Restyling · عوامی Open Skill

Restyling

Restyling Digen پر ایک عوامی Open Skill ہے۔ کھولیں اور دیکھیں یہ ورک فلو کیا تخلیق کرتا ہے۔

@wwws

Auto

1 0 2

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-19/3qcfhsxq.jpg

# Restyling

Turn one uploaded photo into a single restyled fashion image in the app's visual language: a deep charcoal backdrop, a warm key light with a glowing rim along the body, a champagne / beige silk palette, and a small golden particle dissolve where the new garment meets the light. The subject's identity, pose and framing stay intact — only the outfit, its material, and the atmosphere change.

This Skill produces **images only** (no video).

## When to use

- The user uploads a photo and asks to change the outfit, fabric, color, or material of what is worn.
- The user uploads a photo and asks for "that app look": dark studio, warm rim light, particle/dissolve effect, poster title.
- The user asks to adjust one element of a previous result (garment, palette, light, caption, crop) without regenerating the person's identity.

If the user only asks for a plain photo edit with no restyling intent (background removal, upscaling, plain cutout), say this Skill is for restyled fashion images and ask what they want changed.

## Inputs and how each is used

| Input | Role in the prompt |
|---|---|
| Reference image (required for the main flow) | The **source of truth**: identity, face, hair, body proportions, pose, camera angle, framing and existing garment coverage. Passed as `input_urls=[<the uploaded image>]`. Never reinterpret the person. |
| User text (optional) | The **restyle instruction**: target garment, fabric, color, mood, caption word, crop. Translated into the editable slots of the prompt template. |
| Nothing but an image | Default to a palette + material restyle: the existing garment is rebuilt in champagne / warm beige silk with a golden particle dissolve on one seam, on the app's dark backdrop. |

Do not invent garments, props, logos or captions the user did not ask for.

## Steps

1. **Read the request and the photo.** If the uploaded image is ambiguous (dark, cropped, several people, unclear garment), call `recognize_image` once on it and describe what is actually visible — garment type, silhouette, pose, light direction, crop.
2. **Read the prompt conventions.** Call `read_skill_file(name="restyling", path="references/style-and-prompt.md")`. This file holds the locked palette, materials, lighting, composition, layout, must-keep and must-avoid rules, plus the fill-in prompt template. Every generation step must follow it.
3. **Decide the restyle scope.** Call `read_skill_file(name="restyling", path="references/workflow-stages.md")` and follow its scope-decision stage: full outfit swap, fabric/material-only restyle, scene relight, poster title, or a combination. Use its input-check and branch rules instead of guessing.
4. **Write the prompt.** If your scope needs one of the prepared blocks, call `read_skill_file(name="restyling", path="references/prompt-templates.md")` and fill the slots for this specific user and this specific photo. The prompt must name the actual garment, actual fabric, actual pose and actual crop from step 1 — never a generic sentence.
5. **Generate.** Call `generate_image` with:
   - `prompt` = the filled English prompt from steps 2–4 (English only).
   - `input_urls` = `[<uploaded reference image>]`.
   - `model='i2i.hd.lite'` (image-to-image). Only if the user truly gave no image at all, use `model='t2i'` and drop `input_urls`.
   - `orientation='portrait'` by default (fashion poster crop); use `landscape` or `square` only when the source photo or the user asks for it.
   - leave `resolution` unset unless the user asks for a specific sharpness; never pass pixel width/height.
   - `name` = short canvas title, e.g. `Restyle – champagne silk`.
   - `entity_type='restyled_images'`.
6. **Check the result** against the QA checklist in `references/workflow-stages.md` (identity, coverage, hands, particle edge, caption spelling, crop). If one check fails, regenerate **once** with a corrected prompt; report the remaining limitation if it still fails.
7. **Reply** in one or two natural sentences: wh