PortraitCraft: A Unified Evaluation Benchmark for Portrait Composition Understanding and Generation

This article introduces the PortraitCraft benchmark, which is based on approximately 50,000 carefully selected portrait images. It provides multi-level structured annotations, supports two major tasks of composition understanding and generation, fills the gap in specialized evaluation benchmarks for portrait composition, and offers a comprehensive evaluation framework for portrait aesthetic assessment and controllable generation research.

Portrait composition is a core element of portrait aesthetics, determining the balance of the画面, visual flow, and emotional expression. However, existing datasets and benchmarks have limitations: 1. Coarse-grained aesthetic scores lack fine-grained interpretability; 2. General image aesthetic datasets are not designed for portrait composition; 3. Unconstrained portrait generation models rarely consider composition constraints, leading to inconsistent composition quality in results.

PortraitCraft integrates composition understanding and generation into a unified system. The dataset is based on approximately 50,000 selected portrait images and provides multi-level annotations: global composition scores, 13 composition attributes (such as adherence to the rule of thirds, gaze guidance, etc.), attribute-level explanatory text, visual question-answer pairs, and composition-oriented generation descriptions.

Task 1 (Composition Understanding) includes three subtasks: score prediction (predicting composition quality scores), fine-grained attribute reasoning (evaluating performance on 13 composition attributes), and image-based visual question answering (answering questions about composition details); Task 2 (Composition-Aware Generation) requires models to strictly follow composition descriptions to generate portraits that meet the requirements.

The standardized evaluation protocol includes clear data partitioning, evaluation metrics for different subtasks (e.g., correlation coefficients for score prediction, composition fidelity for generation tasks), and baseline results. Academic value supports fine-grained understanding, interpretable evaluation, and controllable generation; practical applications include photography education (real-time feedback), content creation (auxiliary generation), and image editing (intelligent optimization).

Current challenges: balancing subjectivity and objectivity, insufficient fine-grained understanding ability, multi-objective optimization of generation quality and composition constraints. Future directions: multi-modal fusion, personalized aesthetics, cross-style transfer, real-time application optimization.

PortraitCraft fills the gap in specialized evaluation for portrait composition, provides multi-level annotations to support interpretable research, unifies understanding and generation tasks, and establishes standardized protocols and baselines. It lays the foundation for fine-grained composition understanding and controllable generation in the fields of computational photography and generative AI, promoting the progress of portrait photography AI technology.