Q-Align

Why we included this project

Q-Align is a single vision-language model that scores how good an image or video looks, covering image quality, image aesthetics, and video quality through discrete text-defined levels. One checkpoint can replace several task-specific scorers, which helps with quality control, content moderation, or aesthetic ranking. Since it builds on the mPLUG-Owl2 architecture, it fine-tunes efficiently to downstream datasets rather than requiring training from scratch, a real advantage when your scoring criteria differ from the public benchmarks. The repo ships pretrained weights, a model zoo, and a Hugging Face demo, and the model can also be loaded through the IQA-PyTorch metric library, so you can test it before committing to a workflow.