Remove tachie bg

Remove Tachie Background (帶圖去背)

Take one existing anime character standing illustration and produce a clean transparent-background RGBA PNG. Input is an image the user already has — hand-drawn, externally generated, green-screen, or fake-transparent. This skill only removes background; it does not generate.

This is the "bring-your-own-image" de-background path — the counterpart to generate-tachie (which generates and de-backgrounds in one flow). It exists because you often have a finished image from elsewhere that just needs cutting out.

When to use / not

• Use when: the user already has a tachie image and only wants it de-backgrounded.
• Not for: generating tachie (→ generate-tachie), scene backgrounds (→ generate-scene-bg).
• Style limit: anime-style illustrations only. isnet-anime is trained on anime art; semi-realistic / 3D-render / photographic inputs will degrade. Say so upfront rather than silently producing a poor cut.

Prerequisites

rembg[cpu] (with isnet-anime model), Pillow, numpy — already in requirements.txt.

Workflow

Step 0 — Confirm input and background type

Ask the user two things (do not auto-detect — one question is more reliable):

1. Is this an anime-style illustration? (if not, warn about degraded quality)
2. What is the original background? — green (chroma green) / solid (other flat color) / transparent (fake checkered) / other

Background type decides the decontamination strategy in Step 1.

Step 1 — De-background + post-process

Run the script:

python3 scripts/remove_bg.py <input> -o <output> --bg-type <green|solid|transparent|other>

It does: downscale to long-edge 1536 → rembg -m isnet-anime → alpha clamp (α≥120→255, α≤30→0) → if green: green-spill decontamination → save transparent PNG + dark/white check previews.

Step 2 — Dual-background eye check (the core)

The script emits <output>_darkcheck.png (dark 30,30,38) and <output>_whitecheck.png (white). Show BOTH to the user for eye inspection — they catch different defects:

• Dark background exposes bright artifacts (white rim, bright stray strands, gray gap-residue)
• White background exposes dark / green artifacts (residual green spill)

Looking at only one will miss half the defects. This is eyes-first — the previews are for a human to judge, not for a score.

When the dark check shows light patches, tell apart the two causes by comparing against the pre-rembg original: if the patch sits in a slit between hair strands and carries the source background's color/texture (e.g. checker pattern), it is trapped background residue (see Step 3); if the light stroke already exists in the original artwork, it is a painted-in highlight.

Step 3 — Honestly flag the unsolvable

If the dual check shows either of these, say plainly "de-background can't fix this" — do not pretend it's clean:

• Closed silhouette, including its MICRO version — background trapped inside enclosed regions. Macro: twin-tails curling back against the torso. Micro: thin slits between hair strands, hair-vs-neck gaps in dynamic poses — each slit is a tiny semi-enclosed region rembg can't reach. On checkered/fake-transparent sources the trapped background survives as light-gray patches (sometimes with visible checker texture) that only the dark check reveals. Verified 2026-07-03: assets/tachie/hongliu/debg-comparison/report_assets/verify_gap_residue.png
• Painted-in light strokes (flyaway strands, outline highlights drawn into the artwork) → rembg correctly keeps them as foreground; alpha/erosion post-processing is useless

These are design/drawing-stage problems, unsolvable at the de-background stage. The honest move is to name them and point at where they are.

Principles

• Eyes-first: the dual preview is judged by a human. Quantification (green-spill %, edge stats) is only for batch triage when you can't eyeball each one.
• Honest about the unsolvable: what can't be fixed, say so and mark where — don't fake it.
• Background color decides strategy: green needs de-greening, other solid colors need matching decontamination, neutral / fake-transparent is cleanest.

Known Limitations

• isnet-anime is anime-trained; non-anime styles degrade.
• "Perfect de-background" does not exist — closed silhouettes (macro or micro) and painted-in strokes are physically unsolvable. This skill's goal is "as clean as possible + honestly flag the unsolvable", not perfection.
• The main variable is silhouette openness, not background color. An open, non-enclosed pose/hairstyle de-backgrounds cleanly on any background; a gap-heavy one leaves residue on every background. Background color only decides what color the failure is when gaps exist.
• Each background type carries its own debt — background choice is a fuse, not a ranking:
  • Green screen → edge spill (measured 82.4% green edge), but the residue is green → recoverable by de-greening (82.4%→5% verified).
  • Fake-transparent / checkered → zero edge spill, but gap-trapped gray/checker residue has no decontamination — gray is too close to hair highlights and skin to target.
  • Practical call: gap-heavy designs (twin-tails, flyaway dynamic poses) → a green-screen source + de-green is the safer bet because its failure mode has an antidote; clean open silhouettes → fake-transparent stays zero-debt and skips a processing step.
  • Unverified: whether green's high contrast also makes rembg cut micro-gaps cleaner at the source — needs a same-prompt dual-background A/B (2026-07-03).

Web UI (optional)

webapp/ wraps this same pipeline in a local drag-drop browser UI (dual-preview side by side, download button) for manual use — double-click webapp/start.command, see webapp/README.md. It calls scripts/remove_bg.py as a subprocess, so behavior is identical to the CLI path.

Boundaries

• Generate tachie → generate-tachie
• Scene backgrounds → generate-scene-bg
• Batch processing, background auto-detection, alpha-matting refinement, perfect cutout on complex backgrounds → deliberately not in this MVP (v2 scope)