How We See Beauty in Gifts

Scoring Philosophy

Most NFT platforms rank items by rarity alone. We believe visual harmony matters more for collectibles — that's why aesthetic score carries 62% of the final Collector Score.

collector_score = 0.23 × rarity + 0.62 × aesthetic + 0.15 × serial

Rarity is weighted: 65% model + 15% backdrop + 20% symbol. Serial score rewards both low numbers (rank 30%) and beautiful patterns (beauty 70%) — palindromes, meme numbers (69, 420, 1337), repeating digits.

Compatibility Engine

The aesthetic score measures how well a gift's components fit together. Three compatibility pipelines run across every possible attribute pair:

aesthetic = 0.56 × symbol↔model + 0.18 × backdrop↔model + 0.26 × collection↔symbol

Symbol ↔ Model Compatibility (56%)

The primary compatibility — does the symbol make sense with the model? Four AI dimensions are blended:

symbol_model = 0.45 × semantic + 0.28 × visual + 0.10 × strict + 0.17 × cross

• Semantic (45%) — Text embedding similarity via BGE-M3 (1024d). Do they mean similar things?
• Visual (28%) — Image embedding similarity via DINOv2 ViT-L/14 (1024d). Do they look alike?
• Strict (10%) — Multiplicative: semantic × visual. High only when both agree.
• Cross-modal (17%) — SigLIP (1152d): can the AI recognize the model from the symbol's image, and vice versa?

Backdrop ↔ Model Compatibility (18%)

Backdrops are radial gradients — no meaningful image embeddings. We compare them semantically only (BGE-M3 text embeddings). Separately, we compute color separation in dual color space (70% HSV + 30% OKLCH) for the monochrome/contrast achievement system.

Collection ↔ Symbol Compatibility (26%)

Does this symbol fit the collection's theme? Semantic-heavy:

coll_symbol = 0.60 × semantic + 0.20 × visual + 0.15 × strict + 0.05 × cross

All raw cosine similarities are percentile-scaled (p5→0, p95→1) to normalize distributions across entities. Computed in PostgreSQL via pgvector with HNSW indexes (m=16, ef=64).

Color Science

Every model and backdrop image undergoes pixel-level color analysis in multiple color spaces:

Dominant Color Extraction

OKLab k-means clustering (5 clusters, 10K pixel samples, 32 iterations) extracts the dominant color palette. Results stored as OKLCH (Lightness, Chroma, Hue) and HSV.

Monochrome Detection

A 5-stage algorithm in OKLab (perceptually uniform color space) determines whether an image is monochromatic:

1. Gray gate — if p95(chroma) < 0.03, classify as near-grayscale
2. Hue concentration (R) — circular mean of (a, b) vectors; R→1 = single hue dominates, R→0 = scattered
3. Perpendicular spread — p95 deviation from dominant hue axis (threshold: 0.03)
4. Outlier fraction — max 2% pixels may strongly deviate
5. Final verdict — monochromatic if all checks pass

Color Separation

For backdrop ↔ model pairs, we measure how distinguishable the model is against its backdrop. Dual-space blend: 70% HSV (hue 50%, saturation 25%, value 25%) + 30% OKLCH with chroma-gated hue distance (low-chroma colors have unstable hue → gate at C=0.08).

Visual Weight & Rhythm

Rule-based tags from OKLCH values: weight (light/medium/heavy) from 0.5×(1-L) + 0.3×(C/0.15) + 0.2×contrast, and rhythm (calm/dynamic/chaotic) from RMS image contrast thresholds.

Semantic Tagging

Every attribute (model, symbol, backdrop) receives automatic semantic tags derived from LLM-generated keywords + color analysis:

Mood and style use keyword-matching against 150+ trigger words per category, with LLM category fallback (e.g., "animal" → playful, "celestial" → serene + mysterious). Tags power achievements like Mood Curator (80%+ dominant mood) and Chaotic Energy (7+ distinct moods).

3-Layer Search Pipeline

Searching "永遠の炎" (Japanese for "eternal flame") triggers a cascading pipeline:

Entity relevance = max(similarity) across all layers. Gift scoring weights matched attributes: collection +0.03, model +0.04, symbol +0.01, backdrop +0.005. Results cached in Redis (5-min TTL) for instant repeat queries.

AI Model Stack

Total: ~4.8B parameters across 6 models. SigLIP, BGE-M3, and MADLAD-400 run in a dedicated semantic-encoder microservice (7GB RAM, 4 CPU). All vector operations use pgvector HNSW indexes in PostgreSQL.

Achievement System

35+ achievements across 8 categories, many with repeatable instances and diminishing-return scaling (log2, sqrt):

Monochrome/Contrast Lord and God tiers are loss-based — owning a single non-qualifying gift revokes the achievement. This creates meaningful commitment to a collecting strategy.

Infrastructure

Tech Stack

▒███ ▒███ ▒███ ▒███    ▒███ ▒███    ▒███
▒███ ▒███ ▒███ ▒███    ▒███ ▒████  ▒████
▒███ ▒███ ▒███ ▒███    ▒███ ▒█████▒█████
▒███ ▒███ ▒███ ▒████  ▒████ ▒███████████
▒███ ▒███ ▒███ ▒███████████ ▒███████████
▒███ ▒███ ▒███  ▒██████████ ▒███████████
▒███ ▒███ ▒███   ▒█████████ ▒███▒███▒███
▒█████████████         ▒███ ▒███ ▒█ ▒███
▒█████████████ ▒███████████ ▒███    ▒███
▒█████████████ ▒███████████ ▒███    ▒███
▒█████████████ ▒██████████  ▒███    ▒███