サムネ (850 x 400 px).jpeg)
When you train a texture LoRA, you're not just teaching the model to recognize a surface. You're teaching it the process that created that surface. With v1 of SHIFUKU Gold Leaf, we taught the model what gold leaf looks like. With v2, we taught it how gold leaf was applied.
That distinction changes everything.
The Problem v1 Solved (And Where It Stopped)
SHIFUKU Gold Leaf v1 arrived with a clear mission: fix AI's gold leaf problem. You know the problem. You've seen it. 90% of AI-generated gold backgrounds are flat—they're painted gold, or textured gold, but they don't move the way real gold leaf does. They lack the scattered geometry, the atmospheric shimmer, the sense that something physical happened to create that surface.
v1 addressed this by training on 50 carefully curated reference images. The model learned to generate surfaces that looked like gold leaf: the hue, the gloss, the surface variation. It worked. Suddenly your backgrounds didn't look like plastic anymore.
But v1 had a limitation: it taught the model the general appearance of gold leaf, not the specific physics of how different gold leaf techniques produce visually distinct surfaces.
Why Process Matters: The Three Techniques
In nihonga—Japanese traditional painting—gold leaf isn't a monolithic material. The same gold, applied three different ways, creates three completely different textures and light behaviors.
Kiribaku (切箔): The Scatter
Kiribaku means "cut leaf." Small squares of gold leaf—roughly 2–3 cm across—are scattered deliberately across the surface. They overlap. Their edges create a visible grid. When light hits kiribaku, it bounces differently across each square because each sits at a slightly different angle.
When the model generates kiribaku, you see:
Visible edge lines where squares meet
A composition that feels placed, not painted
Overlapping layers that catch light asymmetrically
A texture that reads as discrete objects rather than a continuous surface
Sunago (砂子): The Dust
Sunago is gold dust—microscopically fine particles sprinkled across the surface. Unlike kiribaku, sunago has no edges, no geometry. It's purely atmospheric. When light passes through sunago, it creates a soft, diffuse shimmer, like looking at gold through gauze.
When the model generates sunago, you see:
No distinct shapes, only granular variation
Softer light interaction—less reflection, more diffusion
A sense of depth (the dust sits in a semi-transparent layer)
Atmospheric haze that shifts with viewing angle
Noge (野毛): The Strips
Noge means "wild hair." Thin strips of gold leaf are applied in loose, directional patterns—sometimes following brush strokes, sometimes seemingly random. Unlike the geometric precision of kiribaku, noge has energy. It moves. It catches light along its length.
When the model generates noge, you see:
Linear directionality—the texture flows
Variable width and spacing
High specularity along the strip direction
A sense of applied gesture rather than calculated placement
What Changed in v2
We expanded the training dataset by incorporating the Met Museum Open Access Collection—high-resolution scans of historical East Asian screens and scrolls. This gave the model access to real-world examples of each technique, shot under controlled museum lighting.
The results:
Grain Detail: Individual gold leaf edges are now visible. v1 captured the aggregate appearance; v2 captures the individual components. This matters because it changes how shadows and highlights interact with the surface.
Technique-Specific Light Behavior: v1 applied a general "gold leaf" light response to all prompts. v2 adapts. Kiribaku generates with sharp, angular reflection because that's how scattered squares behave. Sunago generates with softer, more dispersed light because that's how suspended particles work. Noge generates with directional specularity because that's the physics of thin strips.
Oxidation and Aging: Historical gold leaf surfaces show patina—areas where the leaf has darkened or where the ground beneath shows through. v2 captures these imperfections. Your backgrounds don't look freshly applied; they look lived with.
The Haku-Ashi: In traditional application, after the gold leaf is laid, the artist tears away excess leaf with a blade, leaving a raw edge (haku-ashi, 箔足). This edge is visible on close inspection—it's where the backing shows through. v2 generates these micro-details.
How to Use It: Trigger Words Matter
v1 had one trigger word that covered all gold leaf. v2 has four:
gold leaf texture — General, works for mixed techniques
kinpaku — The umbrella term; safe default
kiribaku — Use this when you want visible geometry, overlapping squares, clear edges
sunago — Use this for atmospheric, dusty gold surfaces; great for backgrounds that need depth
noge — Use this for directional, strip-based patterns; good for accent textures
The trigger word is your conversation with the model. Be specific.
Weight Recommendations
0.6–0.8: Subtle gold leaf integration. Good when you want gold to enhance a composition without dominating it.
0.9–1.0: Full effect. The gold leaf dominates. Use this when the gold is the subject.
From Surface to Process
This is the philosophical shift between v1 and v2, and it echoes a broader principle in texture LoRA training: the appearance changes when you understand the process.
This connects to work we've explored in depth elsewhere. In "Style LoRA vs Texture LoRA — They Solve Different Problems", we established that texture is about matière—the physical matter that creates appearance. In "Why AI Doesn't Know the Weight of Paint", we documented how models struggle with the relationship between process and result.
SHIFUKU Gold Leaf v2 is a direct application of the 3-Distance Method—training from multiple angles and distances to capture how a technique reveals itself differently at different scales.
The Subtractive AI approach was essential here too—our captions didn't just describe what we saw in the reference images, but what process created what we saw.
Practical Implications
If you're familiar with v1, v2 is a direct upgrade. The model is more precise, more responsive to specific techniques, and more aware of how light and age affect gold surfaces.
For new users: start with gold leaf texture or kinpaku as your trigger. If your generation feels too uniform, try sunago for depth or kiribaku for geometry. If you're working on a piece that needs directional energy, noge will give you that.
The model works at SDXL 1.0 with weights trained in Kohya_ss. It's commercial-friendly—no credit required, full reuse rights.
The Broader Picture
AI generation of materials is improving because we're getting better at understanding why things look the way they do. This is where texture LoRA development is heading: not toward photorealism, but toward material honesty. A background should feel like it was made by a process, not painted by a diffusion algorithm.
SHIFUKU Gold Leaf v2 is one step down that path.
Model References
SHIFUKU Gold Leaf v2 — Current version
SHIFUKU Gold Leaf v1 — Legacy; foundation for this work
SHIFUKU Kintsugi — Complementary texture LoRA for repaired gold
SHIFUKU Hamon Steel (Beta) — Steel texture using the same training philosophy