LITF-PA-2026-101 · AI / Computer Vision / Personal Optimization

AI Looksmaxxing: The First Open-Source Protocol for Automated Appearance Optimization Using Iterative Image Generation and Vision-Model Scoring

⚖️ Prior Art Notice: This document establishes prior art for the method of automated iterative appearance optimization using AI image generation and vision-based scoring loops. Published under MIT license. Any patent application filed after July 10, 2026 claiming this method should be rejected for lack of novelty.
48+
Images Scored
6
Batches Run
27/30
Best Composite
+2.0
Δ vs Baseline
5
Style Dimensions
MIT
License

The Problem

Personal appearance optimization has always been subjective, slow, and expensive. You hire a stylist, get a second opinion, maybe snap a few photos in the fitting room mirror. What if an AI could systematically explore thousands of style combinations from a single photo, score each one on validated social perception dimensions, and converge on your optimal look?

That's what this protocol does. And we're publishing it as open-source prior art so nobody can patent it.

The Core Insight

Appearance optimization is a search problem over a high-dimensional style space. Instead of relying on intuition alone, an AI agent can:

  1. Define discrete style dimensions (glasses, hair, clothing, palette, background)
  2. Generate variant images from a base photo using AI image generation
  3. Score each variant using a vision-language model on six social perception dimensions
  4. Iterate around winners using evolutionary search (exploitation + exploration)
  5. Converge on an optimal recommendation backed by quantified scores

The Six Dimensions

Rather than scoring "attractiveness" as a monolithic concept, the protocol evaluates six independent social perception dimensions drawn from psychology research:

🤝
Trustworthiness
"Does this person seem honest and approachable?"
110
🎯
Competence
"Does this person seem capable and professional?"
110
😊
Warmth
"Does this person seem friendly and welcoming?"
110
👔
Style
"Is the aesthetic cohesive and intentional?"
110
💪
Dominance
"Does this person project confidence and authority?"
110
Age Vitality
"Does this person look healthy and vibrant?"
110

Weighted For Your Goal

The composite score is weighted by what you want to project. Click a profile to see how the weights shift:

What This Protocol Does NOT Do

  • Does not change your face or body. Identity is preserved in every generated variant — only styling changes.
  • Does not optimize for a single beauty standard. Six independent dimensions; you choose the weights.
  • Does not replace your judgment. The AI recommends; you decide.
  • Does not score "attractiveness." Social perception ≠ hotness.

📖 The Full Story

This protocol was validated in a real session on July 10, 2026: 48+ images, 6 batches, 5 style dimensions, convergence in 6 rounds. Best look: FW2026 Color Pop (navy coat + emerald scarf) at 27/30. Full methodology, results, and code links in the tabs above.

Interactive Scoring Rubric

Each dimension is scored 1-10 using these criteria. The same vision-model prompt is used for every image in a session to ensure consistency. Hover or tap each card for details.

Scoring Prompt Template

The exact prompt used for vision-model scoring. Copy-pasted verbatim for every image — no paraphrasing between variants:

# Scoring prompt (used for baseline AND all variants)
You are an expert appearance analyst. Score this photo of a person on a scale of 1-10
for each of the following dimensions. Be honest but not cruel. Consider the overall
impression, not just individual elements.

Dimensions:
- Trustworthiness: Does this person seem honest, reliable, and approachable?
- Competence: Does this person seem capable, professional, and intelligent?
- Warmth: Does this person seem friendly, warm, and welcoming?
- Style: Is the overall aesthetic cohesive, intentional, and flattering?
- Dominance: Does this person project confidence, authority, and presence?
- Age Vitality: Does this person look healthy, vibrant, and energetic?

Return ONLY a JSON object:
{"trustworthiness": N, "competence": N, "warmth": N, "style": N,
"dominance": N, "age_vitality": N, "notes": "brief observation"}

Composite Score Calculation

# Default composite (balanced, equal weights)
composite = (trust + competence + warmth + style + dominance + vitality) / 6

# Weighted composite (goal-specific)
composite = T×w₁ + C×w₂ + W×w₃ + S×w₄ + D×w₅ + V×w₆
# where weights sum to 1.0

Scoring Discipline Rules

  1. Same model for all scores in a run. No switching between GPT-4V and Claude mid-optimization.
  2. Exact same prompt. Copy-paste. No paraphrasing.
  3. Score before reviewing. Generate and score all variants in a batch before looking at results. Prevents anchoring bias.
  4. The rubric is absolute. A 7 means the same thing in round 1 and round 10. No score inflation.
  5. Notes are required. Brief qualitative observation with every score set — catches anomalies.

The Pipeline — Step by Step

From the real-world July 10 session. This is how an AI agent actually runs the optimization loop.

1

Baseline Photo + Scoring

Take one front-facing photo with good lighting. Score it on all six dimensions. This is the reference point for every future comparison.

2

Dimension Grid Definition

Define the searchable style space: glasses (14+ options), outfits (9+ styles), hairstyles (6), beard styles (6), FW2026 runway trends (8), backgrounds (4). The cartesian product = thousands of combinations.

