Shengwei Tian

Visible-infrared person re-identification (VI-ReID) aims to match pedestrian images captured by visible and infrared cameras, but remains challenging because low illumination and cross-modality discrepancy degrade identity-related appearance cues. Recent text-assisted methods enrich visual embeddings with language semantics. However, their joint representations are still mainly optimized by identity-discriminative supervision and may inherit modality-specific degradation, especially when infrared images contain weak texture responses and missing color cues. This paper investigates whether frozen generative priors can enhance text-assisted VI-ReID without image generation or disruption of the discriminative identity space. We propose ViDiT, a diffusion-prior-guided joint refinement framework that extracts intermediate token maps from a frozen Diffusion Transformer (DiT) and converts them into identity-aware prior features. The prior maps further generate spatial-channel gates and lightweight residual corrections to refine text-visual joint maps. In this way, the final representation preserves visual-semantic discriminability while incorporating structure-aware diffusion priors. To decouple representation learning from post-processing, we report results both without re-ranking and with a multi-source re-ranking pipeline. Experiments on SYSU-MM01, RegDB, and LLCM show that the proposed prior yields its clearest gains on LLCM, where low-light degradation is more severe, while its effect is limited on less degraded or performance-saturated benchmarks.