Silhouette Art Transform · skill publik Open Skill

Silhouette Art Transform

Silhouette Art Transform adalah Open Skill publik di Digen. Buka untuk melihat apa yang dibuat alur ini.

@wwws

Other

0 0 1

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

# Silhouette Art Transform

Turn one uploaded photograph into a single fine-art image in which the subject is fused into one dark silhouette mass while a few precious areas (face, hands, a rim of hair) stay lit, painted on a heavily textured old-master backdrop of teal and ochre.

There is no project state. Every conversion is one `generate_image` call plus a short natural-language reply.

## When to use

- The user uploads a picture and asks for it to be turned into "silhouette art", "silhouette effect", "shadow portrait", "mostly black figure with a lit face", "partial silhouette", "chiaroscuro portrait", or similar.
- The user uploads a picture and asks to restyle it in the look of this skill (teal/ochre painterly background, dark body, lit face).
- The user asks to redo, tweak, add a title text, change framing, or make variants of a silhouette conversion produced earlier in the conversation.

Do not use it for plain photo restoration, background removal, flat vector icons, or a full black silhouette with no visible features.

## Reference files

| File | What it contains | When to read |
|---|---|---|
| `references/style-and-prompt.md` | The visual rules (style, subject, layout, must-keep, must-avoid) and the copy-ready English prompt template and fragments | **Always, before any `generate_image` call** |
| `references/step-details.md` | Input triage, orientation/resolution decisions, prompt assembly checklist, quality gate, text-handling rules, retry and revision branches, failure handling | Every request: read once, then follow the branch that matches the current input |

Read them with `read_skill_file(name='silhouette-art-transform', path='references/...')`. Never generate before `style-and-prompt.md` has been read in the current turn.

## Steps

1. **Read `references/style-and-prompt.md`** and `references/step-details.md`. Do not call `generate_image` before this step is done.
2. **Inspect the input.** Look at the uploaded image (call `recognize_image` on it when you need a closer read of pose, garments, hands, lighting, or when the image is not clearly visible to you). Run the "Input triage" branch in `step-details.md`: confirm there is exactly one primary reference image, note the pose, the lit-feature candidates, and whether the user supplied text.
3. **Draft the prompt.** Fill in the template from `style-and-prompt.md` with the concrete subject, pose, garments, hairstyle, hands, and lighting direction observed in the reference. Follow the prompt assembly checklist in `step-details.md`. The prompt must be in English and must explicitly state which areas stay lit and which become the shadow mass.
4. **Call `generate_image`**:
   - `prompt`: the assembled English prompt.
   - `input_urls`: `[the uploaded reference image]` (single image; this is an image-to-image restyle).
   - `model='i2i.hd.lite'`.
   - `orientation` and `resolution`: take from the decision table in `step-details.md`; default is `orientation='portrait'`, `resolution='2K'`.
   - `name`: a short descriptive canvas title (e.g. `Silhouette — woman in black dress`).
   - `entity_type='silhouette_art'`.
5. **Self-check the result** with the quality gate in `step-details.md` (identity preserved, face readable, shadow mass dominating, backdrop texture present, no stray text). If it fails and one more attempt is justified, regenerate once with a corrected prompt; otherwise show what you have and say what is off.
6. **Reply in natural language**: one or two sentences on what was kept and what was converted into shadow, plus an offer of concrete next moves. Call `set_guidance` in the same turn with 2–4 options (e.g. darker silhouette, more lit face, square crop, add a title). Do not paste long URLs.

## Notes

- Only one reference image is used per conversion; if several are uploaded, use the primary subject image and tell the user which one you used.
- Never invent identity, clothing, jewelry, or props that are not visible in the reference. Never add text unl