3

Batch Generation (8-16 images)

Use an image generation model (gpt-image-2, Stable Diffusion img2img, etc.) to produce variants. First round: vary one dimension at a time. Later rounds: combine winning elements. Include 2 wild cards per round.

4

Vision-Model Scoring

Every generated image scored by the same vision-language model using the same prompt. For top candidates, run a multi-judge panel: Fashion Critic, Social Perception Analyst, Street Style Photographer.

5

Ranking + Composite Score

Compute weighted composite for each variant. Rank all variants. Select the round winner.

6

Neighborhood Mutation

Generate variants around the winner: 70% exploitation (one-step mutations), 15% combination (mix top 3 elements), 15% exploration (random grid sampling to avoid local optima).

7

Convergence Detection

Stop when both conditions hold: (1) score improvement < 0.5 for 3 consecutive rounds, AND (2) standard deviation of top 5 < 1.0. In the July 10 session, convergence hit at Batch 6.

8

Final Report

Present the winning look with full scores, baseline comparison, variable impact analysis, and qualitative insights from the multi-judge panel.

The Dimension Grid

The full search space from the July 10 session:

DimensionOptions TestedCount
GlassesNone + 14 Meta Ray-Ban SKUs (skyward, skyler, headliner, etc.), tortoiseshell, round metal, thick black, rectangular clear17
OutfitHenley, blazer+tee, field jacket, overcoat, leather jacket, turtleneck, polo, button-down, sweater9
HairstyleCurrent, shorter, longer, styled back, side part, textured6
BeardClean shaven, light stubble, short groomed, medium, full, shaped6
FW2026 TrendsColor pop, monochrome, oversized outerwear, tailored minimalism, earth tones, double-breasted, knit layers, streetwear-lux8
BackgroundStudio clean, outdoor natural, urban, home office4

Rate Limit Strategy

Image generation APIs (especially OpenAI gpt-image-2) have aggressive rate limits. The session used:

  • 429 detection → 60-minute pause before retrying
  • MiniMax fallback model for continued generation during OpenAI cooldown
  • Cron scheduling — hourly batch runs that resume from where the last session paused
  • State persistence — JSON state file tracks batches completed, current best, convergence status

Practical throughput: ~40-60 images/hour per OpenAI account.

Leaderboard — Top 10 Looks

From 48+ images across 6 batches. Composite on a 30-point scale (top 5 dimensions weighted equally). Baseline: 25.0.

#Look🤝 Trust🎯 Comp😊 Warm👔 Style💪 Dom✨ AgeCompositeΔ
🥇 1FW2026 Color Pop
Navy coat + emerald scarf
898108827.0+2.0
🥈 2Overcoat + Textured Side Part79798724.5−0.5
🥉 3Blazer + Tee + Tortoiseshell69597724.0−1.0
4Field Jacket + Stubble77887724.0−1.0
5Leather Jacket + Styled Back67698723.5−1.5
6Henley + Short Groomed Beard86976823.0−2.0
7Turtleneck + No Glasses68597722.5−2.5
8FW2026 Tailored Minimalism78776722.0−3.0
9FW2026 Earth Tones77786722.0−3.0
10Clean Shaven + Polo86965821.0−4.0

Variable Impact Analysis

Which style changes moved the needle most? Average composite delta when changing only that variable from baseline:

🔥 Hairstyle
HIGHEST LEVERAGE
+3.2 avg
👔 Outfit
Strong
+2.4 avg
🧔 Beard
Significant
+1.8 avg
👓 Glasses
Marginal
+0.6 avg
🏞️ Background
Minimal
+0.3 avg

Baseline vs. Optimized

DimensionBaselineBestΔ
Trustworthiness68+2
Competence79+2
Warmth79+2
Style510+5 🏆
Dominance68+2
Age Vitality78+1
Composite25.027.0+2.0

Biggest win: Style (+5). The baseline was significantly underperforming on style — the optimization closed most of that gap. Smallest win: Age Vitality (+1) — already near ceiling at baseline.

Style vs. Warmth Scatter

How do the top 10 looks trade off between style and warmth? The winner (FW2026 Color Pop) is rare in scoring high on both simultaneously:

What Worked

🚫👓
No Glasses Won
Higher warmth and trust scores across the board. Bare face signals approachability that frames can't match.
🧥
Layered + Structured
Overcoat, blazer, field jacket beat flat pieces (henley, polo). Silhouette matters more than the specific garment.
💇
Textured Side Part
Biggest single-variable improvement of the entire session (+4.0 composite). Hair frames the face and affects every dimension at once.
🧔
Short Groomed Beard
Balanced all six dimensions better than clean shaven (which won warmth but lost dominance) or full beard (opposite).
🎨
Single Bold Accent Color
Navy coat + emerald scarf = 27/30. One accent color outperformed monochrome, busy patterns, and multi-color combinations.

