HarrisonGoldwyn/distributional_SRCNN: Paper Revision

What's Changed Model is now fully deployable fixed data normalization to use only train set stats fixed bug in stage 2 optimization where parameters were being shared between images, destabilizing stage 3. Image-Specific Covariance model now works at smaller stage-2 regularization (covariances show more sample-to-sample variation) implemented stage 2c: covariance parameters are predicted for out-of-sample test images