Update app.py

app.py

@@ -15,6 +15,27 @@ MODEL_REPO = "rain1011/pyramid-flow-sd3"
 MODEL_VARIANT = "diffusion_transformer_768p"
 MODEL_DTYPE = "bf16"
 
+def center_crop(image, target_width, target_height):
+ width, height = image.size
+ aspect_ratio_target = target_width / target_height
+ aspect_ratio_image = width / height
+
+ if aspect_ratio_image > aspect_ratio_target:
+ # Crop the width (left and right)
+ new_width = int(height * aspect_ratio_target)
+ left = (width - new_width) // 2
+ right = left + new_width
+ top, bottom = 0, height
+ else:
+ # Crop the height (top and bottom)
+ new_height = int(width / aspect_ratio_target)
+ top = (height - new_height) // 2
+ bottom = top + new_height
+ left, right = 0, width
+
+ image = image.crop((left, top, right, bottom))
+ return image
+
 # Download and load the model
 def load_model():
  if not os.path.exists(MODEL_PATH):
@@ -67,12 +88,13 @@ def generate_video_from_image(image, prompt, duration, video_guidance_scale):
  torch_dtype = torch.bfloat16 if MODEL_DTYPE == "bf16" else torch.float32
  
  target_size = (1280, 720)
- image = ImageOps.fit(image, target_size, method=Image.LANCZOS, centering=(0.5, 0.5))
+ cropped_image = center_crop(image, 1280, 720)
+ resized_image = cropped_image.resize((1280, 720))
  
  with torch.no_grad(), torch.cuda.amp.autocast(enabled=True, dtype=torch_dtype):
  frames = model.generate_i2v(
  prompt=prompt,
- input_image=image,
+ input_image=resized_image,
  num_inference_steps=[10, 10, 10],
  temp=temp,
  guidance_scale=7.0,