DORO EPIC SPACE - Energy

Versions

• v1 | DESP_ENERGY_1 - Initial release.

Compatibility

• Illustrious XL 🟢 full

• Pony XL 🟢 full

Quick Start

🏷️ Trigger: DESP_ENERGY_1

🏆 Sweet spot:

• 0.7 - 1.0 - soft stylization, subtle glow and contrast lift

• 1.5 - 2.0 - full Style-ID active: saturated light, deep shadows, atmospheric density ⭐

• 2.1 - 2.5 - max pressure, poster-grade drama, extreme black crush

• 2.5+ - overdriven, color blowout and artifacts

Description

Style LoRA for analog airbrush space art - transfers the physics of cosmic energy onto any subject.

📸 Dataset: 10 hand-picked analog airbrush illustrations (filtered from a pool of 200) - cosmic jets, electrical discharges, infernal vortexes, storm skies. Maximum signal amplitude, zero filler.

✨ Emergent effects:

• Dramatic contrast - deep blacks carved out alongside glowing cyan/orange cores, cinematic depth out of the box

• Energy glow - any light source (fire, magic, stars, electricity) gains physically coherent scatter and bloom

• Structural flow - hair, fabric and FX elements align to dynamic compositional lines: spirals, jets, directed rays

• Color science - saturation boost in complementary pairs (cyan/orange, blue/gold) without muddy midtones

• Character-safe - works on environment and lighting only, does not reshape faces or anatomy. Drop it on any character scene without breaking the subject.

⚠️ Side effect: weights above 2.1 can produce excessive shadow crush in bright daytime scenes. Workaround: reduce weight to 1.5-2.0 or add bright lighting to prompt.

💡 Bonus use: stacks cleanly with other DORO EPIC modules - combine with DORO EPIC AIRBRUSH - Soft Gradients 1 for full analog texture layering.

What happened under the hood

Training used dim=16 rank with alpha=16 (ratio 1.0) - full signal amplitude, no alpha suppression. With only 10 images to learn from, the low rank itself acted as an information bottleneck, forcing the network to discard noise and encode only the highest-amplitude signals: light distribution, contrast structure, and color energy. The result is a style that transfers broadly without leaking subject-specific content.

noise_offset=0.05 was enabled during training - this directly informs the model's ability to produce true deep blacks and extreme shadow crush. The dramatic contrast in outputs is not a prompt artifact; it is baked into the weight distribution.

min_snr_gamma=5 loss weighting was applied to stabilize gradient updates across difficult timesteps, preventing the model from over-fitting to mid-tone noise at the expense of highlight and shadow extremes.

UNet-only mode kept the text encoder fully intact. The LoRA carries no token bias - it doesn't fight the prompt, doesn't inject planets or stars unprompted, and plays well with any subject matter. It acts as an intelligent LUT: pure lighting and atmosphere, zero semantic contamination.

Training followed a high-intensity protocol targeting an exposure energy of ~0.085 (steps × learning rate: 5650 × 1.5e-5). Checkpoint selection was synchronized with cosine_with_restarts scheduler across 3 restart cycles - catching peak weight plasticity before latent space saturation. This is why the sweet spot sits higher than typical LoRAs: the signal is quiet by design and needs weight pressure to fully unlock.

Cosine with restarts scheduler reference

https://huggingface.co/docs/diffusers/api/schedulers/cosine_with_restarts

❤️ Artificial Inspiration by DORO