Summary
Genesis 01 tests whether vision models recognize designer and season from pixels alone.
Key takeaways
@techreport{peace2025genesis01,
title = {Testing AI Vision's Understanding of High Fashion Nuances},
author = {Peace, Kalan},
institution = {Caeliai},
year = {2025},
url = {https://research.caeliai.com/research/genesis-01}
}Collection purity by model.
Source: Genesis 01 paper table (research/genesis-01.tex).
Positive values indicate lower confidence on impostors than on true matches. Shared scale: −0.100 to +0.100.
Source: genesis-01.tex.
Exact-match precision as published in the complete results table.
Source: genesis-01.tex.
Genesis 01 · Research paper
Also published as “A Vision Benchmark for Fashion AI”.
The first study compares how CLIP, SigLIP, and DINOv2 handle fashion images, focusing on uncertainty, impostor detection, and collection cohesion.
Using 12,147 Rick Owens runway images across 23 years, the study runs 3.66 million image comparisons. In this evaluation, SigLIP is the only model that becomes more cautious when the image is ambiguous.
The study is less about raw similarity scores and more about behavior: whether a model can recognize a design family, stay coherent across a collection, and mark uncertainty when it should.
Genesis 01 compares three vision backbones across three tasks: impostor detection, collection cohesion, and exact look matching. The goal is not generic image classification. It is how the models handle uncertain fashion images.
In this evaluation, SigLIP is the only model that becomes more cautious when the image is an impostor.
CLIP and DINOv2 remain more confident on the ambiguous cases. SigLIP is more expensive, but it shows the strongest uncertainty behavior in this evaluation.
Each scene maps to a real product problem: false similarity, weak collection understanding, or failure to retrieve the precise object the user means.
From pixels alone, the model has to recognize that visually adjacent looks are still the wrong designer or wrong season and should trigger uncertainty.
One runway look should retrieve its family, not a group of vaguely similar silhouettes from different years. This separates collection matching from general visual similarity.
Retrieval also has to stay precise. When a user points at a specific look, the system should find that exact target instead of drifting toward a general aesthetic neighborhood.
A vision model needs to show when it is unsure.
If the model cannot mark uncertainty, its output may favor a visually similar but incorrect reference. That is a model-behavior issue, not only a score.
Recommendation systems may use visual similarity to describe products. If a model cannot separate designer identity from visual similarity, its output may become an overconfident approximation.