Krea Image Enhance Edit · مهارة Open Skill عامة

Krea Image Enhance Edit

Krea Image Enhance Edit هي Open Skill عامة على Digen. افتحها لتعرف ماذا ينشئ هذا التدفق.

@wwws

ArchViz

11 1 11

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-19/4v3sw7u6.jpg

---

allowed-tools:
  - recognize_image
  - generate_image
---

# Krea Image Enhance Edit

A precision image-editing skill modelled on the Krea-2-turbo editing app: it takes the user's uploaded image as the source of truth and returns one re-rendered, improved version of that same image. It is an edit/enhance pipeline, not a free generation pipeline — the subject, pose, structure and framing of the upload must survive.

## When to use

- The user uploads (or references on the canvas) an image and asks to improve it, make it look better / higher quality / cleaner.
- The user wants a style or look change applied to their own photo ("make it cinematic", "unify the colors", "studio lighting").
- The user wants the background removed, replaced, blurred or cleaned up.
- The user wants an element added, removed or adjusted inside their photo.
- The user asks for a variant of a previous edit in the same conversation.

Do not use it for pure text-to-image requests with no source image; in that case ask for the image first.

## Steps

1. **Locate the source image.** Use the asset the user just uploaded (or the asset id / s3 uri already on the canvas). Never guess or invent a URL. If several images exist and the target is ambiguous, ask which one before generating.
2. **Read the request precisely.** Write down the edit intent in one line: what must change and what must not. If the user's wording is vague (e.g. "make it better", "fix it"), call `recognize_image` on the source with a short prompt to inspect subject, background, lighting and visible text, then state your interpretation back to the user in one sentence.
3. **Load the conventions.** Before building any prompt, call `read_skill_file` for:
   - `references/style-and-prompt.md` — the style lock, subject handling, layout rules, must-keep / must-avoid lists and the prompt template.
   - `references/edit-operations.md` — the per-intent branch (enhance, restyle, background, add/remove element) with its fragment set, checklist and retry rule.
   Read both; do not generate from memory of this page alone.
4. **Plan multi-part work.** If the request contains more than one edit (e.g. remove background and add a rim light), call `write_todos` with one item per edit, keep exactly one `in_progress`.
5. **Write the English prompt.** Fill the template from `references/style-and-prompt.md` and the matching branch from `references/edit-operations.md`: source description, the single global style lock, the requested change, the preservation clause, and the avoid clause. The prompt describes the *edited result*, not the editing procedure.
6. **Generate.** Call `generate_image` with:
   - `prompt`: the assembled English prompt.
   - `input_urls`: `[source asset id or s3 uri]`.
   - `model='i2i.hd.lite'` (image-to-image editing channel; always this channel for this skill).
   - `orientation`: match the source image — `portrait`, `landscape` or `square`. Do not change the framing unless the user asks for a crop.
   - `resolution`: `1080P` for close-up portraits or when the user asks for high quality; otherwise omit and let the platform default apply.
   - `name`: a short title describing the edit, e.g. `Enhanced portrait`, `Background removed`.
   - `entity_type='edited_images'`.
7. **Check the result.** Use the checklist in `references/edit-operations.md` (identity, structure, requested change, no invented content, no text artifacts). If the result drifts from the source or the request, retry once with the same source image and the corrected clause — do not silently accept a wrong edit.
8. **Reply and guide.** Summarise in natural language what changed and what was preserved; do not paste long URLs. If useful, call `set_guidance` in the same turn with 2–4 next-step options (e.g. another style, background variant, stronger cleanup).

## Notes

- One upload = one edited output per call. For "give me 3 options", run separate calls with distinct style locks instead of one vague prompt.
- Never invent co