Krea 2 Turbo — SDA Diversity LoRA

Demo for F16/krea2-turbo-sda, a rank-32 LoRA for Krea 2 Turbo trained with SDA (Semantic Directional Alignment) to restore the sampling diversity that the Turbo distillation removed — without losing image quality or prompt adherence.

Because the LoRA's effect only appears across seeds, the demo samples one prompt with several seeds twice: plain Krea 2 Turbo on the left, Turbo + SDA LoRA on the right, using the same seeds on both sides.

Gating. The LoRA is trained at a single high-noise sigma node (σ = 0.9567, the second step of the 8-step Turbo schedule) and must only be active for the first 2 of the 8 denoise steps — the demo does this with a callback_on_step_end that calls pipe.disable_lora() after the gate step. The gate is exposed in Advanced settings so you can see the documented quality collapse at gate = 8.

Runs on ZeroGPU with 8-step Turbo sampling, guidance_scale=0.0.

Usage of Krea 2 and its derivatives is subject to the Krea 2 Community License.