yes

app.py

@@ -107,17 +107,38 @@ def encode_sdxl_prompt(prompt, negative_prompt=""):
     ).input_ids.to(device)
     
     with torch.no_grad():
-        # CLIP-L embeddings (768d) - [0] is sequence, [1] is pooled
+        # CLIP-L embeddings (768d) - works fine
         clip_l_embeds = pipe.text_encoder(tokens_l)[0]
         neg_clip_l_embeds = pipe.text_encoder(neg_tokens_l)[0]
         
-        # CLIP-G embeddings (1280d) - [0] is sequence, [1] is pooled (reuse same prompt)
-        clip_g_embeds = pipe.text_encoder_2(tokens_g)[0]
-        neg_clip_g_embeds = pipe.text_encoder_2(neg_tokens_g)[0]
+        # CLIP-G embeddings (1280d) - debug the output structure
+        clip_g_output = pipe.text_encoder_2(tokens_g)
+        print(f"CLIP-G output type: {type(clip_g_output)}")
+        print(f"CLIP-G output length: {len(clip_g_output) if hasattr(clip_g_output, '__len__') else 'no len'}")
+        if hasattr(clip_g_output, '__len__') and len(clip_g_output) > 0:
+            print(f"CLIP-G [0] shape: {clip_g_output[0].shape}")
+            if len(clip_g_output) > 1:
+                print(f"CLIP-G [1] shape: {clip_g_output[1].shape}")
+        
+        # Try different ways to get the sequence embeddings
+        if hasattr(clip_g_output, 'last_hidden_state'):
+            clip_g_embeds = clip_g_output.last_hidden_state
+        elif hasattr(clip_g_output, '__len__') and len(clip_g_output) > 0:
+            clip_g_embeds = clip_g_output[0]
+        else:
+            clip_g_embeds = clip_g_output
+            
+        neg_clip_g_output = pipe.text_encoder_2(neg_tokens_g)
+        if hasattr(neg_clip_g_output, 'last_hidden_state'):
+            neg_clip_g_embeds = neg_clip_g_output.last_hidden_state
+        elif hasattr(neg_clip_g_output, '__len__') and len(neg_clip_g_output) > 0:
+            neg_clip_g_embeds = neg_clip_g_output[0]
+        else:
+            neg_clip_g_embeds = neg_clip_g_output
         
         # Pooled embeddings for SDXL
-        pooled_embeds = pipe.text_encoder_2(tokens_g)[1]
-        neg_pooled_embeds = pipe.text_encoder_2(neg_tokens_g)[1]
+        pooled_embeds = clip_g_output[1] if hasattr(clip_g_output, '__len__') and len(clip_g_output) > 1 else clip_g_output.pooler_output
+        neg_pooled_embeds = neg_clip_g_output[1] if hasattr(neg_clip_g_output, '__len__') and len(neg_clip_g_output) > 1 else neg_clip_g_output.pooler_output
     
     return {
         "clip_l": clip_l_embeds,