ls /tmp/workspace/ 2>/dev/null || mkdir -p /tmp/workspace/photo-edit && echo "ready"

mkdir -p /tmp/workspace/photo-edit && cd /tmp/workspace/photo-edit && curl -s -o input.jpg "https://d361swluvexzqs.cloudfront.net/chats-attachments/1000513430-1784397952820-tajndl.jpg" && echo "Downloaded" && python3 -c "from PIL import Image; img = Image.open('input.jpg'); print(img.size, img.mode)"

pip install --user rembg 2>/dev/null | tail -1; python3 -c "import rembg; print('rembg ok')" 2>&1 | tail -3

pip install --user rembg[gpu] onnxruntime 2>&1 | tail -5

pip install --user rembg 2>&1 | tail -5

python3 << 'EOF' from PIL import Image, ImageDraw, ImageFilter import numpy as np img = Image.open("/tmp/workspace/photo-edit/input.jpg").convert("RGB") w, h = img.size print(f"Image size: {w}x{h}") # The image is rotated sideways (selfie). Let's understand the layout: # 3 main foreground people take up most of the frame # Background people are in the bottom-right area (smaller, further away) # Coordinates: bottom-right cluster of smaller faces # Let's save a copy to analyze regions img.save("/tmp/workspace/photo-edit/analysis.jpg") print("Saved for analysis") # Create a numpy array arr = np.array(img) print("Array shape:", arr.shape) EOF

python3 << 'EOF' from PIL import Image, ImageDraw, ImageFilter import numpy as np import cv2 img = cv2.imread("/tmp/workspace/photo-edit/input.jpg") h, w = img.shape[:2] print(f"Image: {w}x{h}") # The photo is a sideways selfie (900x1600). # Looking at the image visually: # - 3 main subjects fill most of the frame (top-center to bottom-left) # - Background people are the smaller heads in the BOTTOM-RIGHT quadrant # approximately: x: 500-900, y: 950-1600 # Let's use inpainting to remove the background people # First create a mask for the background people region mask = np.zeros((h, w), dtype=np.uint8) # Background people region - the cluster of smaller faces on the right # Based on visual analysis: they appear in the bottom-right area # In the 900x1600 image (portrait): # The background people are roughly in: # x: 530-900, y: 960-1600 (bottom-right) # Draw mask regions for background people # Using polygon shapes for more accurate masking # Region 1: The group of ~4-5 background people (bottom right) pts1 = np.array([ [530, 960], [900, 960], [900, 1600], [530, 1600] ], dtype=np.int32) cv2.fillPoly(mask, [pts1], 255) # Also mask the partial person visible behind the main 3 (top area, right side) # At approximately x:620-900, y:0-200 pts2 = np.array([ [640, 0], [900, 0], [900, 250], [640, 250] ], dtype=np.int32) cv2.fillPoly(mask, [pts2], 255) # Save mask for inspection cv2.imwrite("/tmp/workspace/photo-edit/mask.jpg", mask) # Apply inpainting result = cv2.inpaint(img, mask, inpaintRadius=15, flags=cv2.INPAINT_TELEA) cv2.imwrite("/tmp/workspace/photo-edit/output_v1.jpg", result, [cv2.IMWRITE_JPEG_QUALITY, 95]) print("Done - output_v1.jpg saved") EOF

python3 << 'EOF' from PIL import Image, ImageDraw, ImageFilter import numpy as np import cv2 img = cv2.imread("/tmp/workspace/photo-edit/input.jpg") h, w = img.shape[:2] print(f"Image: {w}x{h}") # From visual analysis: # The image is 900x1600 portrait (but shot sideways) # 3 MAIN foreground people (keep): # - Person 1 (top): face ~y:150-500, x:350-900 # - Person 2 (middle): face ~y:490-680, x:280-620 # - Person 3 (bottom-left, closest): face ~y:680-1050, x:0-580, body extends to y:1600 # # BACKGROUND people to REMOVE (smaller, on the right side): # - 4th person (partial head): approx x:570-720, y:940-1080 # - 5th person: approx x:640-820, y:1010-1200 # - 6th person: approx x:700-900, y:1080-1330 # - 7th person (partial at bottom-right corner): x:750-900, y:1300-1600 # - Also a partial figure at very top-right: x:630-900, y:0-120 mask = np.zeros((h, w), dtype=np.uint8) # Background people cluster (right side, bottom half) # Use a polygon that tightly covers the background people # but avoids the main 3 foreground subjects # The background people are roughly in the right half from y~940 down # with the main person 3's body on the left side of that region # Right-side background group pts_bg = np.array([ [560, 940], [900, 780], # upper right [900, 1600], # bottom right [560, 1600], # bottom left of mask [400, 1350], # curve around main person's body [440, 1100], # [510, 960], # ], dtype=np.int32) cv2.fillPoly(mask, [pts_bg], 255) # Also mask partial figure at top-right background pts_top = np.array([ [640, 0], [900, 0], [900, 130], [640, 130] ], dtype=np.int32) cv2.fillPoly(mask, [pts_top], 255) # Save mask overlay for inspection debug = img.copy() debug[mask > 0] = [0, 0, 255] # red overlay cv2.imwrite("/tmp/workspace/photo-edit/mask_debug.jpg", debug) # Now inpaint with larger radius for smoother fill result = cv2.inpaint(img, mask, inpaintRadius=20, flags=cv2.INPAINT_TELEA) cv2.imwrite("/tmp/workspace/photo-edit/output_v2.jpg", result, [cv2.IMWRITE_JPEG_QUALITY, 95]) print("Done - output_v2.jpg saved") EOF

