Lifts night scenes so the surroundings don't fall into black, and gives skin a matte finish with visible pores. It does not darken the face — it raises the scene around it.
How to use it
Put mttnght at the start of the prompt, describe the subject and name the single light source. That's it. You don't need a paragraph about exposure — the LoRA does that work, and in testing a long "nothing falls into pure black" instruction actually made results worse than a plain prompt. Recommended strength 1.0, which is what all testing was done at.
Example: mttnght, a man in his thirties with brown skin and close-cropped hair, dressed, in a stairwell at night, lit only by the lamp on the half landing above
Training
FLUX.2 Klein 9B, 2000 steps, LR 1e-4, network dim 32, 489 images. This is epoch 10.
Measured
48 images, 3 hard night scenes x 4 seeds, compared against the same prompts without the LoRA:
Face-to-scene brightness ratio: 1.59 with, 1.73 without (lower is better — the face sits closer to the exposure of everything around it)
Worst case: 1.87 with, 2.00 without
Scene brightness: 86.8 with, 72.9 without
Epochs 6, 8 and 10 were within noise of each other; epoch 10 won on worst case and on paired same-seed comparison, 9 of 12.
Description
Lifts night scenes so the surroundings don't fall into black, and gives skin a matte finish with visible pores. It does not darken the face — it raises the scene around it.
Trained on FLUX.2 Klein 9B, 2000 steps, LR 1e-4, dim 32, 489 images. This is epoch 10.
Measured on 48 images (3 hard night scenes x 4 seeds):
- face/scene brightness ratio: 1.59 with LoRA vs 1.73 without
- worst case: 1.87 vs 2.00
- scene brightness: 86.8 vs 72.9
Recommended strength 1.0 (that's what all testing was done at).