one-align

OneAlign is a zero-shot image and video quality and aesthetics assessment model. It unifies image quality assessment (IQA), image aesthetic assessment (IAA), and video quality assessment (VQA) into a single model, achieving state-of-the-art or competitive results across multiple benchmarks without task-specific fine-tuning. The model processes inputs such as the example below and outputs a score in the range [1,5].

Key Strengths

• Unified multi-task: a single checkpoint handles IQA, IAA, and VQA.
• Zero-shot evaluation: no additional training or dataset adaptation needed.
• Outperforms previous SOTA on several datasets (e.g., KonIQ, KADID, LIVE, CSIQ, AVA, LSVQ).
• Supports both image and video inputs.

Benchmark Performance

Results are reported as Spearman/Pearson/Kendall correlations. The model is trained on IQA, IAA, and VQA tasks jointly (referred to as OneAlign in the tables).