y

app.py

@@ -143,7 +143,14 @@ def infer(prompt, negative_prompt, adapter_l_file, adapter_g_file, strength, noi
         pipe.scheduler = SCHEDULERS[scheduler_name].from_config(pipe.scheduler.config)
     
     # Get T5 embeddings for semantic understanding
-    t5_ids = t5_tok(prompt, return_tensors="pt", padding=True, truncation=True).input_ids.to(device)
+    t5_ids = t5_tok(
+        prompt, 
+        return_tensors="pt", 
+        padding="max_length", 
+        max_length=77,  # Match CLIP's standard length
+        truncation=True
+    ).input_ids.to(device)
+    print(t5_ids.shape)
     t5_seq = t5_mod(t5_ids).last_hidden_state
     
     # Get proper SDXL CLIP embeddings
@@ -155,20 +162,7 @@ def infer(prompt, negative_prompt, adapter_l_file, adapter_g_file, strength, noi
     
     # Apply CLIP-L adapter
     if adapter_l is not None:
-        # Ensure tensor shapes match for cross-attention
-        print(f"T5 seq shape: {t5_seq.shape}, CLIP-L shape: {clip_embeds['clip_l'].shape}")
-        
-        # Resize T5 sequence to match CLIP sequence length if needed
-        if t5_seq.size(1) != clip_embeds["clip_l"].size(1):
-            t5_seq_resized = torch.nn.functional.interpolate(
-                t5_seq.transpose(1, 2),
-                size=clip_embeds["clip_l"].size(1),
-                mode="nearest"
-            ).transpose(1, 2)
-        else:
-            t5_seq_resized = t5_seq
-            
-        anchor_l, delta_l, log_sigma_l, attn_l1, attn_l2, tau_l, g_pred_l, gate_l = adapter_l(t5_seq_resized, clip_embeds["clip_l"])
+        anchor_l, delta_l, log_sigma_l, attn_l1, attn_l2, tau_l, g_pred_l, gate_l = adapter_l(t5_seq, clip_embeds["clip_l"])
         gate_l_scaled = gate_l * gate_prob
         delta_l_final = delta_l * strength * gate_l_scaled
         clip_l_mod = clip_embeds["clip_l"] + delta_l_final
@@ -185,20 +179,7 @@ def infer(prompt, negative_prompt, adapter_l_file, adapter_g_file, strength, noi
     
     # Apply CLIP-G adapter  
     if adapter_g is not None:
-        # Ensure tensor shapes match for cross-attention
-        print(f"T5 seq shape: {t5_seq.shape}, CLIP-G shape: {clip_embeds['clip_g'].shape}")
-        
-        # Resize T5 sequence to match CLIP sequence length if needed
-        if t5_seq.size(1) != clip_embeds["clip_g"].size(1):
-            t5_seq_resized = torch.nn.functional.interpolate(
-                t5_seq.transpose(1, 2),
-                size=clip_embeds["clip_g"].size(1),
-                mode="nearest"
-            ).transpose(1, 2)
-        else:
-            t5_seq_resized = t5_seq
-            
-        anchor_g, delta_g, log_sigma_g, attn_g1, attn_g2, tau_g, g_pred_g, gate_g = adapter_g(t5_seq_resized, clip_embeds["clip_g"])
+        anchor_g, delta_g, log_sigma_g, attn_g1, attn_g2, tau_g, g_pred_g, gate_g = adapter_g(t5_seq, clip_embeds["clip_g"])
         gate_g_scaled = gate_g * gate_prob
         delta_g_final = delta_g * strength * gate_g_scaled
         clip_g_mod = clip_embeds["clip_g"] + delta_g_final