(Krea 2 version added: The main description was originally written for the Flux Klein version. Please check 'About this version' for Krea 2-specific notes and comparisons.)
Flux Klein 9B is an excellent base model, but I wanted to push it closer to the kind of imagery I personally enjoy. I made this LoRA mainly for my own taste, so it may not be for everyone. It is designed to enhance style in areas like lighting, shadows, blur, flare, and overall atmosphere. It works particularly well on moody, cinematic, poster-like images and tends to push scenes toward a more dramatic tone. It is not intended for ultra-realistic everyday photography.
Preview comparisons
At Strength 1, the LoRA usually keeps the core idea of the prompt intact while adding its own atmosphere. This level gives the most balanced results and is the safest option for general use.
At Strength 2, the effect becomes much stronger and more stylized. It is better suited for users who want more aggressive cinematic changes, especially in bokeh, haze, glow, and lighting effects
* Test Model: flux-2-klein-9b-kv-fp8.
Description
V0id for Krea 2 is a visual style and atmosphere LoRA focused on lighting, shadows, blur, glow, haze, flare, contrast, camera angles, and overall cinematic mood.
I made it mainly around the kind of imagery I personally enjoy, so it may not be for everyone. It works best with moody, dramatic, poster-like visuals and tends to push scenes toward a more stylized, atmospheric direction. It is not intended for ultra-realistic everyday photography.
Since I’m adding this as a new version, I can’t include preview images directly in the description. You can find the comparison images here: https://imgur.com/a/jjz6SkG
The gallery includes LoRA off, Strength 1, Strength 1.5, and Strength 2 results. I also included some weaker/failed outputs so you can see where the LoRA does not work well. Please check them before downloading to see if it fits your taste and workflow.
Note:
Krea 2 reacts differently from Flux. In my tests, it can absorb multiple visual signals inside a single LoRA very strongly, including both style-related and form-related behavior. This is a big advantage, but when too many different style signals are trained into a single LoRA, balancing them can become more difficult. The model may start blending style and form more strongly than intended.
Compared to the Flux version, I had to reduce some atmospheric effects to keep the V0id feeling usable across different prompts instead of making it work only on images close to the dataset.
I am still testing alternative versions and have already started working on V0id v2 for Krea 2 with revised data, captions, and training settings.