/vt/ - Virtual Youtubers

>>43618589
>Everyone has to use "--no-half-vae" anyways still don't understand why though, so it's probably a wash?
Its quite obvious if you think about it isnt it?
Clearly some of the calculation on the VAE requires FP32 precision to work some of the time, not everytime but often enough to warrant a command like --no-half-vae.
So this is very basic computer science shit but its too complicated to explain here and the wikipedia articles for half and single precision floating point numbers should do it well enough.
I think it eventually winds up into a rounding error that makes the VAE freak out and produce a negative number. Hence why it all turns black. (0,0,0,0) RGBA values.
Example: When dividing by 3 in FP16 the closest value you can divide with is 0.33325195. Not 1/3 = 0.33333... repeating. This is an example of how the inaccuracies in precision could cause mathematical problems in solutions that needs to be more precise and warrant use for FP32 instead. But now you need twice as many bits of space so thats why the fp16 models are twice as small.

Thanks for coming to my ted talk.