import gradio as gr from transformers import OneFormerProcessor, OneFormerForUniversalSegmentation, AutoImageProcessor, AutoModelForDepthEstimation from PIL import Image, ImageFilter import numpy as np import torch from scipy.ndimage import gaussian_filter import cv2 # Load the OneFormer processor and model globally oneformer_processor = None oneformer_model = None try: oneformer_processor = OneFormerProcessor.from_pretrained("shi-labs/oneformer_coco_swin_large") oneformer_model = OneFormerForUniversalSegmentation.from_pretrained("shi-labs/oneformer_coco_swin_large") except Exception as e: print(f"Error loading OneFormer model: {e}") # Load the Depth Estimation processor and model globally depth_processor = None depth_model = None try: depth_processor = AutoImageProcessor.from_pretrained("depth-anything/Depth-Anything-V2-Small-hf") depth_model = AutoModelForDepthEstimation.from_pretrained("depth-anything/Depth-Anything-V2-Small-hf") except Exception as e: print(f"Error loading Depth Anything model: {e}") def apply_gaussian_blur_background(image, mask, radius): """Applies Gaussian blur to the background of the image.""" blurred_background = image.filter(ImageFilter.GaussianBlur(radius=radius)) img_array = np.array(image) blurred_array = np.array(blurred_background) foreground_mask = mask > 0 foreground_mask_3d = np.stack([foreground_mask] * 3, axis=-1) final_image_array = np.where(foreground_mask_3d, img_array, blurred_array) return Image.fromarray(final_image_array.astype(np.uint8)) def apply_depth_based_blur_background(image, mask, strength): """Applies lens blur to the background of the image based on depth estimation.""" resized_image = image.resize((512, 512)) image_np = np.array(resized_image) if depth_processor is None or depth_model is None: return "Error: Depth Anything model not loaded." # Prepare image for the depth estimation model inputs = depth_processor(images=resized_image, return_tensors="pt") with torch.no_grad(): outputs = depth_model(**inputs) predicted_depth = outputs.predicted_depth # Interpolate depth map to the resized image size prediction = torch.nn.functional.interpolate( predicted_depth.unsqueeze(1), size=resized_image.size[::-1], mode="bicubic", align_corners=False, ).squeeze().cpu().numpy() # Normalize the depth map to the range 0-1 depth_norm = (prediction - np.min(prediction)) / (np.max(prediction) - np.min(prediction)) num_blur_levels = 5 blurred_layers = [] for i in range(num_blur_levels): sigma = i * (strength / 5) # Adjust sigma based on strength if sigma == 0: blurred = image_np else: blurred = cv2.GaussianBlur(image_np, (15, 15), sigmaX=sigma, sigmaY=sigma, borderType=cv2.BORDER_REPLICATE) blurred_layers.append(blurred) depth_indices = ((1 - depth_norm) * (num_blur_levels - 1)).astype(np.uint8) final_blurred_image_resized = np.zeros_like(image_np) for y in range(image_np.shape[0]): for x in range(image_np.shape[1]): depth_index = depth_indices[y, x] final_blurred_image_resized[y, x] = blurred_layers[depth_index][y, x] final_blurred_pil_resized = Image.fromarray(final_blurred_image_resized.astype(np.uint8)) final_blurred_pil = final_blurred_pil_resized.resize(image.size) final_blurred_array = np.array(final_blurred_pil) original_array = np.array(image) mask_resized = mask.resize(image.size) mask_array = np.array(mask_resized) > 0 mask_array_3d = np.stack([mask_array] * 3, axis=-1) # Apply the mask to combine the original foreground with the blurred background final_output_array = np.where(mask_array_3d, original_array, final_blurred_array) return Image.fromarray(final_output_array.astype(np.uint8)) def segment_and_blur(input_image, blur_type, gaussian_radius=15, lens_strength=5): """Segments the input image and applies the selected blur.""" if oneformer_processor is None or oneformer_model is None: return "Error: OneFormer model not loaded." image = input_image.convert("RGB") # Rotate the image (assuming this is still needed) image = image.rotate(-90, expand=True) # Prepare input for semantic segmentation inputs = oneformer_processor(images=image, task_inputs=["semantic"], return_tensors="pt") # Semantic segmentation with torch.no_grad(): outputs = oneformer_model(**inputs) # Processing semantic segmentation output predicted_semantic_map = oneformer_processor.post_process_semantic_segmentation(outputs, target_sizes=[image.size[::-1]])[0] segmentation_mask = predicted_semantic_map.cpu().numpy() # Get the mapping of class IDs to labels id2label = oneformer_model.config.id2label # Set foreground label to person foreground_label = 'person' foreground_class_id = None for id, label in id2label.items(): if label == foreground_label: foreground_class_id = id break if foreground_class_id is None: return f"Error: Could not find the label '{foreground_label}' in the model's class mapping." # Black background mask output_mask_array = np.zeros(segmentation_mask.shape, dtype=np.uint8) # Set the pixels corresponding to the foreground object to white (255) output_mask_array[segmentation_mask == foreground_class_id] = 255 # Convert the NumPy array to a PIL Image mask_pil = Image.fromarray(output_mask_array, mode='L') if blur_type == "Gaussian": blurred_image = apply_gaussian_blur_background(image, mask_pil, gaussian_radius) elif blur_type == "Lens": blurred_image = apply_depth_based_blur_background(image, mask_pil, lens_strength) else: return "Error: Invalid blur type selected." return blurred_image iface = gr.Interface( fn=segment_and_blur, inputs=[ gr.Image(label="Input Image"), gr.Radio(["Gaussian", "Lens"], label="Blur Type", value="Gaussian"), gr.Slider(0, 30, step=1, default=15, label="Gaussian Blur Radius"), gr.Slider(0, 10, step=1, default=5, label="Lens Blur Strength"), ], outputs=gr.Image(label="Output Image"), title="Image Background Blur App", description="Upload an image, select a blur type (Gaussian or Lens), and adjust the blur parameters to blur the background while keeping the person in focus." ) if __name__ == "__main__": iface.launch()