Krea 2 Turbo LoRA: Low Mottle

A drop-in replacement for the Krea 2 turbo LoRA with less blotchy noise in flat areas.

Not an official Krea release and not endorsed by Krea.

The problem

Krea 2 Turbo makes images in 8 steps. It leaves a faint, blotchy noise in smooth areas like skies, walls and gradients. It is easy to miss at first. Zoom in or raise the contrast and it shows. It gets stronger when you stack a style LoRA.

The raw model does not do this. The noise comes from the 8-step distillation.

What this LoRA does

It is the original turbo LoRA, fine-tuned to keep smooth areas smooth. You still get 8 steps and CFG 1. Detail and contrast stay about the same.

Same prompt and seeds. The bottom half boosts the fine noise 6x: grey is clean, speckle is mottle.

A scene with flat sky next to fine line detail. The sky is cleaner and the detail is kept. The middle column boosts contrast. Compositions shift a little from the original for the same seed.

How much less mottle

Measured as fine noise in the flattest parts of the image, over 6 seeds per test. Lower is better.

Noise vs original turbo LoRA

Smooth gradient, about 70% lower (about the same as the raw model)

Detailed scene, about 30% lower

Smooth gradient, style LoRA at 1.5, about 50% lower

Detailed scene, style LoRA at 1.5, about 25% lower

How to use

Use it in place of the turbo LoRA in ComfyUI. Do not load both.

1. Download krea2_turbo_lora_rank_64_bf16_low_mottle.safetensors into ComfyUI/models/loras/.

2. Load it where you would load krea2_turbo_lora_rank_64_bf16, at strength 1.0.

3. Keep your turbo settings: 8 steps, CFG 1.

Limits

- It reduces the mottle. It does not remove it, and busy scenes still have more than the raw model.

- Images change slightly from the original turbo LoRA for the same seed.

- On very smooth gradients, a rare seed can show faint vertical banding.

- It does not touch the horizontal streaks some style LoRAs add.

- Tested in ComfyUI with the Krea 2 raw model plus this LoRA. Not tested in other tools.

How it was made

I started from the Krea turbo LoRA and turned down its last three blocks, where most of the mottle comes from. Then I fine-tuned it for 500 steps with TDM distillation (the same kind of training used to make turbo models), plus an extra loss that compares flat-area noise with the raw model's. Training used text prompts only, no images. The code is in https://github.com/rockerBOO/boo-musubi-tuner.

License and attribution

Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.

This LoRA is a modified version of the Krea 2 Turbo LoRA. It is a Derivative under that agreement, and use is subject to its terms, including the revenue threshold, content filtering requirements and Acceptable Use Policy.