Krea 2 — prediction previews

Every step, the model predicts the image it is heading for. This shows that, rather than the noisy latents it currently holds — so there is a whole picture from step one, which then sharpens. Built on the OzzyGT/krea2_preview_blocks modular blocks.

Preview decoder
The tiny VAE costs ~21 ms a step at 1024². The RGB projection is under 1 ms, and blocky.
1 5
Model
Picks the blockset: distilled runs few-step without CFG. Presets set this for you.

🟢 Ready (ZeroGPU)

idle