Image Highlight Cropper

Image Highlight Cropper

Extract the 5 most visually interesting detail regions from a large image, crop them as squares, display them in the chat (2 per row), and save them as downloadable files.

Workflow

Step 1 – Analyse the image

Look carefully at the uploaded image. Identify the 5 most interesting regions based on:

• Visual richness: texture, fine detail, intricate patterns
• Artistic significance: focal points, expressive faces, key objects
• Contrast and color drama
• Unique or surprising elements the viewer might miss at first glance

⚠️ Signature rule: If a painter's signature is visible anywhere in the image, it must always be one of the 5 highlights. Center the crop precisely on the signature so it is fully visible and not cut off.

⚠️ Strict 1:1 square-fit rule: Every highlight must be a subject that naturally fills a square frame. Before selecting a region, ask: "Does this subject fit well into a square?"

• ✅ Good: a face, a bouquet of flowers, a single tree, a lamp post with surroundings, a signature, a wheel, a window, a doorway, a small group of figures
• ❌ Bad: a wide panoramic skyline, a long horizontal street scene, a tall thin tower spanning the full image height — these are inherently non-square and will look like an awkward strip crop
• If a subject is too wide or too tall to feel natural in a 1:1 frame, skip it and choose something else that genuinely fits a square
• The goal: each crop looks intentional and well-composed, as if it were a standalone photograph

For each region, record:

• A short label (e.g. "Gesicht links", "Goldornament", "Signatur")
• Center point (cx, cy) of the subject and a half-size (half the desired square side length)
• A one-sentence explanation of why this area is interesting

Step 2 – Crop with Python (Pillow)

Use the bash_tool + Python to:

1. Load the image from /mnt/user-data/uploads/<filename>
2. For each region: use the save_crop helper below — it handles edge cases automatically
3. Save each crop to /mnt/user-data/outputs/highlight_1.jpg … highlight_5.jpg

Choosing the half-size: Claude decides per region:

• Tiny ornament or signature → half = 100–200 px
• Face or small group → half = 200–350 px
• Large scene or texture area → half = 350–600 px
• Rule of thumb: the crop should feel like a natural close-up. Always use center-based coordinates.

Edge case — subject at image border: If the desired crop extends beyond the image boundary, save_crop takes whatever image content is available and places it centered on a white square canvas (side = longest available side). This keeps the result square and clean with no distortion.

from PIL import Image
import os

img = Image.open("/mnt/user-data/uploads/IMAGE_FILENAME")
w, h = img.size

def save_crop(img, w, h, cx, cy, half, path):
    x1 = max(0, cx - half)
    x2 = min(w, cx + half)
    y1 = max(0, cy - half)
    y2 = min(h, cy + half)
    rect_w = x2 - x1
    rect_h = y2 - y1

    crop = img.crop((x1, y1, x2, y2))

    if rect_w == rect_h:
        # Already square — save directly
        crop.save(path, quality=92)
    else:
        # Place on white square canvas, centered horizontally and vertically
        canvas_size = max(rect_w, rect_h)
        canvas = Image.new("RGB", (canvas_size, canvas_size), (255, 255, 255))
        paste_x = (canvas_size - rect_w) // 2
        paste_y = (canvas_size - rect_h) // 2
        canvas.paste(crop, (paste_x, paste_y))
        canvas.save(path, quality=92)

# Define crops by CENTER point (cx, cy) and half-size
crops = [
    # (label, cx, cy, half)
    ("highlight_1", cx1, cy1, half1),
    ("highlight_2", cx2, cy2, half2),
    ("highlight_3", cx3, cy3, half3),
    ("highlight_4", cx4, cy4, half4),
    ("highlight_5", cx5, cy5, half5),
]

os.makedirs("/mnt/user-data/outputs", exist_ok=True)

for label, cx, cy, half in crops:
    save_crop(img, w, h, cx, cy, half, f"/mnt/user-data/outputs/{label}.jpg")

print("Done")

Step 3 – Display in chat

After saving, use present_files to make all 5 crops downloadable.

Then write a short markdown summary:

## 🎨 5 Highlights aus dem Bild

**1. [Label]** – [Erklärung warum interessant]
**2. [Label]** – ...
...

Show the crops 2 per row by presenting them via present_files and listing them clearly with their labels. Users can download each file individually.

Step 4 – Invite feedback

Ask the user: "Soll ich andere Bereiche auswählen, oder die Größe der Crops anpassen?"

Tips for choosing good crops

• Avoid overlap between the 5 regions as much as possible
• Spread across the image — don't cluster all crops in one corner
• For paintings: prioritize faces, hands, symbolic objects, and texture-rich backgrounds
• For technical drawings: prioritize labels, detail views, and complex intersections
• half-size should be roughly 10–20% of the shorter image dimension so crops feel like genuine close-ups, not tiny stamps

Error handling

• If the image cannot be opened, tell the user and ask them to re-upload
• If Pillow is not installed: pip install Pillow --break-system-packages