I gave ChatGPT SOL ten clips from a grainy 1080p24 source, the full-source MediaInfo report, and my AviSynth+ template. I gave it my tuning method and left the test encodes and visual comparisons to it. I only looked at the final result.
My instruction was to start with two-pass encodes at one fixed bitrate slightly below what I expected the final encode to need. The tighter bit budget makes small differences easier to spot. Once the chosen settings look transparent on a second set of excerpts, lock them and find the highest CRF that still shows no visible quality loss.
SOL ran 11 two-pass candidates at 17,000 kb/s through AviSynth+ → Avs2YUV → standalone x264 in its Linux sandbox. It used veryslow in both passes and kept psy-trellis at zero. It compared AQ modes 1, 2, and 3 at the same strength first. Mode 2 gave the best overall match, so it tested strengths 0.8, 1.0, and 1.2 within that mode and picked 0.8. It also tested deblock and psy-RDO.
SOL compared the encodes against the cropped source at 133 timestamps where every candidate encoded a B-frame. It inspected fine hair strands, detail in deep shadows, skin texture on faces in close-up, and detail in distant backgrounds. It checked smooth gradients for banding and looked at how closely the grain texture matched the source.
It checked the chosen settings at 20,000 kb/s on a second set of excerpts, then compared CRF 16, 17, 17.5, and 18. It picked CRF 17 for these samples, which averaged 19.25 Mb/s. It reported that some grass texture and fine hair started to soften at 17.5.
The run shown took about 19 minutes. The big time saver for me was having SOL inspect the matched frames and make the comparisons itself. For those who tune x264 this way, has anyone else handed off the actual frame comparisons like this?
[Attachment 94409 - Click to enlarge]
[Attachment 94410 - Click to enlarge]
+ Reply to Thread
Results 1 to 1 of 1
-
Last edited by maas3n; 21st Nov 2026 at 20:42. Reason: added more info
Similar Threads
-
FFMPEG: Starting a 1-pass Encode with the mbtree and log stats of a 2-pass?
By skh in forum Video ConversionReplies: 1Last Post: 13th Dec 2023, 14:37 -
how good is ChatGPT at coding video conversions ?
By hydra3333 in forum Video ConversionReplies: 7Last Post: 29th May 2023, 03:39 -
Is x265 half size reduction compared to x264 a real thing ?
By Yertq in forum Video ConversionReplies: 13Last Post: 23rd Jun 2022, 07:04 -
x264 in ffmpeg always uses 3 ref frames
By Travillion in forum Video ConversionReplies: 3Last Post: 5th Jun 2022, 00:30 -
2-pass AVC H.264 encoding = 5% bitrate improvement over 1-pass?
By pxstein in forum Video ConversionReplies: 7Last Post: 25th Jan 2022, 11:34



Quote