Update app.py

app.py

@@ -31,7 +31,7 @@ from huggingface_hub import hf_hub_download
 from safetensors.torch import load_file, save_file
 from diffusers import FluxPipeline
 from PIL import Image, ImageDraw, ImageFont
-from transformers import pipeline
+from transformers import pipeline, T5ForConditionalGeneration, T5Tokenizer
 from refiners.fluxion.utils import manual_seed
 from refiners.foundationals.latent_diffusion import Solver, solvers
 from refiners.foundationals.latent_diffusion.stable_diffusion_1.multi_upscaler import (
@@ -40,6 +40,9 @@ from refiners.foundationals.latent_diffusion.stable_diffusion_1.multi_upscaler i
 )
 from datetime import datetime
 
+model = T5ForConditionalGeneration.from_pretrained("t5-large")
+tokenizer = T5Tokenizer.from_pretrained("t5-large")
+
 def log(msg):
     print(f'{datetime.now().time()} {msg}')
 
@@ -546,11 +549,18 @@ def upscaler(
     return enhanced_image
 
 def summarize_text(
-    text
+    text, max_len=20, min_len=10
 ):
     log(f'CALL summarize_text')
-    summ = pipeline("summarization", model="Falconsai/text_summarization")
-    summary = summ(text, do_sample=False)[0]["summary_text"]
+    inputs = tokenizer.encode("summarize: " + text, return_tensors="pt", max_length=float('inf'), truncation=False)
+    while len(inputs[0]) > max_len:
+        inputs[0][:512] = model.generate(
+            inputs[0][:512], 
+            length_penalty=2.0,
+            num_beams=4,
+            early_stopping=True
+        )
+    summary = tokenizer.decode(inputs[0])
     log(f'RET summarize_text with summary as {summary}')
     return summary
 
@@ -638,7 +648,7 @@ def handle_generation(artist,song,genre,lyrics):
     index = 1
     names = []
     for img in imgs:
-        scaled_by = 4
+        scaled_by = 2
         labeled_img = add_song_cover_text(img,artist,song,height*scaled_by,width*scaled_by)
         name = f'{artist} - {song} ({index}).png'
         labeled_img.save(name)