Multimodal Prompt Learning with Vision--Language Foundation Models for Automated Storyboard Assessment and Formative Feedback in Film Education

Abstract Storyboards articulate cinematic meaning through compositions, actions, captions, and the temporal relations between shots. Assessing them is challenging because recognizable shot attributes do not uniquely determine narrative quality, and persuasive feedback may lack evidence in the submitted panels. We formulate storyboard assessment as criterion-conditioned inference over an editable temporal graph and propose Criterion-constrained Rubric-aware Feedback Transport (CRaFT). The approach couples frozen vision–language encoders with low-rank criterion prompts, unbalanced optimal transport between multimodal evidence and rubric anchors, and ordinal score prediction. A graph-consistent counterfactual objective scores candidate revisions by their localized, criterion-specific effects, while conservative selection can abstain when support or uncertainty is inadequate. On the reported public-data evaluations (MovieShots, FullShots, AVE, and MovieNet), CRaFT obtains macro-F1 scores of 77.4%, 73.1%, 72.3%, and 70.0%, respectively, exceeding the strongest listed prompt baseline by 5.2–5.9 percentage points across these tasks. The reported AVE sequence-order accuracy reaches 80.1%, while the mean ordinal agreement on constructed rubric proxies is 0.69 quadratic-weighted kappa. The same system supports a teacher-aligned five-criterion assessment protocol with independent validation of cited evidence, edit benefit, and model abstention. This separation of measurable cinematic proxies from instructor judgment is central to the proposed evaluation design.