Krea2 Turbo Attention Optimization Boosts Inference Speed by 2-3x

ostrisai · x · 2026-07-10

ostrisai tested the Krea2 Turbo model in ComfyUI and found a significant inference speedup after applying a KV cache reference token attention mechanism. Under the conditions of 1024x1024 resolution and 9 inference steps, using 1 to 3 reference images reduced the generation time from 13-31 seconds down to 7-10 seconds.

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

Original post →

More from coding & agent

coding & agent channel →