arcanea-taste-scorer

name	arcanea-taste-scorer
version	1.0.0
description	AI-powered aesthetic quality scoring for images. 5-dimension evaluation: Canon Alignment, Design Compliance, Emotional Impact, Technical Fit, Uniqueness. Scores 0-100 with automatic tier assignment.
author	frankxai
homepage	https://arcanea.ai
repository	https://github.com/frankxai/arcanea-claw
license	MIT
tags	["quality-scoring","aesthetic-evaluation","ai-vision","curation","image-quality","taste","creative"]
trigger	pipeline
depends_on	["media-process"]
sandbox	true
dependencies	{"python":">=3.11","packages":["google-generativeai","supabase","Pillow"]}
env	{"required":["GEMINI_API_KEY","SUPABASE_URL","SUPABASE_SERVICE_KEY"]}
inputs	[{"name":"batch_size","type":"int","description":"Max images to score per cycle","default":15},{"name":"model","type":"str","description":"Gemini model for scoring","default":"gemini-2.0-flash"},{"name":"dimensions","type":"list","description":"Scoring dimensions and their weights","default":["canon_alignment","design_compliance","emotional_impact","technical_fit","uniqueness"]}]
outputs	[{"name":"scored_count","type":"int","description":"Number of assets scored this cycle"},{"name":"hero_count","type":"int","description":"Number of assets that achieved hero tier (80+)"},{"name":"average_score","type":"float","description":"Average TASTE score across the batch"},{"name":"tier_distribution","type":"dict","description":"Count of assets in each quality tier"}]
tables	["asset_metadata (UPDATE quality_score, quality_tier, quality_dimensions, status)"]

Arcanea TASTE Scorer

AI-powered aesthetic quality scoring that separates hero-tier artwork from the noise. TASTE stands for the five evaluation dimensions that produce a composite 0-100 score.

The Five Dimensions

Every image is evaluated on five axes, each scored 0-20:

Dimension	Weight	What It Measures
T - Technical Fit	0-20	Resolution, clarity, sharpness, web-readiness. Is this image technically sound? No artifacts, proper exposure, clean edges.
A - Aesthetic / Design Compliance	0-20	Composition, color palette, visual hierarchy, professionalism. Does it follow design principles?
S - Story / Canon Alignment	0-20	Does the image fit the world it belongs to? For fantasy projects: does it match the mythology, characters, and aesthetic canon?
T - Transcendence / Emotional Impact	0-20	Does it evoke feeling? Wonder, power, mystery, joy, awe. Does it make you stop scrolling?
E - Exclusivity / Uniqueness	0-20	Is this distinct from generic AI art? Does it have a recognizable signature, style, or perspective?

Composite Score = sum of all five dimensions (0-100).

Quality Tiers

Score Range	Tier	Action
80-100	Hero	Featured placement. Uploaded to CDN. Used for banners, social, marketing.
60-79	Gallery	Standard publication. Uploaded to storage. Browsable in galleries.
40-59	Archive	Thumbnails generated, but not actively published. Available on request.
0-39	Reject	Not published. Flagged for review or deletion.

Installation

Via ClawHub

claw install arcanea-taste-scorer

Via OpenClaw CLI

openclaw skill add arcanea-taste-scorer

As Part of the Full Pipeline

claw install arcanea-media-pipeline
# taste-score is included as step 5 of 8

Usage

In a Pipeline

The scorer runs automatically as step 5 when part of the arcanea-media-pipeline. It picks up images with status='processed' and updates them with scores and tier assignments.

Standalone Scoring

Score a single image programmatically:

from skills.score import TasteScorer

scorer = TasteScorer(model="gemini-2.0-flash")

result = await scorer.score_image("/path/to/image.webp")

print(result)
# {
#   "score": 87,
#   "tier": "hero",
#   "dimensions": {
#     "technical_fit": 18,
#     "design_compliance": 17,
#     "emotional_impact": 19,
#     "canon_alignment": 16,
#     "uniqueness": 17
#   },
#   "reasoning": "Strong composition with dynamic lighting..."
# }

Batch Scoring

from skills.score import TasteScorer

scorer = TasteScorer(model="gemini-2.0-flash", batch_size=15)

results = await scorer.score_batch(
    query={"status": "processed", "quality_score": None},
    limit=15
)

print(f"Scored {results['scored_count']} images")
print(f"Heroes: {results['hero_count']}")
print(f"Average: {results['average_score']:.1f}")

CLI

# Score all unscored processed images
python -m skills.score

# Score a specific file
python -m skills.score --file /path/to/image.webp

# Score with custom thresholds
python -m skills.score --hero-threshold 85 --gallery-threshold 65

Configuration

Basic

taste_score:
  model: "gemini-2.0-flash"
  batch_size: 15

Custom Thresholds

upload:
  hero_threshold: 85    # Raise the bar for hero tier
  gallery_threshold: 65  # Raise the bar for gallery tier

Custom Dimensions

Replace the default TASTE dimensions with your own evaluation criteria:

taste_score:
  dimensions:
    - name: brand_alignment
      weight: 20
      prompt: "Does this image align with our brand guidelines?"
    - name: accessibility
      weight: 20
      prompt: "Is this image accessible (contrast, alt-text potential)?"
    - name: engagement_potential
      weight: 20
      prompt: "Would this stop someone scrolling on social media?"
    - name: technical_quality
      weight: 20
      prompt: "Is this technically sound for web display?"
    - name: originality
      weight: 20
      prompt: "Is this distinct from stock photography and generic AI output?"

Canon Reference

For world-building projects, provide a canon reference file and the scorer will evaluate alignment against your specific universe:

taste_score:
  canon:
    reference: "/path/to/your/CANON.md"
    # The scorer reads this file and uses it as context
    # for the Canon Alignment dimension

How It Works

1. Queries the database for images with status='processed' and no quality score
2. Loads each image and sends it to Gemini Vision with a structured evaluation prompt
3. The prompt asks for JSON output with scores (0-20) for each dimension plus reasoning
4. Parses the response, computes the composite score, assigns a tier
5. Updates the database row with quality_score, quality_tier, quality_dimensions (JSON), and status='scored'

The Gemini prompt is engineered to produce consistent, reproducible scores. It uses few-shot examples calibrated against human-rated reference images.

API Response Format

{
  "asset_id": "uuid-here",
  "score": 87,
  "tier": "hero",
  "dimensions": {
    "technical_fit": 18,
    "design_compliance": 17,
    "emotional_impact": 19,
    "canon_alignment": 16,
    "uniqueness": 17
  },
  "reasoning": "Strong composition with dynamic lighting that evokes a sense of divine power. The color palette aligns with the Fire element canon. Minor deduction on canon alignment due to ambiguous character identity. Highly unique rendering style that distinguishes it from standard diffusion output.",
  "scored_at": "2026-03-17T14:30:00Z"
}

Performance

• Speed: ~2 seconds per image with Gemini Flash
• Cost: Gemini Flash free tier supports ~1,500 images/day (RPM limits)
• Batch size: Default 15 per cycle, configurable up to 50
• Consistency: Scores vary +/- 3 points across runs for the same image

Use Cases

• Content curation: Automatically surface your best work from thousands of generated images
• Quality gates: Block low-quality images from reaching production
• Portfolio building: Let the scorer rank your entire archive
• A/B testing: Compare aesthetic scores across different generation prompts or models
• Social scheduling: Only queue hero-tier images for social media

License

MIT -- use it, fork it, build on it.