Trigger word: ve_style — it is silently ignored if you forget it.
A scenery-first KyoAni style LoRA trained on hand-curated Blu-ray frames from Violet Evergarden (Kyoto Animation).
Recommended weights
- 0.9 — scenery / establishing shots / still life / backgrounds. This is what it is good at.
- 0.6 or lower — when stacking on top of a character LoRA. Above this it starts pushing black hair toward red-brown.
Honest limitations (tested, not guessed)
I ran 54 test images across five rounds (direct stacking, weight ladder, two-pass, ControlNet parameters, denoise ladder). Findings:
- It will NOT restyle a character that is driven by a character LoRA. With a character LoRA at 0.85 in the mix,
ve_stylemostly acts on the color layer (warm tint, hair shifting red-brown) rather than on linework and shading. No amount of weight, two-pass, ControlNet loosening, or denoise increase changed this. - It does render KyoAni-looking characters fine when no character LoRA is competing with it.
- Always state a lighting direction (
high noon,overcast,night, rain). Leave it out and you will get golden-sunset backlight every time — the training set is sunset-heavy. Stating it explicitly overrides this completely.
Training
350 curated frames · 7000 steps · 10 epochs · 1024px · dim/alpha 32/32 · UNet LR 5e-4 · TE LR 5e-5 · AdaFactor · cosine.
Violet herself was capped at 14.3% of the set to keep the style from collapsing onto one face.
This is epoch 7. Convergence was reached around epoch 6 — epochs 8-10 were near pixel-identical to 7, so 4000 steps would have been enough.