CRF algorithm implemented in x264

CRF is no rate distortion algorithm but a bitrate control mode. CRF lets you specify a constant picture QP for the whole sequence like constant QP, but in float. This picture QP is then scaled by the 'complexity' of the local pictures if I remember correctly, someone correct me here please. the mbtree algorithm and aq mode, which work on a macroblock level, then add to this picture QP to derive a QP for a macroblock. I think when mbtree is enabled the complexity of a picture is always 1.0, ie no scale, but again I am not so sure about this one. So I suggest to look at the mbtree, aq mode and bitrate control code. There is a paper on mbtree here:

"CRF" mode without MB-Tree is pretty much just "Quantizer Curve Compression" (qcomp) with a fixed pre-defined scaling factor. As opposed to the "2-Pass" and "1-Pass ABR" modes, which also are based on the very same "qcomp" algorithm but constantly re-adjust the scaling factor in order to hit a specific target average bitrate. In "CRF" mode, the fixed scaling factor is computed from the given CRF value, but is not exactly the same as the given CRF value. The following applies to all RC modes: Simply put, the "qcomp" algorithm increases the QP for more "complex" frames and lowers the QP for less "complex" frames, where the "complexity" is the estimated bit cost of the frame. MB-Tree, which is now enabled by default, replaces "qcomp" and extends that idea to the macro-block level. The QP of an individual macro-block now depends on how much that block is going to be referenced in the future. Adaptive quantization (AQ), which also adjusts the QP's of individual macro-blocks (within the current frame), is kind of "orthogonal" to both, MB-Tree and the traditional "qcomp" algorithm. See also:

Was aElahi referring to rate distortion rather than rate control? That is, the psychovisual enhancements (--psy-rd)?