the thing is with these CRT emulators is that they always look so over-artifacted. I looked at CRTs for a big part of my life and they didn't actually look like this. Maybe if you took a photo of them, but not in full motion.
With real CRTs you don't notice the scanlines so much due to interlacing at double the frame rate. And they are kinda fuzzy so they're not so clear anyway.
I wonder if a really fast, say 240Hz display could simulate this better somehow.
I think another way to do this would be to develop a proper physical simulation that foregoes any notion of realtime. It can take 24 hours to bake 5 minutes of CRT output.
We use the simulator to generate test data that can be used to train an emulator. The emulator can be deployed as a model that provides the CRT effect in real time. We already have things like neural upscaling and denoising. This could be a similar idea.
This must be similar to what seemingly every documentary on Netflix uses in all their editing when showing any sort of footage taken more than 5 years ago.
Some show will be showing a clip of an interview recorded in 2020, and for some reason it’s edited to appear to be a poorly-stored VHS when the original was almost certainly a 1080p or 4K digital recording.
A lot of these are principled simulations, which is fine, but what I wouldn't give for a data-driven reproduction. At the end of the day, it's a learnable function that we have more than enough deep learning machinery to arbitrarily approximate.
the thing is with these CRT emulators is that they always look so over-artifacted. I looked at CRTs for a big part of my life and they didn't actually look like this. Maybe if you took a photo of them, but not in full motion.
With real CRTs you don't notice the scanlines so much due to interlacing at double the frame rate. And they are kinda fuzzy so they're not so clear anyway.
I wonder if a really fast, say 240Hz display could simulate this better somehow.
I think another way to do this would be to develop a proper physical simulation that foregoes any notion of realtime. It can take 24 hours to bake 5 minutes of CRT output.
We use the simulator to generate test data that can be used to train an emulator. The emulator can be deployed as a model that provides the CRT effect in real time. We already have things like neural upscaling and denoising. This could be a similar idea.
This must be similar to what seemingly every documentary on Netflix uses in all their editing when showing any sort of footage taken more than 5 years ago.
Some show will be showing a clip of an interview recorded in 2020, and for some reason it’s edited to appear to be a poorly-stored VHS when the original was almost certainly a 1080p or 4K digital recording.
Does this mean we’ve finally caught up with the Japanese?
FFmpeg filter pipelines for batch simulating vintage CRT TV + VHS/DVD
A lot of these are principled simulations, which is fine, but what I wouldn't give for a data-driven reproduction. At the end of the day, it's a learnable function that we have more than enough deep learning machinery to arbitrarily approximate.