Image Forge

Pixel-precise image editing with three-tier routing: deterministic CLI tools for exact operations, AI models for semantic edits, vision models for analysis.

Routing Decision

Use AI when the edit requires understanding what is in the image. Use ImageMagick when the edit requires knowing exactly what to do to the pixels.

Tier 1: Deterministic (magick, sips, transparent-background)

Use for operations with exact, numeric parameters:

Tier 2: AI Semantic (nano-banana-pro / Gemini)

Delegate to the nano-banana-pro skill when the edit requires understanding image content:

• Remove a person/object from a photo
• Change sky, weather, time of day
• Apply artistic style transfer
• Inpaint/outpaint regions
• Generate new image from scratch
• Content-aware fill after object removal

Nano Banana regenerates, it does not edit. Every call re-rolls the entire frame, even when the prompt asks to preserve regions verbatim. Expect ~40% of pixels to drift outside the target area — measured empirically, not hypothesized. For surgical regional edits (change the face, keep everything else identical), use the hybrid pattern in references/compositing.md § Surgical Regional Edits: let NB produce a drifted reference, then composite only the target region onto the pristine original with a feathered alpha mask.

Nano Banana also supports only discrete aspect ratios (1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9). For inputs at other ratios (e.g. 3:1 panoramics), see references/recipes.md § Aspect Ratio Padding for the pad-generate-crop recipe.

Tier 3: Vision Analysis (Claude Read tool)

Use the Read tool to inspect images before/after edits:

• Verify an edit succeeded
• Describe image contents
• Check composition and framing
• Identify colors, objects, text in image

Background Removal

Use transparent-background (InSPyReNet model) for AI background removal. It produces clean alpha mattes with smooth anti-aliased edges — far superior to ImageMagick's -transparent/-fuzz approach, which leaves white fringing in concave regions and jagged edges.

When to use each:

Python API (recommended):

from transparent_background import Remover
from PIL import Image

remover = Remover()
img = Image.open("input.jpg").convert("RGB")
out = remover.process(img, type="rgba")
out.save("output.png")

With trim (most common workflow):

python3 -c "
from transparent_background import Remover
from PIL import Image
remover = Remover()
img = Image.open('input.jpg').convert('RGB')
remover.process(img, type='rgba').save('/tmp/nobg.png')
"
magick /tmp/nobg.png -trim +repage output.png

Install: uv pip install --system transparent-background

Performance: ~2-3s per image on Apple Silicon (MPS). First run downloads the InSPyReNet model (~170MB).

Note: rembg is an alternative but has a broken gradio dependency as of May 2026. transparent-background is the more reliable option.

Scripts

All scripts are in ~/.claude/skills/image-forge/scripts/. Run with python3.

image_info.py — Inspect Image Metadata

python3 ~/.claude/skills/image-forge/scripts/image_info.py photo.jpg
python3 ~/.claude/skills/image-forge/scripts/image_info.py photo.jpg --field dimensions
python3 ~/.claude/skills/image-forge/scripts/image_info.py photo.jpg --field width

Returns clean JSON: dimensions, format, color space, depth, alpha, DPI, ICC profile, EXIF.

image_pipeline.py — Declarative Edit Pipeline

Write a JSON spec, get a single chained magick command. No intermediate files.

# Create spec
cat > /tmp/pipeline.json << 'EOF'
{
    "input": "photo.jpg",
    "output": "result.png",
    "steps": [
        {"op": "resize", "width": 800, "height": 600, "mode": "fill"},
        {"op": "brightness_contrast", "brightness": 5, "contrast": 10},
        {"op": "annotate", "text": "Title", "gravity": "South", "pointsize": 36, "fill": "white", "stroke": "black", "strokewidth": 2},
        {"op": "composite", "overlay": "watermark.png", "gravity": "SouthEast", "opacity": 25},
        {"op": "quality", "value": 90},
        {"op": "strip"}
    ]
}
EOF

# Dry run (print command)
python3 ~/.claude/skills/image-forge/scripts/image_pipeline.py /tmp/pipeline.json --dry-run

# Execute
python3 ~/.claude/skills/image-forge/scripts/image_pipeline.py /tmp/pipeline.json

Available operations: