diff --git a/web/apps/photos/src/pages/cluster-debug.tsx b/web/apps/photos/src/pages/cluster-debug.tsx
index e5f8148fbc..1f6a1036f6 100644
--- a/web/apps/photos/src/pages/cluster-debug.tsx
+++ b/web/apps/photos/src/pages/cluster-debug.tsx
@@ -70,7 +70,7 @@ export default function ClusterDebug() {
             minBlur: 10,
             minScore: 0.8,
             minClusterSize: 2,
-            joinThreshold: 0.6,
+            joinThreshold: 0.76,
             earlyExitThreshold: 0.9,
             batchSize: 10000,
             offsetIncrement: 7500,
@@ -344,7 +344,7 @@ const ClusterList: React.FC<React.PropsWithChildren<ClusterListProps>> = ({
         index === 0
             ? 140
             : index === 1
-              ? 130
+              ? 110
               : Array.isArray(items[index - 2])
                 ? listItemHeight
                 : 36;
@@ -466,13 +466,10 @@ const ClusterResHeader: React.FC<ClusterResHeaderProps> = ({ clusterRes }) => {
             </Typography>
             <Typography variant="small" color="text.muted">
                 For each cluster showing only up to 50 faces, sorted by cosine
-                similarity to highest scoring face in the cluster.
-            </Typography>
-            <Typography variant="small" color="text.muted">
-                Below each face is its{" "}
-                <b>blur - score - cosineSimilarity - direction</b>.
+                similarity to its highest scoring face.
             </Typography>
             <Typography variant="small" color="text.muted">
+                Below each face is its blur, score, cosineSimilarity, direction.
                 Bad faces are outlined.
             </Typography>
         </Stack>
@@ -544,7 +541,7 @@ const FaceItem: React.FC<FaceItemProps> = ({ faceWithFile }) => {
     return (
         <FaceChip
             style={{
-                outline: isBadFace ? `1px solid wheat` : undefined,
+                outline: isBadFace ? `1px solid rosybrown` : undefined,
                 outlineOffset: "2px",
             }}
         >
diff --git a/web/packages/new/photos/services/ml/cluster.ts b/web/packages/new/photos/services/ml/cluster.ts
index 18a9b52151..e4928aa991 100644
--- a/web/packages/new/photos/services/ml/cluster.ts
+++ b/web/packages/new/photos/services/ml/cluster.ts
@@ -195,6 +195,7 @@ export const clusterFaces = (
         earlyExitThreshold,
         batchSize,
         offsetIncrement,
+        badFaceHeuristics,
     } = opts;
     const t = Date.now();
 
@@ -256,6 +257,7 @@ export const clusterFaces = (
             oldState,
             joinThreshold,
             earlyExitThreshold,
+            badFaceHeuristics,
             ({ completed }: ClusteringProgress) =>
                 onProgress({
                     completed: offset + completed,
@@ -432,6 +434,7 @@ const clusterBatchLinear = (
     oldState: ClusteringState,
     joinThreshold: number,
     earlyExitThreshold: number,
+    badFaceHeuristics: boolean,
     onProgress: (progress: ClusteringProgress) => void,
 ) => {
     const state: ClusteringState = {
@@ -454,7 +457,7 @@ const clusterBatchLinear = (
 
         // Find the nearest neighbour among the previous faces in this batch.
         let nnIndex: number | undefined;
-        let nnCosineSimilarity = joinThreshold;
+        let nnCosineSimilarity = 0;
         for (let j = i - 1; j >= 0; j--) {
             // ! This is an O(n^2) loop, be careful when adding more code here.
 
@@ -464,13 +467,15 @@ const clusterBatchLinear = (
             // The vectors are already normalized, so we can directly use their
             // dot product as their cosine similarity.
             const csim = dotProduct(fi.embedding, fj.embedding);
-            if (csim > nnCosineSimilarity) {
+            const threshold =
+                badFaceHeuristics && fj.isBadFace ? 0.84 : joinThreshold;
+            if (csim > nnCosineSimilarity && csim >= threshold) {
                 nnIndex = j;
                 nnCosineSimilarity = csim;
 
                 // If we've found something "near enough", stop looking for a
                 // better match (A heuristic to speed up clustering).
-                if (earlyExitThreshold > 0 && csim > earlyExitThreshold) break;
+                if (earlyExitThreshold > 0 && csim >= earlyExitThreshold) break;
             }
         }