Conformal No-Reference Image Quality Assessment With Entropy-Adaptive Prediction Intervals

No-reference image quality assessment (NR-IQA) models return a single quality score and give no indication of how far it may be from the human opinion it predicts. We present a conformal wrapper that turns any pretrained NR-IQA predictor into a prediction interval for the mean opinion score with a finite-sample coverage guarantee, without retraining and without distributional assumptions beyond exchangeability. The interval is centered at the original prediction and its half-width is the product of a calibrated quantile and an image-dependent scale computed, without training, from the spatial non-uniformity of local entropy in luminance, texture, and contrast. The guarantee holds for any scale, so the statistic affects only efficiency, and it makes the intervals narrower than the constant-width alternative while their width correlates with the standard deviation of human ratings that the procedure never observes. Per-group calibration extends the guarantee to individual distortion types, and likelihood-ratio weighting restores it under dataset shift.