Genesis 01: Testing AI Vision's Understanding of High Fashion Nuances

The first comprehensive evaluation of vision models for fashion intelligence

Return to Research

Abstract

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

1. Methodology

1.1 Model Selection

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

1.2 Dataset Construction

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

1.3 Technical Infrastructure

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

Three Core Intelligence Tests

Impostor Detection

Task: From pixels only, identify if candidates belong to the same designer and season/look as the query. Good behavior is uncertainty on impostors.

A

Rick Owens
Season: Spring 2019 • Look 27

B

Unknown designer (Impostor)
Not Rick Owens

C

Unknown designer (Impostor)
Not Rick Owens

Models see no brand text. They must match designer + season/look from pixels alone. SigLIP is properly uncertain on B and C; CLIP/DINOv2 are overconfident.

Family Recognition

Task: Given one look from a season, find other looks from the same designer and season (collection cohesion).

Query

Rick Owens • Spring 2010

Rick Owens • Spring 2010

Rick Owens • Spring 2010

Contamination • Rick Owens Fall 2016

SigLIP maintains season purity (63.5%); CLIP/DINOv2 mix in other seasons (e.g., Fall 2016).

Needle Search

Task: Find the exact same look in a large database (designer + season + look must match).

Target

Rick Owens • Fall 2004

EXACT • Fall 2004

Fall 2008

Fall 2013

...

SigLIP finds the exact Fall 2004 look; other models are distracted by near neighbors from 2008/2013.

Performance Metrics & Trade-offs

Technical Implementation

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)
                

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
                

Complete Results Summary

Explore the Research

Part of the Caeliai Genesis research series investigating computer vision approaches to fashion and design understanding.