\documentclass[10.5pt, a4paper]{article}

\usepackage[T1]{fontenc}
\usepackage[top=0.75in, bottom=0.75in, left=0.85in, right=0.85in]{geometry}
\usepackage{xcolor}
\usepackage{hyperref}
\usepackage{booktabs}
\usepackage{array}
\usepackage{enumitem}
\usepackage{titlesec}
\usepackage{graphicx}
\usepackage{lmodern}
\usepackage{amssymb}
\usepackage{listings}
\usepackage{mdframed}
\usepackage{multirow}
\usepackage{colortbl}

% Colors
\definecolor{sage}{HTML}{BAD9B5}
\definecolor{charcoal}{HTML}{2A2A2C}
\definecolor{offwhite}{HTML}{EBEBEB}
\definecolor{codebg}{HTML}{F5F5F5}

% Hyperlinks
\hypersetup{colorlinks=true, linkcolor=charcoal, urlcolor=charcoal,
    pdftitle={Testing AI Vision's Understanding of High Fashion Nuances}}

% Section formatting
\titleformat{\section}{\normalsize\bfseries\color{charcoal}}{}{0em}{\MakeUppercase}
\titlespacing*{\section}{0pt}{10pt}{4pt}
\titleformat{\subsection}{\normalsize\bfseries}{}{0em}{}
\titlespacing*{\subsection}{0pt}{6pt}{2pt}

\setlength{\parindent}{0pt}
\setlength{\parskip}{4pt}
\renewcommand{\arraystretch}{1.2}

% Code listing style
\lstset{
    basicstyle=\ttfamily\fontsize{8}{10}\selectfont,
    backgroundcolor=\color{codebg},
    frame=single,
    framerule=0.4pt,
    rulecolor=\color{charcoal},
    breaklines=true,
    breakatwhitespace=true,
    showstringspaces=false,
    commentstyle=\color{charcoal},
    keywordstyle=\bfseries,
    language=Python,
}

% Key finding box
\newmdenv[
    linecolor=charcoal,
    linewidth=1pt,
    leftline=true,
    rightline=true,
    topline=true,
    bottomline=true,
    backgroundcolor=offwhite,
    innerleftmargin=16pt,
    innerrightmargin=16pt,
    innertopmargin=12pt,
    innerbottommargin=12pt,
]{findingbox}

% Sage accent box
\newmdenv[
    linecolor=sage,
    linewidth=2pt,
    leftline=true,
    rightline=false,
    topline=false,
    bottomline=false,
    innerleftmargin=12pt,
    innerrightmargin=0pt,
    innertopmargin=6pt,
    innerbottommargin=6pt,
]{sagebox}

% ─────────────────────────────────────────────
\begin{document}

% HEADER
\noindent
\begin{minipage}[t]{0.6\textwidth}
    \includegraphics[height=20pt]{../public/caeliai-Offical-Logo.png}\hspace{6pt}{\large\textbf{Caeliai}}\\[2pt]
    {\small\color{charcoal} Research / Genesis-01}
\end{minipage}
\hfill
\begin{minipage}[t]{0.35\textwidth}
    \raggedleft
    {\small\color{charcoal}
    Published: 2025\\
    \href{mailto:contact@caeliai.com}{contact@caeliai.com}
    }
\end{minipage}

\vspace{6pt}
\noindent\rule{\textwidth}{2pt}
\vspace{10pt}

% TITLE
{\fontsize{20}{24}\selectfont\textbf{Testing AI Vision's Understanding\\[4pt]of High Fashion Nuances}}

\vspace{6pt}
{\small\color{charcoal}\textit{The first comprehensive evaluation of vision models for fashion intelligence}}

\vspace{6pt}
\noindent{\color{sage}\rule{\linewidth}{3pt}}
\vspace{10pt}

% ABSTRACT
\section{Abstract}

We present the first comprehensive benchmark evaluating vision models for fashion intelligence, testing \texttt{CLIP}, \texttt{SigLIP}, and \texttt{DINOv2} on \texttt{12,147} Rick Owens runway images across \texttt{23} years of collections. Through \texttt{3.66 million} image comparisons across three evaluation frameworks --- impostor detection, collection cohesion, and exact matching --- we demonstrate that \texttt{SigLIP} achieves superior fashion intelligence with \texttt{+0.079} positive uncertainty detection gap and \texttt{63.5\%} collection purity, making it the only model suitable for production fashion AI systems despite \texttt{9.6x} processing cost.

