About the label of image quality

[Lxinyu1999]

Hello, thank you for your great work! I am a new student in the field of image quality analysis. I think this is a very meaningful work.

I am curious whether DifFIQA operates under a supervised or unsupervised learning paradigm. Does the original dataset used for training the model contain quality labels for the images?

Thank you very much for your help.

[LSIbabnikz]

Thank you for your kind words.

Regarding your question about how DifFIQA works. The basic approach of DifFIQA can be considered as an unsupervised approach, since it uses a diffusion model and any face recognition model to estimate the quality of the samples. While the diffusion model is trained separately, no information about the sample quality is used in its training process. However, the extended DifFIQA(R) approach can be viewed as a supervised approach that distills information from the basic DifFIQA approach into a quality regression model.

I hope this answers your question, otherwise do not hesitate to ask further questions.

[ChuRuaNh0]

Can you share dataset (images) in training regression model DifFIQA(R)? Thanks a lot.

[LSIbabnikz]

The regression model was trained on a subset of the VGGFace2 dataset, you can see which images were used in the file containing the quality score labels. I am not allowed to share the dataset, but I am sure you can find it elsewhere (try the InsightFace repository).