Create app.py

app.py

@@ -0,0 +1,97 @@
+import os
+import cv2
+import numpy as np
+import onnxruntime as ort
+import gradio as gr
+from huggingface_hub import hf_hub_download
+
+# --- 1. Configuration & Model Loading ---
+# Replace with your "username/repo-name"
+REPO_ID = "A123123/AnimeAutoCensor" 
+# Retrieved from Space's Secrets
+HF_TOKEN = os.getenv("HF_TOKEN")
+
+try:
+    # Download .onnx file
+    model_path = hf_hub_download(
+        repo_id=REPO_ID, 
+        filename="model.onnx", 
+        token=HF_TOKEN
+    )
+    # Download .onnx_data file (ONNX Runtime automatically looks for this in the same directory)
+    hf_hub_download(
+        repo_id=REPO_ID, 
+        filename="model.onnx_data", 
+        token=HF_TOKEN
+    )
+    
+    # Initialize Inference Engine (Hugging Face Free Tier supports CPU only)
+    session = ort.InferenceSession(model_path, providers=['CPUExecutionProvider'])
+    input_name = session.get_inputs()[0].name
+    target_size = 640  # Your model training size
+except Exception as e:
+    print(f"Model loading failed. Please check Token and Repository Path: {e}")
+
+# --- 2. Core Processing Functions ---
+def apply_mosaic_mask(image_rgb, mask, mosaic_level=16):
+    h, w = image_rgb.shape[:2]
+    mask = (mask > 0).astype(np.uint8)
+    small = cv2.resize(image_rgb, (max(1, w // mosaic_level), max(1, h // mosaic_level)), interpolation=cv2.INTER_LINEAR)
+    mosaic_image = cv2.resize(small, (w, h), interpolation=cv2.INTER_NEAREST)
+    output_image = image_rgb.copy()
+    output_image[mask == 1] = mosaic_image[mask == 1]
+    return output_image
+
+def process_image(input_img):
+    if input_img is None:
+        return None
+    
+    # Gradio passes images as RGB numpy arrays
+    h_orig, w_orig = input_img.shape[:2]
+    
+    # A. Letterbox Preprocessing
+    scale = target_size / max(h_orig, w_orig)
+    new_h, new_w = int(h_orig * scale), int(w_orig * scale)
+    img_resized = cv2.resize(input_img, (new_w, new_h))
+    
+    canvas = np.zeros((target_size, target_size, 3), dtype=np.uint8)
+    pad_y = (target_size - new_h) // 2
+    pad_x = (target_size - new_w) // 2
+    canvas[pad_y:pad_y+new_h, pad_x:pad_x+new_w] = img_resized
+
+    # B. Normalization & Inference
+    input_tensor = canvas.astype(np.float32) / 255.0
+    input_tensor = (input_tensor - [0.485, 0.456, 0.406]) / [0.229, 0.224, 0.225]
+    input_tensor = input_tensor.transpose(2, 0, 1)[np.newaxis, ...].astype(np.float32)
+
+    outputs = session.run(None, {input_name: input_tensor})
+    pred = outputs[0][0][0]
+
+    # C. Post-processing (Restore Size)
+    mask_valid = pred[pad_y:pad_y+new_h, pad_x:pad_x+new_w]
+    mask_final = cv2.resize(mask_valid, (w_orig, h_orig))
+    binary_mask = (mask_final > 0.5).astype(np.uint8)
+    
+    # D. Apply Mosaic
+    result_rgb = apply_mosaic_mask(input_img, binary_mask)
+    return result_rgb
+
+# --- 3. Gradio Interface ---
+with gr.Blocks(title="AI Anime Auto-Censor") as demo:
+    gr.Markdown("# 🎨 AI Anime Auto-Censor (Trial Version)")
+    gr.Markdown("This tool uses AI to automatically detect and censor specific content. The model weights are protected and not accessible to the public.")
+    
+    with gr.Row():
+        with gr.Column():
+            input_i = gr.Image(type="numpy", label="Upload Image")
+            run_btn = gr.Button("Start Processing", variant="primary")
+        with gr.Column():
+            output_i = gr.Image(type="numpy", label="Censored Result")
+            
+    run_btn.click(fn=process_image, inputs=input_i, outputs=output_i)
+    
+    gr.Markdown("---")
+    gr.Markdown("### Instructions:\n1. Upload an image.\n2. Click 'Start Processing'.\n3. Download the result if satisfied.")
+
+if __name__ == "__main__":
+    demo.launch()