify systematic failure modes, and outline directions for order-aware evaluation and stronger structural priors.</p>