OneAlign

OneAlign is a large multi-modality model for zero-shot visual scoring that unifies image quality assessment (IQA), image aesthetic assessment (IAA), and video quality assessment (VQA) using a discrete-level-based syllabus. Developed by Nanyang Technological University, Shanghai Jiao Tong University, and Sensetime Research, the model is built on the mPLUG-Owl2 architecture and trained on a combination of datasets including KonIQ, SPAQ, KADID, AVA, and LSVQ. Instead of regressing direct scores, it emulates human subjective studies by teaching large multi-modality models (LMMs) with text-defined rating levels, achieving state-of-the-art or competitive performance across all three tasks.

The model scores images and videos on a 1–5 scale. Below is an example input image and the syllabus illustration used during training:

Benchmark Performance

OneAlign achieves leading results on multiple IQA, IAA, and VQA benchmarks. The following table compares its performance against previous state-of-the-art methods across seven IQA datasets, using Spearman/Pearson/Kendall correlations:

On the AVA_test aesthetic benchmark, OneAlign achieves a Spearman correlation of 0.823 and Pearson of 0.819, surpassing prior methods. For video quality assessment, it sets new state-of-the-art on LSVQ_test (0.886/0.886), LSVQ_1080p (0.803/0.837), KoNViD-1k (0.876/0.888), and MaxWell_test (0.781/0.786).

Users must comply with the LLaMA-2 license when using this model commercially. The model is hosted as a managed API on gigarouter, requiring no local installation.