Krea2 Smart Edit · публичный навык Open Skill
Krea2 Smart Edit
Krea2 Smart Edit — публичный Open Skill на Digen. Откройте, чтобы увидеть, что создаёт этот процесс.
Product
0 0 2
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-20/9zemmpbe.jpg
# Krea2 Smart Edit
Turn a user's image plus a text instruction into one edited image that keeps the original subject recognizable while everything (color, material, light, finish) is unified into a single Krea2 Smart Edit visual. There is no structured project state: the reference image, the instruction and the generated result live in the conversation and on the canvas.
## When to use
- The user uploads a photo, product shot, artwork or screenshot and asks for it to be edited, graded, restyled, relit, cleaned up, or turned into an ad / art piece.
- The user asks for fine-grained adjustments in a complex image (one region, one material, one color family) while the rest must stay untouched.
- The user describes the "Krea2" / "Krea 2" / "Smart Edit" look and wants any image (even a new one) rendered in it.
- The user only wants to tweak a previous result ("make it warmer", "keep the face but change the background").
## Inputs and outputs
- Input: a reference image (required for editing) + a text instruction (required). Additional text may describe the desired style, crop, or must-keep details.
- Output: exactly one edited image, delivered on the canvas.
- The reference image goes into `input_urls`; the text instruction is expanded into the English prompt built from `references/style-and-prompt.md`.
## Steps
1. **Load the prompt rules.** Before composing any prompt, call `read_skill_file` with `references/style-and-prompt.md`. Use its Style / Subject / Layout / Must keep / Must avoid blocks and its prompt template as the backbone of every prompt. Do not write a generic "make it beautiful" prompt.
2. **Read the intent and the inputs.** Run the intake checklist in `references/step-details.md` (is there a reference image? is the instruction global or local? which details are declared must-keep?). If the reference image needs to be understood before editing (unknown content, several possible subjects, a busy scene), call `recognize_image` on it and note subject, framing, palette and any text/logos to preserve.
3. **Classify the edit.** Read `references/edit-recipes.md` and pick the closest recipe: global color correction, LoRA-style texture stack, subject restyle, background/lighting rebuild, product/ad composition, or micro detail edit. The recipe tells you what to emphasize, what to freeze, and which clauses to paste.
4. **Write the English prompt.** Assemble it in the order given by `references/style-and-prompt.md`: subject from the reference image, the user's edit instruction translated into concrete visual language, the unified Krea2 style block, layout/framing, must-keep list, must-avoid list. Keep it specific and tied to this image; never fall back on filler adjectives.
5. **Generate the image.** Call `generate_image` with the prompt from step 4:
- `model='i2i.hd.lite'` when a reference image is supplied (the normal case for this skill).
- `model='t2i'` only when the user asks for a fresh image in the Krea2 look with no reference image at all.
- `input_urls=[reference image]` (asset id, `s3://` uri, or the canvas reference) for the `i2i.hd.lite` path; omit it for `t2i`.
- `orientation` matched to the reference image / requested use: `portrait`, `landscape` or `square`. Pass `resolution` only when the user asks for extra sharpness or fine detail work.
- Short English `name` for the canvas card, and `entity_type='edited_images'`.
- Never pass a physical model name; only the channels above.
6. **Verify and, if needed, repair once.** Run the verification checklist in `references/step-details.md`: compare the result with the must-keep list (identity, structure, text, product shape, framing). If something required was lost, rewrite the prompt with the must-keep clauses moved to the front plus one explicit "preserve exactly" sentence and regenerate once with the same channel.
7. **Reply.** Describe what changed and what was preserved in one or two sentences of natural language, offer at most a couple of concr