Remember the 3D mark scene is PERFECT for Deep learning as the training data is literally the same footage for only a few minutes. It isn't as varied as a real game with hundreds of hours of game play and angles for rendering versus a scene that is literally on rails. Also remember sharpness is normally due to lack of AA not extra. AA should smooth surfaces. That is why when you do something like SMAA Ultra you add a sharpening filter to get back some mild texture detail that is normally lost. I did this a lot in Borderlands with 5K downsampling and forced SMAA through sweetfx.