Quantisation of the inputs

cloneofsimo / minRF

Minimal implementation of scalable rectified flow transformers, based on SD3's approach

Apache License 2.0

449 stars 37 forks source link

Quantisation of the inputs #9

Open nicolas-dufour opened 4 months ago

nicolas-dufour commented 4 months ago

Hey, wanted to ask, are you using precomputed quantized int8 VAE states? I see you do it for T5 but not sure for VAE states? If so have you seen any quality degradation?

Thanks!

cloneofsimo commented 4 months ago

The reason is rather funny. Only becase VAE were TINY compared to T5, it really didnt make sense to quantize it hahaha