I have two seperate videos that I’ve tried different settings and all of them add audio lag.
The videos are ~65 mins each before setting the Marks.
I’m running the following:
Windows 11 Pro 24H2
OS build 26100.6584
Processor - 13th Gen Intel(R) Core™ i9-13900KS (3.20 GHz)
RAM - 48.0 GB
Video - NVIDIA GeForce RTX 5090
The video resolutions are 720x480 and I used the following command to capture them: #00 Original Capture:
set FFMPEG=C:\Apps\VHS\ffmpeg\bin\ffmpeg
set OUTDIR=C:\Apps\VHS
set DEVICE_VIDEO=CY3014 USB, Analog 01 Capture
set DEVICE_AUDIO=CY3014 USB, Analog 01 WaveIn
set FILENAME=capture_%date:~10,4%%date:~4,2%%date:~7,2%_%time:~0,2%%time:~3,2%%time:~6,2%.mkv
set FILENAME=%FILENAME: =0%
The audio is perfectly in sync in both videos before using both Starlights.
For Starlight Mini, I choose the following:
Enhancement - On
Output resolution - 2x Upscale
AI Model - Starlight
Model variant - Starlight
For Starlight Sharp, I choose the following:
Enhancement - On
Output resolution - Original size
AI Model - Starlight
Model variant - Starlight Sharp
The weird part is that I beta tested Topaz Video AI 7.2.0.2b and the audio was in sync for Startlight Mini. However, the quality of the converstion was not that great, which this version fixed.
The issue is still present. However, I can split the video into 1-minute segments, which results in only a few milliseconds of delay when merging them back together. I can notice it, but only because I am actively looking for it. Others who were not aware of the issue did not notice it at all. So it is not a fix, but it does serve as a workable workaround.
No, I haven’t tried that. I don’t think it would make a difference anyway, since the issue isn’t with the audio itself. What’s actually happening is that the video gradually gets shorter over time, which causes the audio to fall out of sync and start lagging.
Honestly, I usually work with large videos, anywhere from 30 minutes to 2 hours, and it’s tough to keep a process running nonstop for a couple of days. Because of this bug, I was forced to split the videos into smaller segments, and in hindsight, that’s probably the best thing that could have happened. Now I can close the program or let my computer reboot for Windows updates without losing two full days of processing. At worst, I only lose about 40 minutes, which is how long it takes to process a one-minute clip at 2x with Starlight Mini.
Merging the files back together isn’t difficult and doesn’t take long. I have a simple PowerShell script that uses FFmpeg to handle it.
Yeah, but that approach introduces temporal inconsistencies at the merge points which can lead to visible abrupt changes in the upscaled image (as each segment of course cannot have Intel about the previous/next block).
This is what - depending on the source - is imo highly annoying when using SeedVR2.
