I used 18 FPS, and 25 FPS test videos and ran all the interpolation models at 90, and 120 FPS, I also did a 240 test frame.
But I thought it was a good idea to export them as PNGs, but RIP dropbox as they’re too big, so I’ll encode them in smaller files, and upload them when I’m done, sorry about the delay!
Now for the good news, the 18-90 FPS/25-90 FPS videos turned out fine! Whoo HOO, but better double check just to make sure all the models that I tested behave, as I did not have time to look/do a quick 200-frame encoding job for all of them.
Bad news is that the 18-120/25-120 FPS videos did not do as well…
As the 25–120 FPS video still repeats every sixth frame, and the 18-120 FPS video repeats intermittently @ every 3rd, fourth, second frame, it’s pretty much a mess… At 24–240 FPS it repeats every 6th frame, and at 18–240 FPS is a bloody mess as well!
EDIT: Whelp, I got it under 2 GB, again, sorry for the delay, I get lost in the minutiae of the task and hyper focus a bit…
Thanks for the files. Some of the issues can be explained by model limitations.
Here are some facts about both apollo models, they can only generate fixed number of frames between two frames. Apollo can generate 7 frames and Apollo fast can do 3 frames, so the frame repeats are expected in most of the cases you mentioned above.
I would recommend running multiple passes if you need more frames generated from apollo or apollo fast.
Chronos and Chronos Fast at best can do about 15 frames, anything beyond that will give repeat frames or bad results.
Chronos and Chronos Fast do exactly that. Apollo and Apollo Fast generate frames at fixed intervals, so sometimes it might get mapped to the same frame if the calculations are not exact multiples.
It’s interesting to see though, that Apollo V8 is usually better at handling animation (Apollo gives out almost real calculated frame, instead of a ghost, usually), which has a “foreground” (protagonists, etc.) frame drawn every two “background” frames. I usually use the video’s original frame rate to try combatting the choppy animation, and, for example, the old “Dungeons & Dragoons” cartoon is a great field for experiments, especially because all episodes are available officially on Youtube.
That might change though, as the “engine” behind VEAI gets optimized further.
I’d like to ask about one thing - could VEAI use Nvidia’s “RTX super resolution” (quite interesting video quality enhancer), DLSS and NV’s frame generation (for those, who have Ada Lovelace GPUs) as an option in the future?
This explains a lot, thanks, for the explanation and confirmation!
Well, I figured that 90 is an excellent limit as 60 is too small for my needs, I did double up the 90FPS test video, and it did just fine, the only downside is that I have to use maths, for any de-dup’d source videos, as maths are my kryptonite!
I’m wondering if a new future model can have full frames, as this really is the best interpolation method that I’ve come across, but then again, I’m just glad you guys employ a custom interpolation method instead, as optical flow is rubbish!
I hate optical flow, by the way!
Ahh, thanks for the confirmation, I suspected as much!
Time to bust out the calculator then, as that makes so much more sense; now I have the proper info going forward, though maths don’t care for me that much!
Still 100x better than being stuck with and using optical flow!
Preview processing is still a bit (or actually again) broken. On occasion its stops at 99/100% for a bit then the frames have to catch up and then after 10-20s it becomes green and plays…
EDIT:
Also sometimes it doesn’t let you skip though the frames like it did with 3.3.6. Have to wait until it finishes the preview render…
yes, i can confirm that. I’m waiting over 30 seconds for the ‘last frame’ to finish.
Also, the previews (and exports) don’t play in sync currently, this from the last few updates.