ls /tmp/workspace/photo-edit/

python3 << 'EOF' from PIL import Image, ImageDraw, ImageFilter import numpy as np import cv2 img = cv2.imread("/tmp/workspace/photo-edit/input.jpg") h, w = img.shape[:2] print(f"Image: {w}x{h}") # Keep only the 2 FRONT people (Person 1 top-center, Person 2 middle) # Remove: Person 3 (bottom-left large face) + all background people on right # # Person 1 face: top area, x:350-900, y:100-500 # Person 2 face: middle, x:220-620, y:490-680 # Their bodies: x:0-900, y:0~800 (left half) # # REMOVE everything below y~750 on the left (person 3 body/face) # AND remove the background people on the right from y~780 down # # Strategy: mask out person 3 (y:680-1600, x:0-580) + background people (x:530-900, y:780-1600) # Then inpaint with background texture mask = np.zeros((h, w), dtype=np.uint8) # Person 3 region (the frontmost large face + body, bottom portion) # Face starts around y:680, body extends to bottom pts_p3 = np.array([ [0, 780], [580, 780], [620, 960], [580, 1600], [0, 1600], ], dtype=np.int32) cv2.fillPoly(mask, [pts_p3], 255) # Background people on the right side (y: 780 - 1600, x: 530-900) pts_bg = np.array([ [530, 780], [900, 780], [900, 1600], [530, 1600], ], dtype=np.int32) cv2.fillPoly(mask, [pts_bg], 255) # Also partial top-right background figure pts_top = np.array([ [640, 0], [900, 0], [900, 140], [640, 140] ], dtype=np.int32) cv2.fillPoly(mask, [pts_top], 255) # Debug overlay debug = img.copy() debug[mask > 0] = (0, 0, 220) cv2.imwrite("/tmp/workspace/photo-edit/mask_v3_debug.jpg", debug) # Inpaint - use TELEA with large radius for smooth background fill result = cv2.inpaint(img, mask, inpaintRadius=25, flags=cv2.INPAINT_TELEA) cv2.imwrite("/tmp/workspace/photo-edit/output_v3.jpg", result, [cv2.IMWRITE_JPEG_QUALITY, 96]) print("Saved output_v3.jpg") EOF

python3 << 'EOF' # Let's do a more refined version with better mask boundaries # Also do multiple passes of inpainting for better quality from PIL import Image, ImageDraw, ImageFilter import numpy as np import cv2 img = cv2.imread("/tmp/workspace/photo-edit/input.jpg") h, w = img.shape[:2] # Looking at the image carefully: # The 2 FRONT people (keep): # - Person 1: Largest face, top of frame (y:100-520, x:330-900) # - Person 2: Second face below (y:450-700, x:200-660) # Their clothing/body fills left portion # # REMOVE: # - Person 3 (bottom large face + body): y>650 on left side # - All background people on right: from y~760 onwards on right half # - Top-right partial background person mask = np.zeros((h, w), dtype=np.uint8) # Bottom-left: Person 3's face and body # His face center is around y:800-1000, x:0-520 pts_p3_face = np.array([ [0, 700], [560, 700], [640, 850], [620, 1600], [0, 1600], ], dtype=np.int32) cv2.fillPoly(mask, [pts_p3_face], 255) # Right side background people (from about y:750 downward, x:520-900) pts_bg_right = np.array([ [520, 750], [900, 750], [900, 1600], [520, 1600], ], dtype=np.int32) cv2.fillPoly(mask, [pts_bg_right], 255) # Top-right partial background figure pts_top_right = np.array([ [640, 0], [900, 0], [900, 150], [640, 150] ], dtype=np.int32) cv2.fillPoly(mask, [pts_top_right], 255) # Dilate mask slightly to avoid fringing kernel = np.ones((5,5), np.uint8) mask_dilated = cv2.dilate(mask, kernel, iterations=2) # Double-pass inpainting for better results result1 = cv2.inpaint(img, mask_dilated, inpaintRadius=20, flags=cv2.INPAINT_TELEA) result2 = cv2.inpaint(result1, mask_dilated, inpaintRadius=15, flags=cv2.INPAINT_NS) cv2.imwrite("/tmp/workspace/photo-edit/output_final.jpg", result2, [cv2.IMWRITE_JPEG_QUALITY, 96]) print("Saved output_final.jpg") # Also save a side-by-side comparison side = np.hstack([ cv2.resize(img, (450, 800)), cv2.resize(result2, (450, 800)) ]) cv2.imwrite("/tmp/workspace/photo-edit/comparison.jpg", side) print("Saved comparison.jpg") EOF

ls -lh /tmp/workspace/photo-edit/output_final.jpg /tmp/workspace/photo-edit/comparison.jpg