Update utils.py

utils.py

@@ -53,7 +53,7 @@ pipe.load_ip_adapter(
 #     cache_dir=CACHE_DIR
 # )
 
-def generate_sticker(input_image: Image.Image, prompt: str):
+def generate_sticker(input_image: Image.Image, style: str = Form("chibi")):
     """
     Given a user image and a prompt, generates a sticker/emoji-style portrait.
     """
@@ -66,7 +66,7 @@ def generate_sticker(input_image: Image.Image, prompt: str):
     # ).to(DEVICE)
 
     # Preprocess the image (resize, etc)
-    face_img = input_image.convert("RGB").resize((224, 224))
+    face_img = input_image.convert("RGB").resize((512, 512))
     # inputs = image_processor(images=face_img, return_tensors="pt").to(DEVICE)
     # with torch.no_grad():
     #     image_embeds = vision_encoder(**inputs).image_embeds
@@ -77,15 +77,19 @@ def generate_sticker(input_image: Image.Image, prompt: str):
     # IP-Adapter expects the reference image via image_embeds, which is produced by this function:
     # image_embeds = pipe.prepare_ip_adapter_image_embeds(face_img)
 
+    prompt == (f"A set of twelve {style}-style digital stickers"
+               "each with a different expression: laughing, angry, crying, sulking, thinking, sleepy, blowing a kiss, winking, surprised, happy, sad, and confused. "
+                "Each sticker has a bold black outline and a transparent background, in a playful, close-up cartoon style."
+    )
     # Run inference (low strength for identity preservation)
     result = pipe(
         prompt=prompt,
         image=init_image,
         # image_embeds=image_embeds,
         ip_adapter_image=face_img,
-        strength=0.65,
-        guidance_scale=7.5,
-        num_inference_steps=30
+        strength=0.6,
+        guidance_scale=8,
+        num_inference_steps=40
     )
     # Return the generated image (as PIL)
     return result.images[0]