Not recently loaded
Download
1 variant available
bf16 SafeTensor
X4LSFUMATO_krea2_CodexX.safetensors
BF16, good balance • 223.83 MB
Verified: 18 hours ago
The creator of this LoRA has set this version to Early Access and as such it is only available for people who purchase it. This LoRA will be available for free in . If you want to know more, check out our article here.
Notify me when it's available.Codex X – first release.
Trained for 10 epochs on 72 of my own urban exploration photographs. Strong contrast and deep colors, stays stable even at very high strength up to 2.5.
Show more
In Memory of Leonardo
A style LoRA built entirely from my own urban exploration photography, years of abandoned villas, chapels, workshops and forgotten rooms, condensed into one look.
The name is a quiet tribute. Sfumato, the soft transition from light into shadow, is what defines these places for me: a single beam through a broken window, dust in the air, and everything else slowly dissolving into darkness.
Trigger word
X4LSFUMATO
Style prefixes (from the three training sets)
X4LSFUMATO, dark cinematic color grading, deep shadows, heavy vignetteX4LSFUMATO, warm muted film tones, soft faded colorsX4LSFUMATO, natural warm light, documentary photoRecommended settings
- Base model: Krea 2 Turbo (trained on Krea 2 Raw)
- Strength: 1.0 – 1.4, up to 2.5 for an extreme painterly look
Tips
- Keep prompts short and factual: the place, the objects, where the light comes from, the camera angle. Long style or lighting descriptions override the LoRA.
- One light source works best. A single beam onto the main subject with the rest of the room falling into shadow is exactly what it was trained on.
- For the punchy dark look, add: teal blue shadows and warm orange highlights, high local contrast
- It also adds decay to scenes that are not abandoned at all, but stays subtle when the prompt clearly describes an active place.
Known limitations
- Torn fabric (curtains, tarps) sometimes gets unnaturally round holes.
- On surfaces like signs or license plates the trigger word can show up as text. Avoid those in the prompt or retry with another seed.
Dataset: 72 of my own photographs, captioned by hand and with a local vision model, rank 32.
A MOSS version is planned.

