hugging-apps/krea2-turbo-4step-lora-demo

Krea 2 Turbo — 4-Step Distillation LoRA

A demo of lvladikov/Krea2-Turbo-Distill-4step-LoRA, a trained (not extracted) LoRA that reduces Krea 2 Turbo's minimum usable step count from 8 to 4 while recovering the high-frequency detail a 4-step schedule loses.

How it works

• —Base model: krea/Krea-2-Turbo — the 8-step distilled Krea 2 pipeline.
• —LoRA: Applied on top of Turbo's transformer at scale 1.0 via load_lora_adapter + set_adapters.
• —Inference: 4 steps, guidance scale 0.0 (Turbo is CFG-free), timestep shift mu=1.15 (Turbo's deployment grid).

The LoRA was trained by progressive distillation with Krea 2 Turbo as its own teacher — the student learns to cover two teacher steps in one, halving the step count while preserving detail.

Settings

Example prompts

The examples below are the five test prompts from the LoRA's model card, chosen to stress different failure modes: skin and hair, frozen motion and fine droplets, legible text and reflections, many small faces and repeated objects, and close-range texture.