DORO EPIC SKIES - Red Dwarf

DORO EPIC SKIES - Red Dwarf

Versions

• v1 | DES_RED_DWARF_1 - Initial release

Compatibility

• Illustrious XL 🟢 full

• Pony XL 🟢 full

Quick Start

🏷️ Trigger: DES_RED_DWARF_1

🏆 Sweet spot:

• 0.7 - 0.9- color correction mode - deepens reds, enriches shadows

• 1.0 - 2.0 - full concept, balanced quality ⭐

• 2.0+ - extreme FX - surreal / abstract

Description

FX-LoRA that transforms any scene into a deep-space environment lit by a dying red star.

📸 Dataset: 7 handpicked images of red dwarf stars, cooling giants, and solar activity - flat dataset, no scene splits. Focused on plasma texture, corona structure, and chromatic contrast between cold space and hot stellar surface.

✨ Emergent effects:

• Emissive Boost - automatically finds light sources, neon, plasma or glowing elements in the scene and applies dying-star physics: deep crimson core with soft shadow falloff

• Material Interaction - reflective surfaces (metal, latex, glass, gemstones) correctly pick up the red spectral shift

• Atmospheric Depth - at higher weights, midtone detail compresses into high-contrast noir, ideal for dramatic and melancholic compositions

• Spectral Targeting - effect anchors to anything in the red spectrum: lava gets denser, neon more aggressive, magical FX (fire, spells) gain cosmic scale

⚠️ Side effect: At weights above 1.6 facial detail and skin tones are absorbed into deep shadow. Workaround: reduce LoRA weight or use negative prompting to restore face detail.

💡 Bonus use: Works as a pure color grading layer at 0.1-0.5 - no sky needed, just a warm crimson tint over any existing composition.

What happened under the hood

Training was UNet-only (no text encoder) using Noise Offset at 768px resolution, 3000 steps, Alpha/Dim 16/16, cosine scheduler with 3 restarts. The absence of text encoder training means the LoRA doesn't rewrite token semantics - it operates purely at the diffusion level, reshaping light distribution and color temperature across the entire latent space.

Because the signal is injected globally rather than tied to a specific object class, it behaves as a scene-wide lighting engine rather than a background replacement. Any element with physical plausibility (reflective surface, emissive material, atmospheric particle) responds to the stellar environment as if it were genuinely lit by a red giant.

The result is a modular FX layer: drop it into any composition and the scene reorients around the dying star - without overwriting the subject. Method based on Noise Offset LoRA training for atmospheric lighting control.

Noise Offset - original technique:

https://www.crosslabs.org/blog/diffusion-with-offset-noise

❤️ Artificial Inspiration by DORO