The AI approach is interesting but there are just too many problems with it. MozJPEG is actually using something similar called trellis quantization or soft-th

The AI approach is interesting but there are just too many problems with it. MozJPEG is actually using something similar called trellis quantization or soft-thresholding. The idea is to further quantize DCT coefficients in noisy areas. There are some limitations [1] though. But it's not an issue with locally sensitive metrics and can be improved with edge detection. Another challenge here is to estimate the current image quality and the target error. [1] https://github.com/mozilla/mozjpeg/issues/129 https://github.com/mozilla/mozjpeg/issues/129