Experimental structurally pruned releases of Kroma v0.3 Turbo by lodestones. I was just curious as to how the removal of pieces of architecture would affect the model, so just experimental models with a guarantee of quality loss. Once again, my goal is to reduce peak VRAM/RAM usage, especially for lower end hardware.

Pruned v1

Structural changes:

• Removed 1 full transformer block: original block 2

• Main block count reduced from 28 → 27

• Pruned 1,536 MLP channels from every remaining block

• MLP width reduced from 16,384 → 14,848

The pruning is physical: the removed block and MLP channels are no longer stored in the checkpoint rather than simply being bypassed.