[mp3dom]

Originally Posted by Sagittaire

x264 never encode high frequency at q5 and low frenquency at q25 with AQ. Moreover dark area can have low and high frequency for block. There are scaling between frequency and complexity for block. AQ with Luma/contrast are typicaly psy optimisation.

With the command line posted in first post it happends what I've said (probably not everytime but at least on the frame I've checked). Verified with StreamEye. In the 4th screenshot the zones with artefacts (and I cannot trust you if you say that you can't see that ringing) I have a qp>25 (some blocks in the contrast area have, if I remember correctly, a qp near 29).

but you use boost because without boost it's really hard te see differences ... isn't it?)

No it isn't. Read my point #2 in previous post. I say again that 3rd and 4th screenshot artefacts are visible to me during playback but probably not for a casual viewer (who can have wrong calibrated equipment, or not a trained eye). But as a video compressionist my job is to have a real good encode without taking into account other variables like user experience, user trained eyes or user equipments.

Perfect codec must (in HVS sense) make encoding like the source master ... but it's really not the case for compressionist ... simply because lossy codec 4/2/0 can't never make encoding like the maters source 4/4/4 or 4/2/2. Pre-filtering is here for degrade less important part of source (for my eyes) and privilegy the most important part of the source (always for may eyes) if the codec can't make perfect encoding and it's always thes case for lossy codec. And it will be always like that ... for lossy codec.

Preprocessing in the sense of dithering down a 10bit master to 8bit is needed and necessary, filtering to keep original color smoothing is another good choice to preserve the original intent of the author. Denoising a master to help the encoder, to me, is not a good choice. Just to know, do you cut some audio frequencies in a PCM file to help a Dolby encoder to output a lossy Ac3? Or do you apply filters that eliminate everything under a -40/-50 dB just because "it can't be heard"? I don't think so... and I don't think audio engineers do this.

If the other codec make good job for your eyes, use simply the other codecs. Anyways encoding grain in dark part area will be always time loss (and bytes loss) for the pro ... it's like that.

Just to point out, the screenshots posted as a 'source' are in reality the output from another AVC encoder that evidently keep more details in dark areas. It's not the same as the real source (the 4th screenshot of the real source is better), but very close and I've decided to use it as a 'reference' for what I mean/would like to have from x264. It's not encoded by me (because I don't own that encoder) but by another studio of another country and another company (japanese animation are licensed for every country at different companies). That encode also starts in a 'worst' position than a general x264 encode because (accordingly to StreamEye) it was made with only 2 bframes and no 4x4 partitions so it's even more 'inefficient'. If another encoder can do that job, I think even x264 should do at least a similar job.
The x264 'engine' is really very good because it can find vectors/references/partitions and handle B frames in a very better/efficient way than any other AVC encoder. Efficiency that seems to not take account of dark areas. It's just that, to me, seems more 'enhanced' to provide good quality at low bitrate (something that any other AVC encoder can't provide) but it continues to work in the same manner even when the bitrate is very generous.

The question is: Is it possible, tuning parameters (which?), to obtain a result more close to the source even in dark areas? Or it's something that x264 is not designed for? For now I've tried to lower deadzones to 2, disabling mbtree, lowering aq-strength to 0.5 and lowering psy-rd. While the result is a bit better, it's not enough as 3rd and 4th screenshot continues to have that kind of artefacts.