Ghosted Double Exposure Portrait · publiczny skill Open Skill
Ghosted Double Exposure Portrait
Ghosted Double Exposure Portrait to publiczny Open Skill w Digen. Otwórz go, by zobaczyć, co tworzy ten workflow.
Other
1 0 0
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4150474/2026-10-15/saruhwg3.jpg
# Ghosted Double Exposure Portrait
Produces one cinematic vertical portrait: a single adult subject in cool teal/cyan studio light, with a semi-transparent ghost echo of the *same* face floating behind and above the main head, dissolving into haze. There is no project state — everything happens in one image generation call.
**Before writing any prompt, read `references/ghosted-style.md` with `read_skill_file`. It is the authoritative style spec (layout, lighting, palette, ghost layer, prompt template, negative list). Do not improvise the look from memory.**
## When to use
- User asks for a "ghosted", "double exposure", "spirit echo", or "cinematic blue portrait" image.
- User uploads a portrait photo and wants that same person turned into this style (identity kept, clothing/pose/background rebuilt).
- User wants a variation of a portrait already generated in this conversation (new jacket, new mood, female subject, tighter crop, different theme) while keeping the ghosted look.
## Inputs
| Input | Required | Notes |
|---|---|---|
| Text | Yes | Who the subject is plus the theme/mood/wardrobe/background direction. If the user only says "make me one of these", infer a neutral cinematic setup from the reference spec. |
| Reference photo | No | The person to keep. Used as the `i2i` input. Only identity is carried over — clothing, pose, framing, background and lighting are rebuilt. |
## Output contract
- One image, vertical **3:4**, chest-up / half-body framing, centered subject.
- **Text-free**: no titles, captions, letters, logos, watermarks, borders, arrows, or side-by-side "before/after" insets. Keep the lower third as clean dark negative space for later typography.
- Mood: cinematic still, cool teal-and-cyan, directional light, one ghost only.
## Steps
1. **Parse the request.** Fix four things: subject (who, age range, gender presentation), theme/mood (e.g. "arctic calm", "psychological thriller"), wardrobe, and any background variation. If a reference photo is attached, treat it as the identity source.
2. **Read the identity.** If a reference photo is attached and you need concrete traits for the prompt, call `recognize_image(image_url=<reference asset>, prompt="Describe this adult person's face, hair, facial hair, eye color, skin tone and build in one short factual paragraph. No speculation.")` — or describe the traits directly if they are already evident. Never invent facial features.
3. **Read the style spec.** `read_skill_file(name='ghosted-double-exposure', path='references/ghosted-style.md')`, then fill its copy-ready prompt template. Every prompt must contain all seven blocks described there: subject + identity anchors, wardrobe, main-layer pose/gaze, ghost layer, lighting, backdrop/atmosphere, grade + negative list.
4. **Call `generate_image`** exactly once for the first result:
- With a reference photo (image-to-image):
`generate_image(prompt=<full style prompt>, input_urls=[<reference asset_id or s3 uri>], model='i2i', orientation='portrait', aspect_ratio='3:4', name='<short title>', entity_type='portraits')`
- Text only (no reference):
`generate_image(prompt=<full style prompt>, model='t2i', orientation='portrait', aspect_ratio='3:4', name='<short title>', entity_type='portraits')`
- `aspect_ratio='3:4'` is this style's canonical frame and overrides `orientation`; never pass pixel `width`/`height`.
- `name` is a short human title (e.g. "Ghosted Portrait — Arctic Blue"). Do not pass `group_id` or `index`.
5. **Reply.** One or two natural sentences: what was made and the one choice that matters (ghost layer, light direction, or wardrobe). Do not paste URLs or describe the whole prompt back.
6. **Follow-ups (optional).** If useful, in the same final turn call `set_guidance` with 2–4 concrete variations tied to this image (e.g. "Same man, black leather jacket", "Female version, colder palette", "Push the ghost further back"). Skip it if the user only asked for one image and is done.
## Pr