Multi-reference image generation aims to synthesize images by integrating attributes from multiple reference images under textual instructions. As the number of references increases, the task necessitates complex semantic comprehension, such as correctly associating attributes with the intended subjects and planing out coherent spatial arrangement between subjects and their environments. Existing approaches, which rely solely on natural language instruction, often fail to capture these complex intentions precisely, leading to semantic misalignment and inconsistent generation. We identify two key factors behind these limitations: natural language instructions are often verbose and ambiguous, and high-quality multi-reference data is scarce. To address these issues, we propose StructGen, which employs a structured, dictionary-like format to encode multiple reference images, thereby enabling explicit and unambiguous specification of generation intentions. To support this design, we construct a structured dataset based on high-quality real images and develop a corresponding training framework, along with a dedicated benchmark for challenging multi-reference scenarios. Extensive experiments on both public benchmarks and our proposed benchmark demonstrate that StructGen consistently outperforms existing methods on both semantic alignment and detailed reference-generation consistency, especially under complex instructions with multiple references. The code is available at https://jianingpeng0382.github.io/StructGen/
Jianing Peng, Mengyu Wang, Henghui Ding et al.· 0 citations
Automated depression assessment from clinical interviews is a challenging intelligent healthcare task because depressive states are reflected not only in verbal responses, but also in acoustic and visual-behavioral cues. Existing multimodal methods have improved prediction performance, but many still process interviews as generic long sequences and provide limited structured evidence for interpreting model outputs. This paper presents MIDAS, an interview-aware multimodal framework that preserves the question-response structure of clinical interviews for depression assessment. MIDAS organizes interviews into question-response pairs to preserve the interaction structure of clinical interviews, and integrates linguistic, acoustic, and visual-behavioral representations through gated multimodal fusion and shared large language model-based contextual modeling. The framework jointly supports depression classification, PHQ-8 severity prediction, and structured multimodal evidence summarization. Experiments on the DAIC-WOZ development set show that MIDAS achieves a Macro-F1 of 88.10% for depression classification, with Dep-F1 and Con-F1 scores of 85.76% and 90.45%, respectively. For PHQ-8 severity prediction, MIDAS achieves an MAE of 3.03 and an RMSE of 3.93. Ablation results indicate that acoustic and visual-behavioral information provide complementary benefits over linguistic-only modeling. Qualitative examples further show that MIDAS can organize modality-specific behavioral cues into structured evidence summaries associated with model predictions. These results demonstrate the potential of interview-aware multimodal modeling as a research framework for depression assessment, while the structured evidence summaries are intended to support the interpretation of model outputs.
