Why Most Style LoRAs Fail at Texture
Most texture or material LoRAs are trained on one of two things: AI-generated images that approximate a style, or flat photographs downloaded from the internet. Both have a fundamental limitation — they capture appearance, not physical structure.
Real gold leaf doesn't just "look gold." It has directional grain from the hammering process, micro-cracks from aging, and a specific way light diffracts through its approximately 0.1–0.3 micron thickness. A style LoRA trained on golden images will give you gold color. A texture LoRA trained on the real material gives you gold leaf behavior.
The same applies to any physical material: lacquer, shell, stone, paper, fabric.
The 3-Distance Photography Method
The key insight: a material's texture exists at multiple scales simultaneously. A single photograph captures only one scale. Training on one scale teaches the model one thing. Training on three scales teaches it how the material actually works.
Distance 1: Macro (~20 images)
Extreme close-up photography of the material surface. At this distance, you see:
Grain structure and directionality
Surface micro-irregularities
How light interacts at the material level
Aging patterns, patina, wear
Distance 2: Mid-Range (~20 images)
The material applied to a surface or small object. At this distance, you see:
How texture tiles and repeats
Edge behavior where material meets other surfaces
Medium-scale pattern variation
How the material catches light across a larger area
Distance 3: Full Object (~10 images)
The complete artifact or surface. At this distance, you see:
Overall material impression
How texture reads at viewing distance
The material's role in composition
Interaction with surroundings
Captioning Strategy
Each distance receives different caption tags because different information is visible at each scale. At macro, you describe grain and surface. At mid-range, you describe application and behavior. At full distance, you describe impression and context.
This isn't a secret — it's the logical consequence of treating materials as multi-scale phenomena rather than single-image styles.
Results
Gold Leaf (金箔)
The gold leaf LoRA captures the hammered grain structure and the way real kinpaku diffracts light differently from metallic paint or digital gold effects.
Mother-of-Pearl (螺鈿)
Raden's iridescence comes from thin-film interference in shell layers. The LoRA reproduces the angle-dependent color shifting that makes real mother-of-pearl distinctive.
Kintsugi (金継ぎ)
The repair lines in kintsugi follow the original break pattern — they're never perfectly straight or uniform. The LoRA learns this organic irregularity from actual kintsugi pieces.
What's Next: 25+ Traditional Japanese Materials
The full project covers materials used in Japanese traditional crafts:
Metal: Gold leaf, silver leaf, copper leaf, Japanese sword steel
Shell & Stone: Mother-of-pearl, mineral pigments (iwaenogu)
Lacquer: Urushi (black, red, gold), maki-e (gold-powder lacquer)
Fiber: Washi paper, silk, indigo-dyed fabric
Ceramic: Raku glaze, celadon, Kintsugi repair
Wood: Paulownia, cypress, bamboo
...and more
Each LoRA follows the same 3-distance methodology, creating a coherent library of physically-accurate material textures.
---
Case Studies
Gold Leaf v2 — 3 Distances × 3 Techniques
Each technique (kiribaku, sunago, noge) reveals different information at each distance. At macro: individual particle/edge detail. At mid-range: the overall pattern. At full view: the technique dissolves into atmospheric shimmer. → Model
Hamon Steel — 4 Material Surfaces × 3 Distances
A single sword contains hamon, jigane, saya, and tsuba — each surface tells a different story at each scale. → Model
Explore the full SHIFUKU library:
SHIFUKU Gold Leaf v1 — Physical Texture LoRA for SDXL
SHIFUKU Gold Leaf v2 — Kiribaku / Sunago / Noge — 3 gold leaf techniques
SHIFUKU Kintsugi — Physical Texture LoRA for SDXL
SHIFUKU Hamon Steel (Beta) — Japanese sword textures
Free LoRAs available on Civitai and HuggingFace.
TextureLoRALab — Art Studies (public art university, Japan) + MA Museum Studies (UK, Merit). Training texture LoRAs from real materials.
X: @TextureLoRALab