\vspace{8pt}

\begin{findingbox}
\begin{center}
{\large\textbf{Key Finding}}\\[8pt]
\textbf{SigLIP} is the only model with proper uncertainty detection, achieving \textbf{+0.079} positive noise gap while \texttt{CLIP} (\textcolor{charcoal}{$-0.015$}) and \texttt{DINOv2} (\textcolor{charcoal}{$-0.070$}) show dangerous overconfidence.
\end{center}
\end{findingbox}

\vspace{10pt}

% METHODOLOGY
\section{1. Methodology}

\subsection{1.1 Model Selection}

We evaluated three state-of-the-art vision models representing different architectural approaches to multimodal understanding:

\begin{itemize}[leftmargin=*, itemsep=3pt]
    \item \textbf{SigLIP} (\texttt{google/siglip-so400m-patch14-384}): Google's enhanced CLIP variant with sigmoid loss, \texttt{1152}-dimensional embeddings, \texttt{400M} parameters.
    \item \textbf{CLIP} (\texttt{openai/clip-vit-base-patch32}): OpenAI's original contrastive model, \texttt{512}-dimensional embeddings.
    \item \textbf{DINOv2} (\texttt{facebookresearch/dinov2\_vitb14}): Meta's self-supervised vision model, \texttt{768}-dimensional embeddings.
\end{itemize}

\subsection{1.2 Dataset Construction}

Our dataset comprises \texttt{12,147} Rick Owens runway images spanning \texttt{23} years (\texttt{2002--2025}) across multiple collections:

\begin{sagebox}
\textbf{Fall Collections:} 2002, 2003, 2004, 2008, 2013, 2015, 2016, 2021, 2024\\
\textbf{Spring Collections:} 2004, 2006, 2007, 2010, 2012, 2013, 2015, 2016, 2017, 2019, 2020, 2023, 2025\\
\textbf{Coverage:} Menswear, Ready-to-Wear, Beauty campaigns
\end{sagebox}

\subsection{1.3 Technical Infrastructure}

Processing performed on \texttt{M4 Mac} with \texttt{Apple Silicon GPU}. Embeddings stored in \texttt{Pinecone} vector database for similarity search. All models implemented using \texttt{PyTorch} with \texttt{transformers} library.

\vspace{8pt}

% THREE CORE TESTS
\section{2. Three Core Intelligence Tests}

\renewcommand{\arraystretch}{1.3}
\begin{tabular}{>{\bfseries}p{0.22\textwidth} p{0.35\textwidth} p{0.35\textwidth}}
Test & Task & What good looks like \\
\midrule
Impostor Detection & From pixels only, identify if candidates belong to the same designer and season/look as the query & Proper \emph{uncertainty} on impostors --- not overconfident matching \\[6pt]
Family Recognition & Given one look from a season, find other looks from the same designer and season & High collection purity --- low cross-season contamination \\[6pt]
Needle Search & Find the exact same look in a database of 12,147 images & Rank 1 match = exact designer, season, and look number \\
\end{tabular}
\renewcommand{\arraystretch}{1.2}

\vspace{10pt}

% PERFORMANCE METRICS
\section{3. Performance Metrics \& Trade-offs}

\subsection{3.1 Uncertainty Detection}

Positive gap = model is properly less confident on impostors than true matches. Only SigLIP passes.

\vspace{6pt}
\begin{tabular}{>{\bfseries}p{0.15\textwidth} p{0.22\textwidth} p{0.22\textwidth} p{0.3\textwidth}}
Model & Uncertainty Gap & Verdict & Interpretation \\
\midrule
SigLIP  & \textbf{+0.079} & \textcolor{black}{PASS} & Proper uncertainty on impostors \\
CLIP    & $-0.015$        & \textcolor{charcoal}{FAIL} & Overconfident on noise \\
DINOv2  & $-0.070$        & \textcolor{charcoal}{FAIL} & Dangerously overconfident \\
\end{tabular}

