diff --git a/comfy/ldm/ace/ace_step15.py b/comfy/ldm/ace/ace_step15.py
index 69338336d..1d7dc59a8 100644
--- a/comfy/ldm/ace/ace_step15.py
+++ b/comfy/ldm/ace/ace_step15.py
@@ -1110,7 +1110,7 @@ class AceStepConditionGenerationModel(nn.Module):
 
         return encoder_hidden, encoder_mask, context_latents
 
-    def forward(self, x, timestep, context, lyric_embed=None, refer_audio=None, audio_codes=None, is_covers=None, **kwargs):
+    def forward(self, x, timestep, context, lyric_embed=None, refer_audio=None, audio_codes=None, is_covers=None, replace_with_null_embeds=False, **kwargs):
         text_attention_mask = None
         lyric_attention_mask = None
         refer_audio_order_mask = None
@@ -1140,6 +1140,9 @@ class AceStepConditionGenerationModel(nn.Module):
             src_latents, chunk_masks, is_covers, precomputed_lm_hints_25Hz=precomputed_lm_hints_25Hz, audio_codes=audio_codes
         )
 
+        if replace_with_null_embeds:
+            enc_hidden[:] = self.null_condition_emb.to(enc_hidden)
+
         out = self.decoder(hidden_states=x,
                            timestep=timestep,
                            timestep_r=timestep,
diff --git a/comfy/model_base.py b/comfy/model_base.py
index 3bb54f59e..3aa345254 100644
--- a/comfy/model_base.py
+++ b/comfy/model_base.py
@@ -1552,6 +1552,8 @@ class ACEStep15(BaseModel):
 
         cross_attn = kwargs.get("cross_attn", None)
         if cross_attn is not None:
+            if torch.count_nonzero(cross_attn) == 0:
+                out['replace_with_null_embeds'] = comfy.conds.CONDConstant(True)
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
 
         conditioning_lyrics = kwargs.get("conditioning_lyrics", None)