diff --git a/src/drivers/nl_sm2135eh.c b/src/drivers/nl_sm2135eh.c
index 5a14e07..50081d2 100644
--- a/src/drivers/nl_sm2135eh.c
+++ b/src/drivers/nl_sm2135eh.c
@@ -87,6 +87,15 @@ void nl_sm2135eh_set(const uint8_t ch[NL_CH_COUNT])
         uint8_t chan = dark ? SM_CHANNEL_STBY : 0x00u;
         uint8_t d[6] = {SM_ADDR_CURRENT, s_current, chan, ch[NL_CH_B], ch[NL_CH_G], ch[NL_CH_R]};
         sm_send(d, 6);
+        /* The WY data registers are cleared here too, even though this is the RGB branch:
+           the chip keeps whatever white value was last written, and a direct white -> colour
+           change never passes through the all-dark frame above. The stale white then bleeds
+           through as soon as the colour dims - the same defect as the colour -> white case
+           below (fixed in 00d5a4c, 2026-09-13), just mirrored: that fix made the WY branch
+           also write the (always-zero) RGB registers, this makes the RGB branch also write
+           the (always-zero) WY registers. Both directions had it; only one was caught first. */
+        uint8_t w[3] = {SM_ADDR_WY, 0, 0};
+        sm_send(w, 3);
     } else {
         /* The RGB data registers are written along with the channel select, even though they
            are always zero here: the chip keeps whatever colour was last written, and a direct