Krea2 Edit Training Introduces Isolated Reference Attention

ostrisai · x · 2026-07-10

Developers have added an isolated reference token attention option for Krea2's edit training. This mechanism leverages cached reference KV to boost inference speed by 2-3x without increasing training time. LoRA models trained with this feature will deliver improved overall performance.

Related event: Krea2 Boosts Inference Speed with Isolated Reference Attention(2 posts)→

Original post →

More from Multimodal

Multimodal channel →