[BlueSwordM]

One simple way to improve rav1e at all quality levels is to simple improve deblocking: while the strength selection algorithm is already very optimal, the distortion metric used is SSE, and is not influenced in any way by other forms of more advanced psy RDO metric in rav1e.

That means it blurs blocks to a larger degree than necessary in my testing in anything grainy, and unlike aomenc (unless I can't read the deblock.rs source properly), it doesn't actually include a bias to counteract that to a small degree.

It should be relatively easy to lower deblocking strength somewhat to counteract this in the very short term, and in the mid term, implement frequency weighted SSE to make the deblocking filter application a lot more effective.
This could either be determined through the non-flatness metric used, or by using tx_domain_dist in an interesting way to map out the block frequencies.

Long term wise, it would be possible to even integrate the current psycho-visual RDO metric into it, although that would require some new piping and tuning, but it would allow reuse of data that is already used in activity masking.