Seamless Image Restore · Open Skill 公開技能

Seamless Image Restore

Seamless Image Restore 是 Digen 上的公開 Open Skill。開啟後即可了解此工作流能製作什麼。

@HarperDiaz

Custom

0 0 0

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284956/2026-10-17/qctnpasf.webp

# Seamless Image Restore

Missing pieces, healed. Damaged, erased or incomplete areas are restored so the image reads as one seamless natural photograph or artwork.

## When to use

- The user wants damaged missing or erased areas of an image naturally restored into one seamless result (image restoration, inpaint complete, heal damaged photo)
- A variation on a previous result: 3:4

Not for: Video output (use a video skill).

## Inputs and gates

| Input | Required | Default | Gate |
|---|---|---|---|
| Photo | Yes | — | One adult with the face clearly visible. Ask for another photo if the face is tiny, blurred, heavily covered or in profile |
| Image | No | — | the damaged or incomplete photo to restore |
| Description | No | tuned default below | What the user wants; merged into the prompt |
| Ratio | No | `keep source` | keep source (default); 3:4, 1:1 on request |

- If the person may be under 18, stop and decline politely. Real-person photos must be of adults the user has the right to use.
- Fashion, dark and horror themes are fine as stills; refuse explicit sexual content and anything sexualizing minors.

## Steps

1. Read `references/image-edit.md` for prompt structure.
2. **Identity read**: `recognize_image(image_url=<photo>, prompt='Describe this adult for identity preservation: face shape, eye shape and color, brows, nose, lips, skin tone, apparent age range, hair color and length, distinctive marks, glasses. One compact line.')`.
3. **Generate**: `generate_image(prompt=<PROMPT>, input_urls=[photo], model='i2i.hd', orientation='portrait', aspect_ratio='3:4', name='Seamless Image Restore', entity_type='skill_generation')`. Merge the user's details and options into the template.
4. **Self-check, one retry**: look at the result. If a person photo was used and the face no longer reads as the same person, regenerate with the identity clause first. If someone glancing at it would not instantly see the intended change (it reads as a light filter, the old clothes or background remain, or the style is faint), regenerate once with the change described bigger and more concretely.
5. Reply in one or two sentences, e.g. "All set. Your seamless image restore is on the canvas. Next I can do 3:4." Results appear on the canvas; never paste URLs.

## Prompt templates

**PROMPT**

```text
The same adult from the reference: <identity line>. Same face structure, eyes, skin tone and recognizable likeness; no beautification or face reshaping. Intentional changes: the uploaded image with missing damaged or erased regions naturally completed and restored into a single seamless coherent result, matching original style lighting grain and perspective, no visible seams patches or repeating artifacts, faithful to the undamaged content. <user details and options>. Apply the style to the entire frame with its signature traits (medium, texture, palette, linework); the person stays recognizable under the style. Lighting and materials described concretely; composition suited to portrait 3:4. Photoreal unless the style says otherwise, fully clothed, no text unless requested.
```

Quality bar: Name the style's visual traits (medium, brushwork or render, palette, line quality) and apply them to the whole frame; keep the face's likeness readable under the style.

## Failure handling

- `success=false`: say what failed in one sentence; retry once only if transient; never invent links.
- The face changes: regenerate once with the identity clause first; then say so honestly.

## Examples

**Default**: the user says "this one" with a photo. Use ratio `keep source`, 3:4. Reply: "All set. Your seamless image restore is on the canvas."
**Variation**: "next variation". Keep the same inputs and change one option (3:4).
**Missing input**: no photo was sent. Ask: "Drop in one photo of yourself (face clearly visible) to get started."

## References

- `references/image-edit.md`