Chen Zhao 赵晨

Structured Stateful Relational Scripts make persistent identities, evolving states, sound-source bindings, and cross-shot event continuity explicit. MSAV-Script-100K provides 100,000 curated native multi-shot audio-visual clips with structured supervision. This work was carried out at Nanjing University and Meituan.

TCA-Captioner improves audiovisual captioning through iterative data refinement and a benchmark that evaluates audio-visual binding and temporal reasoning.

LUVE combines motion generation, latent upsampling, and dual frequency experts to generate ultra-high-resolution videos with coherent motion and fine detail.

SHINE adapts pretrained diffusion priors for training-free image composition and introduces ComplexCompo to assess challenging lighting and reflection conditions.

UltraHR-100K introduces a curated ultra-high-resolution image dataset and frequency-aware post-training to improve fine-detail synthesis in text-to-image diffusion models.

SN-DCR combines semantic and style contrastive regularization with spectral normalization to preserve structure and texture in unpaired image translation.

This study decomposes UHD image restoration into zero-, low-, and high-frequency stages and introduces the LSUHDIR benchmark for detailed evaluation.

PA-Diff guides a diffusion transformer with underwater imaging priors and implicit neural representations to recover clearer underwater images.

SPWM uses a spiking dual-pyramid wavelet architecture to capture long-range image structure while reducing restoration compute and energy consumption.

ERR restores UHD images through three coordinated stages for global enhancement, coarse content recovery, and high-frequency texture refinement.

WF-Diff combines wavelet-Fourier information interaction with residual diffusion to restore low- and high-frequency information in underwater images.

CCLformer removes rain without paired training images through cycle and location contrastive learning, transformer features, and SAM-guided structural consistency.

SFGNet couples dense spatial-frequency fusion with a gradient-aware corrector to improve underwater image colors, details, and structure.

OmniSF combines spatial and frequency representations with dynamic feature fusion to reconstruct face images efficiently.

AddSR uses timestep-adaptive adversarial diffusion distillation and prediction-based self-refinement to accelerate blind super-resolution while preserving perceptual quality.

STAR uses text-to-video diffusion priors, local detail enhancement, and a dynamic frequency loss for temporally consistent real-world video super-resolution.

SDIR predicts precipitation by preserving large-scale weather structure and progressively refining fine-scale details under physical spectral constraints.

STTrack combines persistent temporal state tokens with spatial fusion to track objects across RGB and complementary thermal, depth, or event modalities.

UMMF uses uncertainty-aware training and fusion to combine structural MRI, functional MRI, and diffusion tensor imaging for neurocognitive impairment prediction.