app.py

import os
import subprocess

# --- Ensure SAM checkpoint is present ---
SAM_CHECKPOINT = "sam_vit_h_4b8939.pth"
if not os.path.exists(SAM_CHECKPOINT):
    print("Downloading SAM checkpoint...")
    subprocess.run([
        "wget",
        "https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth"
    ])
# ----------------------------------------
import gradio as gr
import numpy as np
from PIL import Image
import cv2
import torch
from segment_anything import sam_model_registry, SamPredictor

# --- CONFIG ---
SAM_CHECKPOINT = "sam_vit_h_4b8939.pth"
SAM_MODEL_TYPE = "vit_h"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
BLUR_RADIUS = 10
# --------------

# Load SAM once
sam = sam_model_registry[SAM_MODEL_TYPE](checkpoint=SAM_CHECKPOINT)
sam.to(device=DEVICE)
predictor = SamPredictor(sam)

def soft_alpha(mask_uint8, blur_radius=10):
    blurred = cv2.GaussianBlur(mask_uint8, (0,0), sigmaX=blur_radius, sigmaY=blur_radius)
    return (blurred.astype(np.float32) / 255.0).clip(0.0, 1.0)

def make_overlay(img_rgb: np.ndarray, mask_uint8: np.ndarray) -> Image.Image:
    """Return a purple-tinted overlay on the selected region."""
    # Ensure 0/255 mask and uint8 image
    mask = (mask_uint8 > 127).astype(np.uint8)
    overlay = img_rgb.copy()

    # Create a purple layer
    purple = np.zeros_like(img_rgb, dtype=np.uint8)
    purple[..., 0] = 180  # R
    purple[..., 1] = 0    # G
    purple[..., 2] = 180  # B

    # Blend only on selected pixels
    sel = mask.astype(bool)
    blended = cv2.addWeighted(overlay[sel], 0.6, purple[sel], 0.4, 0)
    overlay[sel] = blended

    return Image.fromarray(overlay)

def isolate_with_click(image: Image.Image, evt: gr.SelectData):
    # Guard: if no image or no click event, return original image and a subtle info overlay
    if image is None or evt is None:
        return None, None

    img_rgb = np.array(image.convert("RGB"))
    predictor.set_image(img_rgb)

    # SAM expects input points as numpy array [[x,y]]
    x, y = evt.index  # (x, y) from Gradio click
    input_point = np.array([[x, y]], dtype=np.float32)
    input_label = np.array([1], dtype=np.int32)  # 1 = foreground

    masks, scores, _ = predictor.predict(
        point_coords=input_point,
        point_labels=input_label,
        multimask_output=True
    )

    # If SAM didn't return masks, show original plus a faint marker
    if masks is None or len(masks) == 0:
        # Create a simple marker overlay to indicate click
        overlay = img_rgb.copy()
        cv2.circle(overlay, (int(x), int(y)), 12, (180, 0, 180), thickness=3)
        return None, Image.fromarray(overlay)

    # Pick the highest score mask
    best_idx = int(np.argmax(scores))
    best_mask = masks[best_idx].astype(np.uint8) * 255

    # Soft alpha for RGBA cutout
    alpha = soft_alpha(best_mask, blur_radius=BLUR_RADIUS)

    # Crop to bounding box (with small pad)
    ys, xs = np.where(best_mask == 255)
    if len(xs) == 0 or len(ys) == 0:
        # If mask is empty, still return overlay with click marker
        overlay = img_rgb.copy()
        cv2.circle(overlay, (int(x), int(y)), 12, (180, 0, 180), thickness=3)
        return None, Image.fromarray(overlay)

    x0, x1 = xs.min(), xs.max()
    y0, y1 = ys.min(), ys.max()
    pad = int(max(img_rgb.shape[:2]) * 0.02)
    x0 = max(0, x0 - pad); x1 = min(img_rgb.shape[1]-1, x1 + pad)
    y0 = max(0, y0 - pad); y1 = min(img_rgb.shape[0]-1, y1 + pad)

    fg_rgb = img_rgb[y0:y1+1, x0:x1+1]
    fg_alpha = alpha[y0:y1+1, x0:x1+1]

    # Compose RGBA correctly once
    rgba = np.dstack((fg_rgb, (fg_alpha * 255.0).astype(np.uint8)))
    cutout = Image.fromarray(rgba)

    # Build and return the purple overlay on the original image
    overlay_img = make_overlay(img_rgb, best_mask)

    return cutout, overlay_img

# --- Gradio UI ---
with gr.Blocks() as demo:
    gr.Markdown("### SAM Object Isolation\nUpload an image (or pick the demo), then click on the object to isolate it. The last panel shows a purple overlay of the mask.")
    inp = gr.Image(type="pil", label="Upload image", interactive=True)
    out_cutout = gr.Image(type="pil", label="Isolated cutout (RGBA)")
    out_overlay = gr.Image(type="pil", label="Segmentation overlay preview")

    # Click-to-segment
    inp.select(isolate_with_click, inputs=[inp], outputs=[out_cutout, out_overlay])

    # Demo example at the bottom that populates the upload image
    gr.Examples(
        examples=["demo.png"],   # ensure demo.png is in your repo
        inputs=inp,
        label="Try with demo image"
    )

demo.launch(share=True)