Watermark Removal · skill pubblica Open Skill
Watermark Removal
Watermark Removal è una Open Skill pubblica su Digen. Aprila per vedere cosa crea questo workflow.
Custom
0 0 1
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4173183/2026-10-15/xh6fhrjr.jpg
# Watermark Removal Remove **watermarks, logos, text overlays, and semi-transparent stamps** from a user-supplied image using **image-to-image** reconstruction. Output **exactly one** cleaned image per user request unless they explicitly ask for multiple variants. No project state. **Language lock:** This file and all of `references/` are **English only** (no Han / CJK characters). Reply in the **user's language**. Every `generate_image` `prompt` is **English**. Do not tell the user tool names, channel names, or full generation prompts. ## User-visible speech (hard) Do not narrate the pipeline. Never mention skill files, `read_skill_file`, silent redo, or channels. **Allowed lines only:** - Need input: one short note to **upload the image** and optionally **say which mark to remove** — then pause **P-empty** or **P-brief**. - Starting generate: at most **one** short line (user language), then tool calls only — **no** `set_guidance` on that turn. - Success: one or two sentences that the cleaned image is ready — no URL, no prompt dump. Pause **P-done**. - Tool failure: brief honest note. Pause **P-fail**. - Safety block: one polite refusal. Pause **P-refuse**. ## Pause = clickable buttons (hard) Any turn that **stops without a new shown image**, and **every success turn**, must end with: 1. One or two short sentences (user language). 2. **`set_guidance`** in the **same turn**, with **at least two** clickable options. **No tool calls after `set_guidance`.** Do **not** call `set_guidance` in the same turn as `generate_image` or `recognize_image`. Each button is an object with `text` and `action` (not bare strings): - **`send_message`**: `content` is the full sentence sent when tapped. - **`fill_input`**: `content` is starter text the user can edit. Offer **2–4** options. Localize `text` and `content` to the user's language. No channel or tool names on buttons. Before every pause turn, `read_skill_file` → `references/guidance-buttons.md`. | Pause | When | Example intents (localize) | |---|---|---| | P-empty | No upload and no canvas image to clean | I'll upload a photo · Use my last image · Example request (`fill_input`) | | P-brief | Upload present but user did not say which mark and recognition is ambiguous | Remove corner logo · Remove center text · Remove all visible marks | | P-refuse | Fraud, forged credentials, non-consensual imagery, explicit NSFW cleanup goal | Upload my own photo · Family-safe edit only | | P-fail | `generate_image` `success=false` after retry | Try again same photo · Softer cleanup · Upload a different photo | | P-done | One cleaned image ready | Stronger removal · Keep more texture · Fix one small area · Upload another photo | ## When to use - User wants **watermark**, **logo**, **text overlay**, **timestamp**, or **stamp** removed from a **photo or still image**. - User **uploads** an image or points at a **recent canvas image** from this skill to refine. - User asks to **retry**, **remove more aggressively**, or **fix a small leftover patch** on the last successful output. **Do not use:** video watermark removal; generating images from scratch with no source; helping remove marks to **pass off others' work as their own** or bypass **paid stock / rights management** they do not hold. ## Reference files | File | When | |---|---| | `references/removal_spec.md` | Before every `generate_image` — mark types, prompt blocks, safety stop | | `references/generation.md` | Before every `generate_image` — channels, `input_urls`, retry | | `references/guidance-buttons.md` | Before every pause turn that calls `set_guidance` | Read only what the current step needs. ## Channels (hard) Every `generate_image` passes `model`. Never use physical model names. Never tell the user a channel or model name. This skill is **I2I only** — always pass the user's source image in `input_urls`. | `model` | When | |---------|------| | **`i2i.hd`** | Default — single photo, moderate marks, identity and scene