Krea-2 Turbo 8-Step Distillation LoRA (SVD Extract)

Makes Krea-2 Raw behave like Krea-2 Turbo (8-step, CFG-free) without downloading the 24 GB Turbo checkpoint. This is an experimental extraction, not an official Krea release.

How to use

1. Load the Krea-2-Raw transformer (not Turbo) in ComfyUI or diffusers.

2. Apply one LoRA to the diffusion transformer.

3. Use 8 steps, CFG 0–1, and timestep shift mu 1.15.

Files

• krea2_turbo_distill_r64.safetensors: rank 64, ~0.47 GB. Smallest, rougher fit (use if VRAM or disk is tight)

• krea2_turbo_distill_r128.safetensors: rank 128, ~0.94 GB. Recommended start

• krea2_turbo_distill_r256.safetensors: rank 256, ~1.87 GB. Closest fit

Notes

• This is an approximation. It recovers part of the Raw→Turbo shift, so it won't match native Turbo pixel for pixel.

• Evaluate at the Turbo settings above.

• Keys use the ComfyUI Krea2 convention (diffusion_model.blocks.0.attn.wq.lora_A.weight). Alpha = rank.

• Method: truncated SVD of (W_turbo − W_raw) on each 2D layer, from krea/Krea-2-Raw and krea/Krea-2-Turbo.

License: derived from Krea-2 weights, so it inherits the Krea-2 community license. See Krea's licensing for commercial terms.

Krea 2 technical report: https://www.krea.ai/blog/krea-2-technical-report