I'm skeptical of this (I think they avoid adding grain to the AV1 stream which they add to the other streams--of course all grain is artificial in modern times), but even if true--like, all grain is noise! It's random noise from the sensor. There's nothing magical about it.
The grain’s got randomness because distribution and size of grains is random, but it’s not noise, it’s the “resolution limit” (if you will) of the picture itself. The whole picture is grain. The film is grain. Displaying that is accurately displaying the picture. Erasing it for compression’s sake is tossing out information, and adding it back later is just an effect to add noise.
I’m ok with that for things where I don’t care that much about how it looks (do I give a shit if I lose just a little detail on Happy Gilmore? Probably not) and agree that faking the grain probably gets you a closer look to the original if you’re gonna erase the grain for better compression, but if I want actual high quality for a film source then faked grain is no good, since if you’re having to fake it you definitely already sacrificed a lot of picture quality (because, again, the grain is the picture, you only get rid of it by discarding information from the picture)
If you’re watching something from the 70s, sure. I would hope synthesized grain isn’t being used in this case.
But for anything modern, the film grain was likely added during post-production. So it really is just random noise, and there’s no reason it can’t be recreated (much more efficiently) on the client-side.
> Are other open weight video models also this small?
Apples models are weights-available not open weights, and yes, WAN 2.1, as well as the 14B models, also has 1.3B models; WAN 2.2, as well as the 14B models, also has a 5B model (the WAN 2.2 VAE used by Starflow-V is specifically the one used with the 5B model.) and because the WAN models are largely actually open weights models (Apache 2.0 licensed) there are lots of downstream open-licensed derivatives.
> Can this run on a single consumer card?
Modern model runtimes like ComfyUI can run models that do not fit in VRAM on a single consumer card by swapping model layers between RAM and VRAM as needed; models bigger than this can run on single consumer cards.
I have a Pi3b in a 3D printer, and compiling the software, but also simply apt upgrade feels like it takes forever. Most day to day operations work just fine though.
At work we have a display with a Pi3 (not B) connected, just showing websites in rotation. Websites even with a simple animation are laggy, startup takes a few minutes.
Both of these usecases don't need more than 1 GB of ram, but I want to speed of a 4 or 5.
reply