Update app.py

app.py

@@ -14,6 +14,7 @@ import random
 import time
 import requests
 import pandas as pd
+import torch.nn as nn
 
 # Disable tokenizer parallelism
 os.environ["TOKENIZERS_PARALLELISM"] = "false"
@@ -56,6 +57,9 @@ pipe_i2i = AutoPipelineForImage2Image.from_pretrained(
 MAX_SEED = 2**32 - 1
 
 pipe.flux_pipe_call_that_returns_an_iterable_of_images = flux_pipe_call_that_returns_an_iterable_of_images.__get__(pipe)
+# Replace text embedding generation with pre-computed Longformer embeddings
+pipeline.text_encoder = None  # Disable the default CLIP text encoder
+pipeline.longformer_embeddings = transformed_embeddings  # Inject your embeddings
 
 def process_input(input_text):
     # Tokenize and truncate input
@@ -69,6 +73,17 @@ def process_input(input_text):
 input_text = "Your long prompt goes here..."
 inputs = process_input(input_text)
 
+# Get Longformer embeddings
+with torch.no_grad():
+    longformer_embeddings = longformer_model(**encoded_input).last_hidden_state
+
+# Create a transformation layer to match CLIP's embedding dimension
+transform_layer = nn.Linear(pooled_embeddings.size(-1), 512)
+transformed_embeddings = transform_layer(pooled_embeddings)  # Shape: [batch_size, 512]    
+
+# Pass your embeddings to the pipeline during generation
+image = pipeline(prompt=None, text_embeddings=pipeline.longformer_embeddings)
+
 class calculateDuration:
     def __init__(self, activity_name=""):
         self.activity_name = activity_name