DORO EPIC COVERS - Sepultura 93 - Chaos AD
Versions
v1 |
DEC_CHAOSAD_1- Initial release. Full biomechanical world reskin distilled from the original production master file - the highest resolution scan made from the painting for album release.
Compatibility
Illustrious XL 🟢 full
Pony XL 🟡 partially
Quick Start
🏷️ Trigger: DEC_CHAOSAD_1
⚠️ High-offset LoRA - effective range starts at 1.5+, not the usual 0.5-1.0. This is intentional: UNet-only training preserves composition integrity, prevents anatomy bleed, and maximizes cross-model compatibility. Not broken. Just needs more pressure.
🏆 Sweet spot:
1.5- passive threshold, minimal effect without booster tags2.0-2.5- GOLDEN ZONE, full world stylization ⭐3.0- extreme surreal overload, structure holds but reality bends
Booster tags (for weight < 1.5):
bandages, cables, wires, machinery, pipes, brick wall, many faces
Description
A style LoRA that acts as a visual virus - transforming environments and objects into dense biomechanical and technogenic aesthetics.
📸 Dataset: 16 images, one master painting - Cacophony (1993) by Michael R. Whelan, used as the cover for Sepultura's Chaos A.D. Sourced from the original production master file - the highest resolution scan made from the painting for album release. Mixed media: photocopies as collage base, painted over with airbrush. Dataset structured into 4 sub-scenes: shroud textures, wall souls, techno-organic cables, and compositional anchors.
✨ Emergent effects:
Material replacement - stones, rubble and plain surfaces become dense tangles of bronze/golden pipes, tubes and cables
Background transformation - generates complex technogenic panels and organic structures with relief faces embedded in walls
Airbrush texture overlay - imposes soft gradient quality and near-photographic material depth
Color palette lock - deep indigo and violet shadows, warm ochre and rust metallic highlights, sharp temperature contrast
⚠️ Side effect: Selective activation - works only on elements structurally similar to the training (pipes, wires, mechanisms, hard backgrounds). Faces and human anatomy are largely protected and stay clean. Only environment and equipment are transformed. Workaround: none needed - this is by design.
💡 Bonus use: Stack with airbrush-style LoRAs for maximum material depth and painterly quality.
What happened under the hood
This LoRA was trained on a single painting distilled into 16 carefully composed images across 4 sub-scenes. The dataset isolation - one visual source, one atmosphere, one color logic - is what gives it such a decisive stylistic grip. There's no averaging, no drift, no compromise.
UNet-only training (no Text Encoder) means the LoRA operates purely on the visual signal, not the language side. This is why it needs higher weights to activate - the text encoder isn't pushing it. What you gain is anatomy protection: the model has no semantic handle on "person", so it doesn't touch faces or skin. The biomechanical transformation happens around the character, not on them.
Alpha/Dim ratio is 1.0 (dim=32, alpha=32), which gives full signal without scaling compression. The high effective weight range is a direct consequence of UNet-only architecture plus the strong visual coherence of the training source - not a flaw in calibration.
Chaos A.D. - Sepultura (Wikipedia):
https://en.wikipedia.org/wiki/Chaos_A.D.
Michael Whelan (official site):
Kohya LoRA training - UNet-only:
https://github.com/kohya-ss/sd-scripts
❤️ Artificial Inspiration by DORO