Image Background Transparency (ImageMagick Flood Fill) Skill

Image Background Transparency (ImageMagick Flood Fill)

Use this skill to remove a solid or pastel-colored background from a raster image and produce a clean transparent PNG. The core technique is ImageMagick flood fill from the four corners, which works reliably even with pastel backgrounds that are spectrally close to subject pixels — situations where chroma-key and automatic background removal tools typically fail.

When to use this skill

A generated image (sticker, illustration, icon, avatar, etc.) has a solid or near-solid background that needs to become transparent.
Chroma-key removal (-transparent, -fuzz with exact color) produces halos, bites into edges, or fails with pastels.
The background color is known or can be sampled and is roughly uniform but may have slight color variation, anti-aliasing, or gradient edges.
The user explicitly asks for 去背景, 抠图, 背景变透明, transparent background, remove background, or similar.

Do not use this skill when the background is a complex photographic scene, or when the user needs subject cutout from a busy image. This technique shines with clean, solid or pastel backgrounds typical of generated sticker sheets, UI mockups, icons, and flat illustrations.

Inputs to collect

Ask only for what is missing:

Input image path: local file to process.
Background color(s): the dominant background color(s) to flood, e.g. #FFF5F7. If unknown, the agent should sample corner pixels from the image.
Output path: where to save the transparent PNG. Default: same directory with -transparent suffix.
Keep shadows?: whether soft drop shadows around the subject should be preserved (requires lower fuzz) or removed entirely.

Core command

magick input.png -alpha set -fuzz N% -fill none \
  -draw 'color 0,0 floodfill' \
  -draw 'color W,0 floodfill' \
  -draw 'color 0,H floodfill' \
  -draw 'color W,H floodfill' \
  output.png

where W = width - 1 and H = height - 1.

How to get image dimensions

identify -format '%w %h' input.png

Fuzz testing workflow (recommended)

Test multiple fuzz values and visually inspect each, then pick the cleanest one:

W=$(identify -format '%w' input.png)
H=$(identify -format '%h' input.png)
W=$((W - 1))
H=$((H - 1))

for f in 1 2 3 4; do
  magick input.png -alpha set -fuzz ${f}% -fill none \
    -draw "color 0,0 floodfill" -draw "color ${W},0 floodfill" \
    -draw "color 0,${H} floodfill" -draw "color ${W},${H} floodfill" \
    output-f${f}.png
done

Review each output and select the one with cleanest edges. Fine-tune in smaller steps (0.2%–0.5%) around the best value if needed.

Fuzz threshold guide

Fuzz selection principle: maximize fuzz without eating into subject pixels.

General thresholds:

| fuzz | typical outcome | | ------- | ----------------------------------------------------- | | 0.3%–1% | background remnants left around edges | | 2% | cleanest balance for most illustrations (Recommended) | | 3%–4% | starts biting into light-colored subject areas | | >5% | significantly erodes subject, usually unusable |

Common pitfalls

Light-subject bleed: When the subject has white or very light areas (clothing, outlines, highlights), the background and subject are spectrally close. Keep fuzz ≤ 2%, test increments of 0.2%, and accept a slight background halo rather than eating into the subject.
Shadow edge treatment: Soft drop shadows around the subject produce a gradient halo. Start at 2%, then fine-tune in 0.2% steps (e.g. 1.2%, 1.5%, 1.8%) if the shadow fringe is too thick. If shadows should remain, accept a slightly lower fuzz.
Checkerboard backgrounds: Some intermediate outputs may show a faux checkerboard pattern (alternating #F6F6F6 / #FCFCFC). Flood fill works on these as long as both checker colors are within the fuzz tolerance of each other — push fuzz up slightly (3–4%) and verify no subject damage.
Non-uniform backgrounds: If the background has a gradient or vignette, corner flood fill may still work if the gradient is gentle. Sample all four corners — if they differ significantly, flood from each corner with the sampled color as the anchor, or consider a different approach.

Execution guidance

Verify imagemagick is installed (which magick). If missing, instruct the user to install it.
Sample the background color from at least two corners using magick input.png -crop 1x1+0+0 txt:- or similar.
Get image dimensions.
Run the fuzz testing loop above.
Ask the user to review outputs, or visually inspect with read if available.
Once the best fuzz is identified, produce the final output.
Report the output path and fuzz value used.

Agent Skills: Image Background Transparency (ImageMagick Flood Fill)

Install this agent skill to your local

Skill Files