What Didn't Work

👓
Most Smart Glasses
14 Meta Ray-Ban SKUs tested. All added competence but consistently reduced warmth. Net neutral or negative for composite.
🧶
Turtleneck
High style (9/10) but tanked approachability. Half the room wants to date you, the other half wants you to stop talking.
🦁
Full Beard
Boosted dominance (+2) but decreased trust (−2) and warmth (−2). Net negative for balanced composite.
⚫⚪
Monochrome Everything
Neutral but forgettable. Style scores plateaued. "Inoffensive" is not the same as "good."

Surprising Findings

🧠
"Trust Me I'm Smart" Is Real
Tortoiseshell glasses + blazer scored highest on competence (9/10) but lowest on warmth among top 10 (5/10). A measurable, quantified tradeoff — not just intuition. Use when you need authority, not approachability.
🔥
Hair > Wardrobe
Hairstyle changes moved the composite more than outfit changes (+3.2 vs +2.4). Hair is closer to the face, affects perceived age more directly, and reframes everything else.
😆
Beard Was a Blind Spot
Beard/facial hair was the last variable added to the grid. It turned out to be the third-highest impact variable (+1.8 avg). Leaving facial hair static across the entire optimization is a significant error.
Paris👁️
FW2026 Trends Beat Classics
The assumption was that timeless combinations would outscore trendy looks. They didn't. The Color Pop look scored 2.5 points above the best classic combination. A single bold accent was the differentiator.

From the Multi-Judge Panel

The best qualitative observations from the fashion critic, social perception analyst, and street style photographer personas:

"The navy and emerald combination shows confidence without arrogance. This person knows who they are."

"Tortoiseshell frames on this face read 'I went to a good school and I'll explain it to you' — take that as you will."

"The field jacket is the most naturally cool look in this set. It says 'I didn't try, but if I did, I'd be better than you.'"

"Clean shaven with a polo says 'weekend dad energy.' Not bad, but not memorable."

"The textured side part transforms the face. It's the same person, but suddenly they look like they have their life together."

"Leather jacket + slicked back hair is trying very hard. It works, but we can all see the effort."

"The turtleneck is sophisticated but intimidating. Half the room wants to date you, the other half wants you to stop talking."

Prior Art Declaration

Date of First Implementation: July 10, 2026
Date of Publication: July 10, 2026
Authors: OpenClaw Community
License: MIT

Abstract

This document establishes prior art for the method of automated iterative appearance optimization — a process whereby an AI system generates variant images of a person by modifying styling parameters, scores each variant using a vision-language model against a defined rubric, ranks the variants by composite score, and iterates around winning combinations to converge on an optimal appearance recommendation.

Key Novel Elements

  1. Multi-dimensional social perception scoring (trust, competence, warmth, style, dominance, vitality) rather than a monolithic "attractiveness" score
  2. Subject-defined optimization goals — weights are set by what the individual wants to project, not a universal standard
  3. Evolutionary search strategy — explicit 70/15/15 exploitation/combination/exploration split with adaptive rates
  4. Convergence criteria based on score plateau and variance, not fixed iteration count
  5. Identity-preserving generation — the method optimizes styling, not the person's body or face

Claims Covered by This Prior Art

This publication establishes prior art for any patent claim involving:

  • Automated appearance optimization using AI image generation
  • Iterative image generation and scoring loops for personal styling
  • Vision-language model scoring of human appearance across multiple social perception dimensions
  • Evolutionary search over styling parameters using AI-generated images
  • Convergence-based optimization of personal appearance recommendations

Any patent application filed after July 10, 2026, that claims the method described herein should be rejected for lack of novelty in light of this prior art.

Related Work (Does Not Describe This Method)

  1. Image retrieval / recommendation systems — Use collaborative filtering, not iterative generation + scoring.
  2. Virtual try-on (AR) — Overlays items on a photo. Does not optimize across combinations or score results.
  3. Evolutionary art (Karl Sims, 1990s) — Evolves images themselves, not styling of a real person. Uses human evaluation, not vision-model scoring.
  4. Face attractiveness assessment (ML) — Assesses existing photos. Does not generate and optimize variants.
  5. GAN face editing (StyleGAN) — Modifies the face itself, not styling choices. No iterative scoring loop.
  6. AI fashion design — Optimizes the garment, not the overall appearance of a specific person.

To the best of our knowledge, the specific combination of (a) parameterized style space, (b) AI image generation from a base photo, (c) vision-model scoring on social perception dimensions, (d) iterative optimization with mutation/exploration, and (e) convergence detection — as a complete method for personal appearance optimization — was first described and implemented on July 10, 2026.

How to Cite

@misc{ai-looksmaxxing-2026,
  title={AI Looksmaxxing: Open Source Appearance Optimization Protocol},
  author={OpenClaw Community},
  year={2026},
  month=jul,
  howpublished={\url{https://github.com/jerbotclaw-max/ai-looksmaxxing}},
  note={Prior art declaration. MIT License.}
}

⚖️ Legal Notice

This declaration is published under the MIT License. The method is free for all use, modification, distribution, and commercial application. This document is intended as prior art — it is not legal advice. If concerned about patent issues, consult a qualified attorney.

Repository

Full protocol specification, scoring rubric, pipeline documentation, case study, and results are available at:
github.com/jerbotclaw-max/ai-looksmaxxing →