From 694f860b6de101c41b2717286e7a2582b7360c6d Mon Sep 17 00:00:00 2001
From: Holden <holden@arch.fyi>
Date: Sun, 14 Jun 2026 02:52:12 +0000
Subject: [PATCH] Raise near-duplicate dedup threshold from 0.10 to 0.20

0.10 only removed burst shots (distance 0.01-0.05). Same-event photos with
similar pose and lighting sit at 0.10-0.20 and were passing through,
producing visually similar training images especially for people with small
datasets. 0.20 removes these while still preserving genuinely different
poses, expressions, and lighting conditions.
---
 winnow/diversity.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/winnow/diversity.py b/winnow/diversity.py
index 2d026b9..6e766db 100644
--- a/winnow/diversity.py
+++ b/winnow/diversity.py
@@ -304,7 +304,7 @@ def _select_by_embedding(
 # Near-Duplicate Removal
 # =============================================================================
 
-_DEDUP_THRESHOLD = 0.10  # cosine distance — burst shots are ~0.01-0.05 apart
+_DEDUP_THRESHOLD = 0.20  # cosine distance — burst shots ~0.01-0.05, same-event similar shots ~0.10-0.20
 
 
 def _dedup_embeddings(