Gargoyles & Grotesques [Chroma, Flux, Illu, Klein9B, Z-Image]
![Gargoyles & Grotesques [Chroma, Flux, Illu, Klein9B, Z-Image]](/models/gargoyles/cover.png)
## Description
Architectural guardians in seven materials – stone, bronze, crystal, ice, jade, clockwork, and coral. A style LoRA that transforms any subject into grotesque sculpture perched on gothic architecture.
Gargoyles & Grotesques is trained on 20 diverse gargoyle archetypes spanning creature types (dragon, wolf, eagle, bat, toad, ram, serpent, mantis, monkey, fallen angel), materials (stone, crystal, ice, jade, brass, coral, wood, bone), and architectural settings (Gothic cathedrals, medieval fortresses, castle ruins, crypts, bridges, temples).
The training dataset was synthetically generated using OpenAI’s gpt-image-1 API, with detailed natural-language captions describing each gargoyle’s creature anatomy, material properties, architectural context, and lighting. This gives the LoRA strong concept understanding – it knows that a gargoyle is an architectural guardian creature perched on a structure, not just a generic monster, and it renders appropriate material textures, weathering, and environmental context for each variant.
The LoRA generalizes well beyond its training data. Prompt for materials (clockwork brass, amethyst crystal, rusted cyberpunk metal) and settings (steampunk factories, Chinese temples, arctic glaciers, underwater reefs) not in the training set and it adapts naturally.
Available for five base models: Flux, Illustrious, Klein9b, Zimage, and Chroma. The v2 versions (klein9b, zimage, chroma) bring improved material rendering and architectural detail at lower step counts.
100% SFW – dark fantasy architectural creatures, no human figures required.
## Gallery
## Download
### Flux v1
### Illustrious v1
### Klein9b v2
### Zimage v2
### Chroma v2
## Recommended Parameters
### Strength
- Flux: 0.80 – 0.90 (0.85 recommended)
- Illustrious: 0.75 – 0.85 (0.80 recommended)
- Klein9b: 0.75 – 0.85 (0.80 recommended)
- Zimage: 0.75 – 0.85 (0.80 recommended)
- Chroma: 0.75 – 0.85 (0.80 recommended)
Higher strength pushes harder into the gargoyle aesthetic with more dramatic stone textures and architectural integration. Lower strength (0.5 – 0.6) gives more subtle sculptural influence on existing subjects. The Flux version handles high strength (0.95) gracefully; the other versions are best kept under 0.85.
### Guidance Scale
- Flux: 3.0 – 4.0
- Illustrious: 5.0 – 7.0
- Klein9b: 5.0 – 7.0
- Zimage: 5.0 – 7.0
- Chroma: 3.0 – 4.0
### Trigger Words
No trigger word needed. The LoRA activates on gargoyle/grotesque concepts naturally.
## Creature Types That Work Well
- Classic – dragon, demon, bat-winged humanoid, chimera
- Beasts – wolf, lion, eagle, bull, bear, ram, boar
- Exotic – serpent, toad, mantis, monkey, spider, octopus
- Mythic – fallen angel, phoenix, griffin, stag, wyvern
- Material hybrids – crystal, ice, clockwork, jade, coral, wood, bone
## Prompt Tips
- Name the creature and material: “A [material] [creature] gargoyle” gives the strongest results
- Include architecture: The gargoyle’s perch reinforces the concept – cathedral, fortress, castle, temple, bridge, tower
- Weathering language helps: “weathered,” “moss-covered,” “cracked,” “rain-stained,” “lichen” trigger realistic aging
- Lighting matters: “moonlit,” “torchlight,” “starlight,” “bioluminescent” set the mood
- Cross-domain works: Cyberpunk, steampunk, Chinese temple, underwater – the LoRA adapts
- Illu/Klein9b/Zimage: Use booru-style tags:
gargoyle, dragon, stone, gothic cathedral, perched, bat wings, moonlit, dark fantasy, no_humans, masterpiece, best quality
## Sample Prompts
### Flux – Dragon Cathedral
A massive stone gargoyle with dragon features crouching on the corner of a Gothic cathedral, bat wings folded against its back, moonlit night, weathered limestone with rain stains and patches of dark lichen, dark fantasy
### Flux – Clockwork Steampunk
A clockwork brass gargoyle perched atop a Victorian factory chimney, its body an intricate mechanism of gears and pistons, steam venting from joints, copper patina on exposed surfaces, gaslight glow from below, steampunk dark fantasy
### Flux – Cyberpunk Sentinel
A rusted metal gargoyle perched on a concrete skyscraper ledge in a post-apocalyptic cyberpunk city, neon signs reflected in rain puddles below, exposed rebar and crumbling concrete, night rain, dark fantasy urban decay
### Illustrious – Crystal Cave
gargoyle, amethyst, crystal cave, geode, purple crystals, translucent, prismatic light, cavern, stalactites, no_humans, masterpiece, best quality, dark fantasy, magical
### Illustrious – Ice Aurora
gargoyle, ice, frozen, glacier, frost, icicles, blue ice, arctic, aurora borealis, northern lights, spread wings, no_humans, masterpiece, best quality, dark fantasy, winter
### Chroma/Zimage – Cathedral Guardian
A weathered stone gargoyle dragon perched on a Gothic cathedral flying buttress, wings spread wide, moonlit night, moss growing in the cracks, rain-slicked stone, medieval city visible below, dark fantasy, atmospheric
## Training Details
| Parameter | Flux v1 | Illustrious v1 | Klein9b v2 | Zimage v2 | Chroma v2 |
|---|---|---|---|---|---|
| Base model | FLUX.1-dev | Illustrious XL | Klein9b | Zimage | Chroma |
| Method | LoRA | LoRA | LoRA | LoRA | LoRA |
| Rank | 4 | 4 | 4 | 4 | 4 |
| Alpha | 4 | 1 | 4 | 4 | 4 |
| Steps | 3500 | 2800 | 2000 | 2000 | 2000 |
| Optimizer | Prodigy | AdamW 8-bit | AdamW 8-bit | AdamW 8-bit | AdamW 8-bit |
| Training images | 20 | 20 | 20 | 20 | 20 |
### Checkpoint Selection Note
The Illustrious version uses step 2800 rather than the final step 3500. Testing across checkpoints (2500–3500) showed that material texture richness, environmental complexity, and atmospheric detail peaked at step 2800, with subtle quality regression after that point. The Flux version uses the final checkpoint – its Prodigy optimizer + EMA smoothing prevented overtraining at 3500 steps. The v2 versions (klein9b, zimage, chroma) converge faster at 2000 steps due to improved base model quality.
Note: The klein9b v2 version is currently in testing status and may be updated before release.
