You are correct, this only adds 3d effect without adding necessary distortion for 180SBS viewing… though you may be able to use some ffmpeg filters (available in iw3 the “-vf” box), pre-process the video or use settings in your player to give it the distortion it needs (lens/projection settings in heresphere come to mind)
I personally find the 3D effect (especially with newer models like VDA) quite compelling and immersive, especially with POV content, but can understand it’s not everyone’s cup of tea.
The conversion time can be long as well, but once the process is dialed in it only takes a couple clicks so I’ll usually queue something up before bed.
I will also look at the VDA_L you mention in a previous comment. Now I understand what to expect I will continue trying a few different videos and also using funflowscript to try and put together some acceptable Maimy ASMR videos. If I succeed I will share them on this forum.
I would strongly recommend anyone doing conversions to play around with the Processor settings when using Distill_Any_L. I went from ~10.3 fps => 15.5 fps (50% performance uplift). Your mileage may vary though depending on your hardware. My GPU was underutilized on the default settings.
The 2 parameters to adjust are Depth Batch Size and Worker Threads. From my understanding, Depth Batch Size is the number of frames that are pulled in a batch and worked on at a time and Worker Threads are the number of parallel processes working on separate frames in the batch.
Depth Batch Size directly increases the GPU dedicated memory usage (not the same as system RAM). Setting it too high caused my computer to lock up. Had to hard restart it.
I found the best results by increasing Depth Batch Size to be as high as possible, then incrementing Worker Threads until I saw a performance decrease. For testing, I picked a video and waited for it to do about 300-500 frames before recording the results down.
All my existing settings above are applicable to VDA - the new checkbox options for Scene Boundary Detection and Preserve Screen Border are recommended as well.
I do not personally tweak the depth batch size as I only have a 12GB GPU, VDA_L is very large and I am often processing 4k content which stretches my GPU thin, especially when using my PC for desktop purposes.
I have also moved to using avi output (utvideo lossless codec) and then using handbrake for the final encode since I can use my own custom encoder parameters but that is a bit out of scope for this topic I imagine.
Anything else need to be changed other than avi for the output? I tried iw3 a bit and tried .avi and the final size of a 6 min pmv was 78 gigs. When run for mp4 it was only 2 gigs I thought avi was supposed to be smaller but… woah
AVI is is the middle step. It’s a lossless codec (UTVideo) and massive file size intentionally.
You can then take that AVI and use it with a program like handbrake or ffmpeg to do the final compressed encode to h.265 or av1. This is only superior if you use custom encoding parameters compared to just using CPU with slow preset in iw3.
Regarding the one suggestion for hevc_nvenc, this will speed up processing but produce a file size 25-50% larger compared to CPU encoder.
Most ai support AMD with rocm now. Is this going to get AMD support or is this project finished? only asking becouse i have a rx 6900 xt which is capable of running almost all ai workflows ive encountered sometime only with zluda support that is. (tho I could always use my GF’s pc as she uses a rtx 2070 with 8gbs vram. and doesn’t mind my randomly installing ai stuff on it.) my pc has a lot more storage on it only reason i’m asking.
When run for mp4 it was only 2 gigs I thought avi was supposed to be smaller but… woah
That’s the refresh rate for someone on a heroic dose of amphetamines…
Got it. I’ll test those settings against my other with the file I just did. Thx