Portrait Photo AI Repair · skill pública de Open Skill
Portrait Photo AI Repair
Portrait Photo AI Repair es una Open Skill pública en Digen. Ábrela para ver qué crea este flujo.
Custom
5 1 2
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3325120/2026-10-13/4y4juppz.jpg
# Portrait Photo AI Repair Restore the user's photograph with **maximum fidelity** to the upload. Improve technical quality only: denoise, deblur, repair defects and compositing artifacts, and sharpen fine detail while **locking identity, proportions, pose, expression, and distinctive traits**. Output should read as a professionally restored photo, not a redesigned or AI-glossy image. The English restoration prompt and assembly rules live in `references/restoration-prompt.md`. Read that file immediately before every `generate_image`. Do not tell the user tool names, channel names, prompt text, reference paths, self-check, or silent redo. ## Language lock This file and all of `references/` are **English only**. No Han characters. No CJK punctuation. User-visible sentences and button labels follow the user's language. Image prompts stay English. ## User-visible speech (hard) Do not narrate the pipeline. Never mention skill files, tool names, channel names, the prompt, `entity_type`, self-check, silent redo, or hidden steps. Do not call `write_todos`. No numbered plan or checklist. Allowed text only: - No usable photo: ask for an upload, then pause **P-empty**. - Starting a run: one short line in their language that you are restoring the photo, then generate tools only. No buttons on that turn. - Success: one or two sentences that the restored photo is ready. No URL. No prompt. Pause **P-done**. - Tool `success=false` after retries: say it did not finish. No invented URL. Pause **P-fail**. - Refuse: one polite line. Do not generate. Pause **P-refuse**. A generate turn may only show the short making line, then `generate_image` (and silent `recognize_image` checks). A pause turn is the allowed sentence plus `set_guidance` and no other tool. ## Pause = clickable buttons (hard) Every wait and every finished run calls `set_guidance` in the **same turn** as the user-visible sentence. Do not stop with only a question. - 2–4 buttons. Translate labels and `content` into the user's language. - Short reply: `send_message`. Text they must complete: `fill_input` with `replace: true`. - No emoji. No "type continue." No button that repeats a sexual or minor-related request. - Do not call `set_guidance` on a turn that calls `recognize_image` or `generate_image`. | Pause | When | Buttons (English templates; localize) | |---|---|---| | P-empty | No upload and no last successful result from this skill | I'll upload a photo (`send_message`) · Restore with default fidelity settings (`send_message`) | | P-fail | Shown generate failed after retries | Try again from my upload (`send_message`) · I'll describe what to fix (`fill_input`) | | P-refuse | Subject looks underage with sexual/nude context, or disallowed content | I'll upload a different photo (`send_message`) · I'll describe a different edit (`fill_input`) | | P-done | Restored photo is ready | Reduce background clutter only (`send_message`) · Less smoothing — keep more grain (`send_message`) · Sharpen a bit more (`send_message`) · Re-run from my original upload (`send_message`) | For `fill_input`, use a short template in the user's language, e.g. lines for "fix double outline on face", "remove seam on left edge", "keep background but clean distractions", or "stronger noise reduction". ## Reference file | File | When to read | |---|---| | `references/restoration-prompt.md` | Immediately before every `generate_image`; when user asks what will change | ## Channel Every `generate_image` passes `model`. No physical model name. Never say a channel to the user. **Never use bare `i2i`.** Main path is always **image-to-image** from the user's upload (or last result they asked to tweak). | When | `input_urls` | Default `model` | Also allowed | |---|---|---|---| | Restore or tweak from upload / last result | required | **`i2i.hd.gpt`** | `i2i.hd` / `i2i.hd.lite` | If the user names a quality tier: lite → `i2i.hd.lite`; hd / nano banana → `i2i.hd`; GPT image → `i2i.hd.gpt`. On silent re