This is a repository for all experimental work.

**Hybrid Pruned fl2va_ref2va**
* Graft of fl2va as base with ref2va as donor from blocks(25-49)
  * bf16
  * int8_convrot

**VAE**
* Created a w4a8 vae to bring down the size of the VAE, i haven't done A/B testing on the decode speed, for me it was just about saving memory. Kept it true to how the to_qv for vae works on ComfyUI