Originally posted by mattt Do you suppose that's due to the baked raw?
My guess is that the answer is yes, but one should have more information on the exact machine learning workflow behind DeepPrime. My guess (again, pure speculation) is that it consists of a generic deep learning model that is trained on their full database of cameras and sensors to learn a "universal" set of weights. The model architecture is most likely a convolutional neural network of some sort. This pretrained model is then fine-tuned on each camera model, perhaps by only adjusting the weights of the final layers of the network. This is a relatively common workflow in AI applications for image processing (the idea is related to the concept of
transfer learning https://en.wikipedia.org/wiki/Transfer_learning). It is more or less established that Pentax models with the accelerator chip apply a level of preprocessing to the RAW files that is "well above average". This makes Pentax files a sort of outlier in this dataset. It is therefore reasonable to expect that the model adapts best to the most common type of RAW files that are not preprocessed so aggressively. Naturally, due to some level of fine-tuning of the model by retraining on a small camera-specific database (I suppose), it works acceptably well even on these "peculiar" raw files. As a matter of fact, I am under the (subjective) impression that Deep Prime somewhat treats the files from my K3 "better" than the K-3iii...
I think that maybe one of the reasons behind Pentax's choice of not allowing to disable the accelerator chip is that handling of Pentax raws by noise reduction algorithms would become a nightmare, because the software should infer from the EXIF whether or not in-camera NR was applied, and treat the two types of files as if they came from two different camera models. Only very basic noise reduction algorithm with no camera-dependent profiles would work reliably, meaning that the user should manually adapt the available parameters. But anything more sophisticated would be a mess unless there is the option to select different sub-profiles with and without in-camera NR.
All in all, I am not a fan of this choice, the simple solution would be to have it
always off in RAW and apply it to Jpeg only...