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. 1. Load the image from INLINECODE1
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 … INLINECODE4

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.

CODEBLOCK0

Step 3 – Display in chat

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

Then write a short markdown summary:

CODEBLOCK1

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: INLINECODE8