> Quantizing the latents from floating point to 8-bit unsigned integers by scaling, clamping and then remapping them results in only very little visible reconst

> Quantizing the latents from floating point to 8-bit unsigned integers by scaling, clamping and then remapping them results in only very little visible reconstruction error. This might actually be interesting/important for the OpenVINO adaptation of SD ... from what I gathered from the OpenVINO documentation, quantizing is actually a big part of optimizing as this allows the usage of Intels new(-ish) NN instruction sets.