\subsection{3.2 Collection Purity}

\begin{tabular}{>{\bfseries}p{0.15\textwidth} p{0.22\textwidth} p{0.5\textwidth}}
Model & Purity & Note \\
\midrule
SigLIP  & \textbf{63.5\%} & Maintains season cohesion \\
CLIP    & 48.8\%  & Cross-season contamination \\
DINOv2  & 48.6\%  & Near-random collection grouping \\
\end{tabular}

\subsection{3.3 Processing Speed}

Designer Gauntlet Test (50 comparisons):

\begin{tabular}{>{\bfseries}p{0.15\textwidth} p{0.15\textwidth} p{0.55\textwidth}}
Model & Speed & Trade-off \\
\midrule
CLIP    & 1.9 min  & Fast but unreliable \\
DINOv2  & 2.7 min  & Fast but unreliable \\
SigLIP  & 18.7 min & 9.6x cost, only production-ready option \\
\end{tabular}

\vspace{10pt}

% TECHNICAL IMPLEMENTATION
\section{4. Technical Implementation}

\subsection{4.1 SigLIP Embedding Generation}

\begin{lstlisting}
from transformers import AutoModel, AutoProcessor
import torch

# Load SigLIP model
siglip_model = AutoModel.from_pretrained(
    "google/siglip-so400m-patch14-384",
    torch_dtype=torch.float16
).to("mps")  # Apple Silicon GPU

# Generate embedding
image = Image.open("rick-owens-fall-2011-menswear-39.jpg")
inputs = siglip_processor(images=image, return_tensors="pt")
with torch.no_grad():
    image_features = siglip_model.get_image_features(**inputs)
    embedding = image_features / image_features.norm(dim=-1, keepdim=True)
\end{lstlisting}

\subsection{4.2 Vector Search Implementation}

\begin{lstlisting}
async def identify_image(self, image_embedding: List[float]):
    """Find most similar runway image using cosine similarity."""

    # Query Pinecone vector database
    query_result = self.img_index.query(
        vector=image_embedding,
        top_k=1,
        include_metadata=True,
        filter={"file_type": "look"}  # Only runway looks
    )

    # Calculate confidence
    similarity_score = match.score
    if similarity_score >= 0.95:
        confidence = 99.0  # Near-perfect match
    elif similarity_score >= 0.90:
        confidence = 95.0  # Strong match
\end{lstlisting}

\vspace{8pt}

% COMPLETE RESULTS
\section{5. Complete Results Summary}

\begin{tabular}{>{\bfseries}p{0.12\textwidth} p{0.22\textwidth} p{0.18\textwidth} p{0.18\textwidth} p{0.2\textwidth}}
\toprule
Model & Uncertainty Detection & Collection Purity & Needle Precision & Processing Speed \\
\midrule
\rowcolor{offwhite}
SigLIP  & +0.079 gap (PASS) & 63.5\% & 100\% accuracy & 18.7 min (10x cost) \\
CLIP    & $-0.015$ gap (FAIL) & 48.8\% & 90.0\% accuracy & 3.1 s \\
DINOv2  & $-0.070$ gap (FAIL) & 48.6\% & 71.4\% accuracy & 4.2 s \\
\bottomrule
\end{tabular}

\vspace{10pt}

% CONCLUSION
\section{Conclusion}

SigLIP is the only vision model suitable for production fashion AI. Its \texttt{+0.079} uncertainty gap and \texttt{63.5\%} collection purity demonstrate genuine semantic understanding of designer aesthetic, not just visual similarity matching. CLIP and DINOv2's negative uncertainty gaps --- meaning they are \emph{more} confident on impostors than true matches --- make them categorically unsuitable for high-precision fashion retrieval.

The \texttt{9.6x} processing cost of SigLIP is a real constraint but acceptable at production scale for high-value fashion intelligence applications. The benchmark establishes SigLIP at \texttt{google/siglip-so400m-patch14-384} as the foundational model for any serious fashion AI system requiring reliable similarity search across large runway archives.

\vspace{12pt}
\noindent\rule{\textwidth}{2pt}
\vspace{6pt}

\noindent{\small\color{charcoal}Caeliai Research $\cdot$ \href{https://caeliai.com/research}{caeliai.com/research} $\cdot$ \href{mailto:contact@caeliai.com}{contact@caeliai.com}}

\end{document}