diff --git a/crates/paint/examples/big_dab_bench.rs b/crates/paint/examples/big_dab_bench.rs
index b8231b9003..94ee7b8e8d 100644
--- a/crates/paint/examples/big_dab_bench.rs
+++ b/crates/paint/examples/big_dab_bench.rs
@@ -46,7 +46,11 @@ fn main() {
     // Two points a pixel apart: the generator places the first dab at once (a lone point waits).
     r.push(&[StrokePoint::new(3000.0, 2000.0, 1.0), StrokePoint::new(3001.0, 2000.0, 1.0)]);
     let push = t.elapsed();
-    println!("  push (rasterize+accumulate): {:>5.1} ms  (accumulate ≈ {:.1} ms)", push.as_secs_f64() * 1e3, push.saturating_sub(raster).as_secs_f64() * 1e3);
+    println!(
+        "  push (rasterize+accumulate): {:>5.1} ms  (push − warm rasterize, first dab ≈ {:.1} ms)",
+        push.as_secs_f64() * 1e3,
+        push.saturating_sub(raster).as_secs_f64() * 1e3
+    );
     let t = Instant::now();
     let dmg = r.composite(&pre, &mut target, None, false, false);
     let comp = t.elapsed();
diff --git a/crates/paint/examples/coverage_bench.rs b/crates/paint/examples/coverage_bench.rs
index 469738c4c8..97746a7f0a 100644
--- a/crates/paint/examples/coverage_bench.rs
+++ b/crates/paint/examples/coverage_bench.rs
@@ -1,6 +1,8 @@
 //! Isolate dab accumulation over a 6016×6016 (36 MP) plane, without rasterization/compositing.
 //! Run in release; reports median of five runs after warmup. Compare the same harness on both
 //! revisions. Resident tiles model a long stroke; moving dabs also exercise new tile allocation.
+//! A final line times a large soft dab (3457x3457, about 3025 tiles) marching 500 px per dab on
+//! fresh maps: first dab and steady-state median, in ms.
 use std::hint::black_box;
 use std::time::Instant;
 
@@ -33,6 +35,31 @@ fn run(size: i32, moving: bool) -> (f64, f32) {
     (micros, checksum)
 }
 
+/// Large soft dab marching along x on a fresh map: (first dab ms, median of the rest ms).
+fn run_large() -> (f64, f64) {
+    let size = 3457;
+    let vals: Vec<f32> = (0..size * size)
+        .map(|i| {
+            let x = (i % size) as f32 + 0.5 - size as f32 / 2.0;
+            let y = (i / size) as f32 + 0.5 - size as f32 / 2.0;
+            (1.0 - (x * x + y * y).sqrt() / (size as f32 / 2.0)).max(0.0) * 0.1
+        })
+        .collect();
+    let mut map = CoverageMap::new(3);
+    let mut times = Vec::new();
+    for i in 0..30 {
+        let x = i * 500;
+        let start = Instant::now();
+        map.accumulate(black_box(Rect::new(x, 0, x + size, size)), black_box(&vals), 0.8, None, None);
+        times.push(start.elapsed().as_secs_f64() * 1e3);
+    }
+    black_box(map.get(0, 0));
+    let first = times.first().copied().unwrap_or(0.0);
+    let mut rest: Vec<f64> = times.into_iter().skip(1).collect();
+    rest.sort_by(f64::total_cmp);
+    (first, rest.get(rest.len() / 2).copied().unwrap_or(0.0))
+}
+
 fn main() {
     for size in [20, 50, 128, 256] {
         for moving in [false, true] {
@@ -50,4 +77,15 @@ fn main() {
             }
         }
     }
+    let (mut firsts, mut steadies) = (Vec::new(), Vec::new());
+    for _ in 0..5 {
+        let (first, steady) = run_large();
+        firsts.push(first);
+        steadies.push(steady);
+    }
+    firsts.sort_by(f64::total_cmp);
+    steadies.sort_by(f64::total_cmp);
+    if let (Some(first), Some(steady)) = (firsts.get(2), steadies.get(2)) {
+        println!("large dab 3457x3457 first_ms={first:.2} steady_ms={steady:.2}");
+    }
 }
diff --git a/crates/paint/src/render.rs b/crates/paint/src/render.rs
index 83243bdafa..4a919fe7a0 100644
--- a/crates/paint/src/render.rs
+++ b/crates/paint/src/render.rs
@@ -559,21 +559,24 @@ impl CoverageMap {
                 let keys: Vec<_> = keys.into_iter().zip(touched).filter_map(|(key, touched)| touched.then_some(key)).collect();
                 if parallel(keys.len().saturating_mul(n)) {
                     use rayon::prelude::*;
-                    // Only genuinely parallel work takes ownership of tiles out of the map.
+                    // Only genuinely parallel work takes ownership of tiles out of the map. A new tile's
+                    // zeroing is the expensive part of a large dab, so it happens on the workers.
                     let mut tiles: Vec<_> = keys
                         .into_iter()
                         .map(|(key, tr)| {
                             self.dirty.insert(key);
-                            let tile = self.tiles.remove(&tile_key(key)).unwrap_or_else(CovTile::new);
-                            (key, tr, tile)
+                            (key, tr, self.tiles.remove(&tile_key(key)))
                         })
                         .collect();
-                    tiles.par_iter_mut().for_each(|(key, tr, tile)| {
+                    tiles.par_iter_mut().for_each(|(key, tr, slot)| {
+                        let tile = slot.get_or_insert_with(CovTile::new);
                         dab.prepare(tile);
                         dab.tile(*tr, *key, tile);
                     });
                     for (key, _, tile) in tiles {
-                        self.tiles.insert(tile_key(key), tile);
+                        if let Some(tile) = tile {
+                            self.tiles.insert(tile_key(key), tile);
+                        }
                     }
                 } else {
                     self.accumulate_serial(keys.into_iter(), dab);