Swimsuit Candid Snapshot · publiczny skill Open Skill
Swimsuit Candid Snapshot
Swimsuit Candid Snapshot to publiczny Open Skill w Digen. Otwórz go, by zobaczyć, co tworzy ten workflow.
Doc
1 3 4
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284862/2026-10-16/cjtsrxmb.jpg
# Swimsuit Candid Snapshot
Turn a photo the user uploads into a candid summer swimsuit snapshot with the look of 1950s American vintage film. This is an image-to-image restyle: the person and the identity in the picture come from the user's photo, while the wardrobe, the scene and the film look come from this skill.
There is no structured project state. Everything the skill needs arrives in the conversation as one reference image plus optional text, and every produced image is placed onto the canvas automatically — the final reply only needs a short natural-language summary.
## When to use
- The user uploads or pastes a photo and asks for a swimsuit snapshot, a beach/holiday photo, a candid summer snapshot, or a vintage film-photo version of that picture.
- The user wants an existing snapshot adjusted (another location, another swimsuit, a different candid moment, warmer or more faded film look) without changing the person.
- The user asks for the same treatment on a pet, a car or an object photo — the same rules apply to that subject.
- Do not use for a plain text-to-image scene with no reference photo unless the user explicitly insists after being asked for one (see `references/step-details.md`, "Edge cases").
## Steps
1. **Collect the inputs.** Two things are needed: (a) a reference photo — an uploaded image asset or an image URL — and (b) whatever text the user wrote (location, swimsuit style or colour, pose, mood, how many snapshots). If there is no photo, ask for one and stop; do not invent a person. If the photo is there but the text is thin, do **not** interrogate the user: pick sensible defaults from `references/style-and-prompt.md` and generate.
2. **Look at the reference photo before generating.** Call `recognize_image` with the photo as `image_url` and a short prompt asking for the number of people, the identity features worth preserving (hair, glasses, freckles, moles, build), the existing pose and framing, the background and the light direction. Skip this step only when the photo is an obvious single clear portrait and the user has already described it. Keep what you learn as short notes for the prompt.
3. **Read the visual contract.** Call `read_skill_file('swimsuit-candid-snapshot', 'references/style-and-prompt.md')` before the first generation of a conversation — it holds the film look, the styling rules, the Must-keep / Must-avoid lists and the prompt skeleton. Read `references/prompt-templates.md` when you want a ready-filled template for a common case, and `references/step-details.md` for the stage checklists, retry rules and edge cases. Do not read all of `references/` at once if the task only needs one of them.
4. **Compose the prompt in English.** Use the skeleton from `references/style-and-prompt.md` and fill the slots: subject (from the photo), period swimwear, candid action, location, time of day and light. Text the user wrote in another language is translated into the prompt — never pasted as-is. The user's words map to specific slots (location words → location slot, colour words → swimsuit slot, "black and white" → monochrome film phrase); the mapping table is in that reference file.
5. **Generate with `generate_image`**, one call per snapshot:
- `prompt=<the composed English prompt>`
- `input_urls=[<reference photo asset_id or s3:// uri>]` — the reference is the identity and composition anchor
- `model='i2i.hd.lite'`
- `orientation='portrait'` by default (the candid snapshot reads best vertically); use `landscape` when the user's photo is landscape, the scene is wide, or several people are in frame; `square` only if the user asks for a retro print feel
- `resolution='1080P'` by default for a finished snapshot; no need to go higher
- `name='Swimsuit candid — <location>'` (short canvas title)
- `entity_type='characters'`
For several snapshots, make several `generate_image` calls with a different candid moment each — never try to get multiple images out of one call.