Using vmaf_v1.0.16_hfr_3d0h on 59.94 FPS content provides extremely high scores for extremely high CRF values, with results that are predictably devoid of fine details.
At a supposed score of 95.6, the resulting encode has a CRF of 36.5.
For 29.97 content, using the non-HFR model, the score given for nearly transparent encode is 94.6 for a CRF of 29.
So there is zero consistency of VMAF scores across different FPS and models: normal FPS content is underscored while high FPS content is extremely over-scored. These new models are all over the place and provide zero reliability.
Going back to 0.6.1.
Using
vmaf_v1.0.16_hfr_3d0hon59.94FPS content provides extremely high scores for extremely high CRF values, with results that are predictably devoid of fine details.At a supposed score of
95.6, the resulting encode has a CRF of36.5.For
29.97content, using the non-HFR model, the score given for nearly transparent encode is94.6for a CRF of29.So there is zero consistency of VMAF scores across different FPS and models: normal FPS content is underscored while high FPS content is extremely over-scored. These new models are all over the place and provide zero reliability.
Going back to
0.6.1.