Krea2 Boosts Inference Speed with Isolated Reference Attention

Krea2 has introduced an isolated reference token attention mechanism for edit training. By utilizing cached reference KV, this feature significantly boosts inference speed by 2-3x without increasing training time.

2026-07-10 ~ 2026-07-10 · 2 related posts

• Krea2 Edit Training Introduces Isolated Reference Attention — ostrisai · 2026-07-10
• Krea2 Turbo Attention Optimization Boosts Inference Speed by 2-3x — ostrisai · 2026-07-10