A large-scale visual distortion assessment instruction-tuning dataset and benchmark for UGC images (11.5K images, 36K distortion boxes, 534K instruction data) that unifies distortion grounding, low-level perception, and quality description, built with a grounding- and perception-guided chain-of-thought pipeline.

A modular agentic framework for explainable pairwise IQA: a Planner, an Executor, and a Summarizer collaborate with dimension-specific analysis tools, trained with SFT and RL, to reach traceable, expert-aligned preference decisions.

A one-step diffusion Real-ISR framework that feeds both the LR image and its noisy version into an image-editing diffusion model, with Negative-aware Feature Fine-Tuning (NF²T) that uses aggregated IQA rewards to build positive and negative optimization directions.

Report of the Detailed IQA track of MIPI 2025, which I organized; the track benchmarks distortion grounding, low-level perception, and quality reasoning.

IOVQA fine-tunes Qwen2.5-VL for AIGC video quality scoring with integer-only labels and a target-mask loss on the score tokens; 3rd place in the VQualA 2025 GenAI-Bench challenge.