Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Towards Syn-to-Real IQA: A Novel Perspective on Reshaping Synthetic Data Distributions
Authors: Aobo Li, Jinjian Wu, Yongxu Liu, Leida Li, Weisheng Dong
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments across three cross-dataset settings (synthetic-to-authentic, synthetic-to-algorithmic, and synthetic-to-synthetic) demonstrate the effectiveness of our method. The code is available at https://github.com/Li-aobo/Syn DR-IQA. |
| Researcher Affiliation | Academia | Aobo Li Jinjian Wu Yongxu Liu Leida Li Weisheng Dong Xidian University EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 Distribution-aware Diverse Content Upsampling Strategy Algorithm 2 Synthetic Data Generation Algorithm 3 Density-aware Redundant Cluster Downsampling Strategy |
| Open Source Code | Yes | The code is available at https://github.com/Li-aobo/Syn DR-IQA. |
| Open Datasets | Yes | We conduct experiments on eight IQA datasets: four synthetic distortion datasets LIVE [20], CSIQ [21], TID2013 [22], KADID-10k [4], three authentic distortion datasets LIVEC [23], Kon IQ-10k [24], BID [25], and the dataset PIPAL [26] with both synthetic and algorithmic distortions. |
| Dataset Splits | Yes | Following standard IQA protocols [27, 28], we employed an 80/20 split by reference images for intra-dataset experiments, repeated ten times with median SRCC/PLCC reported. |
| Hardware Specification | Yes | All experiments are implemented in Py Torch and on a server equipped with a 2.10GHz Intel Xeon(R) CPU E5-2620 v4 processor and four NVIDIA GTX 1080 Ti GPUs. |
| Software Dependencies | No | All experiments are implemented in Py Torch and on a server equipped with a 2.10GHz Intel Xeon(R) CPU E5-2620 v4 processor and four NVIDIA GTX 1080 Ti GPUs. |
| Experiment Setup | Yes | The mini-batch size is set to 32, with a learning rate of 2 x 10^-5. The Adam optimizer, with a weight decay of 5 x 10^-4, is used to optimize the model for 24 epochs. During training, one 224 x 224 patch is randomly sampled from each image, with random horizontal flipping applied for data augmentation. ... The distorted-feature distances threshold Tdf and the ground-truth distances threshold Tg are set to 0.1, and 1 (for MOS values in [0, 10]), respectively. ... The best results are obtained when k = 5 and Trf = 0.05 ... The model performs best with Tu = 20 |