dawkinscm wrote:Did you do the Rife 4.25 vs Topaz comparison?
Yes, I did.
Topaz 7.x with "Chronos fast" model seems to give best performace vs quality for Full HD 25 to 60 FPS conversion.
On my specs, it takes about 2x time of input video, with same output resolution.
Converting to native LCD FPS is primary benefit. Resolution is not so important because HW upscaling is doing good enough job, it makes processing time much slower and final file much larger.
Very user friendly, usually works out of the box, price is a bit higher but final quality is better.
It can do motion deblur, stabilization etc, which is more HW demanding (and out of reach for live playback on current high-end GPUs).
If I want to go for quality, Topaz is the way. Or preparing video so I can play it somewhere else without relevant GPU for SVP.
SVP can play on the fly, which I consider it's main benefit and best choice for overall usage.
So I use both.
I have noticed some remarks that those doing RIFE models are now interested in other stuff so not sure if we will see any newer RIFE models soon.
Perhaps I may be wrong.
There is other exiting stuff, kind of exploring future ways of enhancing videos.
Topaz released it's Starlight Mini model, which is diffusion model for video upscaling, locally rendered if you have relevant HW (HUGE GPU with a LOT of VRAM). It is kind of next level.
Those can keep even NVIDIA GPU Baseboard 4 H200 busy for a while, I guess.
Experimenting with ComfyUI.
A node-based application for generative AI is different aproach and video frame interpolation is only part of it.
It is open source, hence free and with all negatives of it - not working straight away, compatability issues with tons of modules, Python versions etc.
And usually painfully slow.
Found interesting paper called "A Survey on Future Frame Synthesis: Bridging Deterministic and Generative Approaches".