DORO EPIC AIRBRUSH - Soft Gradients
Versions
v1 |
DEA_SOFT_GRAD_1- Initial releasev2 |
DEA_SOFT_GRAD_2- more concentrated and powerful than v1
Compatibility
Illustrious XL 🟢 v1 full | 🟢 v2 full
Pony XL 🟡 v1 partially | 🟢 v2 full
Quick Start
🏷️Trigger v1: DEA_SOFT_GRAD_1
🏷️Trigger v2: DEA_SOFT_GRAD_2
⚠️ High-offset LoRA - effective range starts at 2.0+, not the usual 0.5–1.0
🏆 Sweet spot v1:
0.7–1.9- subtle, barely visible2.0–3.0- effect kicks in, full style ⭐3.0+- overpowered, style compression
🏆 Sweet spot v2:
0.7-1.0- subtle atmosphere, clean polishing range1.5-2.5- full effect: dramatic lighting, deep shadows, vortex glow ⭐3.0+- overpowered, latent saturation, loss of fine detail
Description
📸 Dataset: 15 abstract airbrush gradient crops - no objects, pure tonal transitions and color blending. Trained at 768px, 600 steps, AdamW8bit, cosine_with_restarts.
✨ Emergent effects:
Gradient surfaces - clouds, smoke, fog, fire, atmosphere: deep analog painterly quality, as if airbrushed on paper
Smooth surfaces - skin, plastic, metal: surface-blur-like effect, removes micro-noise and texture artifacts, evens gradients
Object edges - sharpened and stylized, more "painted" feel
Background-first - at moderate weights affects mainly background; at high weights touches subjects too
⚠️ Side effect: Smoothing suppresses fine texture (pores, grain, fabric). Not ideal when texture detail is the goal. Workaround: generate smooth, then add noise + slight blur in Photoshop.
💡 Bonus use: Pre-upscale prep - smooths surfaces and reduces artifacts for a cleaner upscale input.
What happened under the hood
This LoRA was trained on abstract gradient crops with no recognizable objects - and that turned out to be the key.
The model couldn't learn any specific object, so it learned pure rendering principles: how to blend tones, transitions, and light. When applied, it rewrites the model's rendering language across all volumetric, gradient-by-nature subjects - clouds, smoke, fire - because those subjects are gradients at their core.
This is a case of spontaneous feature disentanglement: an abstract dataset forced style and content to separate. The result is a universal style modifier, not a content LoRA - similar in principle to implicit style-content separation described in B-LoRA (ECCV 2024).
B-LoRA paper:
https://arxiv.org/abs/2403.14572
❤️ Artificial Inspiration by DORO
Description
Version 2.0 notes: Expanded and curated dataset. High-step low-Dim training (Hard Micro Distillation) produces a far more concentrated signal - noticeably more powerful than v1 at equivalent weights.
Expanded dataset + Hard Micro Distillation: high steps at low Dim force the model to distill only the mathematical core of gradient physics - no noise, no bloat. The result is a significantly more concentrated and powerful signal than v1.
🏷️ Trigger: DEA_SOFT_GRAD_2
⚠️ High-offset LoRA - effective range starts at 2.0+, not the usual 0.5-1.0.
This is intentional: UNet-only training + Alpha/Dim = 0.5 scaling keep the signal
quiet by design - preserving composition integrity, preventing anatomy bleed,
and maximizing cross-model compatibility. Not broken. Just needs more pressure.
🏆 Sweet spot:
- 0.5-1.0 - subtle atmosphere, clean polishing range
- 1.5-2.5 - full effect: dramatic lighting, deep shadows, vortex glow ⭐
- >3.0 - overpowered, latent saturation, loss of fine detail