lxe/holosomnia-krea2-style-lora

HoloSomnia Krea 2 Style LoRA

A lightweight style LoRA trained to carry the creator's HoloSomnia visual language into Krea 2: luminous multicolored clouds, strong light rays, saturated reflected color, sculpted cel-shaded forms, and deep atmospheric landscapes.

This is the exact Krea 2 checkpoint used to create the reference images for the Skythread R2V experiment.

File

• holosomnia-krea2-style-v3-full.safetensors
• SHA-256: 2c725cee0fccc5b9d5432a2f6f5129d119aa057c45383a140a004646d63019c3
• Training step: 1250
• Training base: krea/Krea-2-Raw
• Intended inference model: krea/Krea-2-Turbo
• Trigger word: none

Suggested strength

Start around 0.4. That is the strength used for Skythread and gave the style a visible influence without overpowering composition or subject details.

A practical range is roughly 0.3 to 0.6:

• Lower values keep more of Krea's base rendering.
• Higher values push the HoloSomnia color, clouds, light, and surface treatment harder, but may reduce structural fidelity.

ComfyUI

Place the checkpoint under your LoRA model directory, for example:

ComfyUI/models/loras/krea2/holosomnia-krea2-style-v3-full.safetensors

Load it after the Krea 2 Raw-to-Turbo LoRA in a compatible Krea 2 workflow. The complete workflow, prompts, seeds, and reference PNGs used for Skythread are in the Skythread repository.

Rights and limitations

The training material is based on original HoloSomnia artwork and is not included here. This LoRA is not a standalone image model and requires a compatible Krea 2 setup. Users are responsible for following the Krea 2 license and the terms of the tools used for inference.