yes

app.py

@@ -89,12 +89,20 @@ def infer(prompt, adapter_l_file, adapter_g_file, strength, noise, gate_prob, us
     if noise > 0:
         clip_g_mod += torch.randn_like(clip_g_mod) * noise
 
+    # Combine embeddings
     prompt_embeds = torch.cat([clip_l_mod, clip_g_mod], dim=-1).to(dtype)
     neg_embeds = torch.zeros_like(prompt_embeds)
 
+    # Compute pooled embeds (mean pooling as default fallback)
+    pooled_prompt_embeds = prompt_embeds.mean(dim=1)
+    pooled_neg_embeds = neg_embeds.mean(dim=1)
+
+    # SDXL generation with required pooled embeddings
     image = pipe(
         prompt_embeds=prompt_embeds,
+        pooled_prompt_embeds=pooled_prompt_embeds,
         negative_prompt_embeds=neg_embeds,
+        negative_pooled_prompt_embeds=pooled_neg_embeds,
         num_inference_steps=20,
         guidance_scale=5.0
     ).images[0]