Halloween Pet Costume · Open Skill 公開技能

Halloween Pet Costume

Halloween Pet Costume 是 Digen 上的公開 Open Skill。開啟後即可了解此工作流能製作什麼。

@eggs

Manhua

0 1 0

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284862/2026-10-16/x8p65sk7.jpg

# Halloween Pet Costume

Turns an uploaded pet photo into an image of that same pet wearing a Halloween costume. One input image plus optional text, one output image. There is no structured project state: anything that must survive between turns lives in the chat text and in the canvas assets already produced.

## When to use

- The user uploads (or points at) a pet photo and asks for a Halloween, costume, dress-up, or trick-or-treat version of it.
- The user asks to change the costume, palette, or scene of a Halloween pet-costume image this skill already produced.
- Not for: creating a pet with no reference photo (that is plain text-to-image, not this skill), restyling photos of people, or non-pet animals treated as something other than a companion animal.

## Inputs and output

| Input | How it is used |
|---|---|
| Reference image (required) | Passed as the first item of `input_urls` to `generate_image`. It is the identity anchor: the prompt must be written as "the same {species/breed} with {fur colors, markings, eyes, ears} as the reference photo", and the costume is drawn *onto* that animal rather than replacing it. |
| Text (optional) | Mapped into prompt slots by `references/prompt-templates.md`: costume, palette, scene, must-keep and must-avoid clauses. Unspecified slots fall back to the defaults in `references/style-and-prompt.md`. |

Output: one generated image per call (`batch_size` is locked to 1), added to the canvas as an `characters` asset.

## Steps

1. **Check the reference.** If the user did not attach an image and there is no pet image on the canvas, ask for a pet photo before doing anything else. If the message has no pet in it (person only, empty room, screenshot), explain what this skill needs instead of generating.
2. **Read the conventions before designing anything.** Call `read_skill_file(name="halloween-pet-costume", path="references/style-and-prompt.md")` and `read_skill_file(name="halloween-pet-costume", path="references/workflow-stages.md")`. Do not generate from memory of this file alone — the style file holds the locked palette, lighting, framing, must-keep and must-avoid rules.
3. **Optional identity pass.** When the reference is busy (several animals, heavy motion blur) or when the user asks for a faithful likeness, call `recognize_image(image_url=<reference>, prompt="Describe the pet's species, breed traits, fur colors and markings, eye color, ear and muzzle shape, collar or tag, pose, and camera angle.")` and carry those details into the prompt's identity clause.
4. **Design the shot.** Pick the costume and scene from the catalogs in `references/style-and-prompt.md` (user text wins; otherwise use a theme default and vary it between turns). Fill the template in `references/prompt-templates.md` and read `references/workflow-stages.md` for the pre-generation checklist and branch rules (multiple pets, human in frame, offline/refused requests).
5. **Generate.**
   `generate_image(prompt=<filled template>, input_urls=[<reference image>], model='i2i.hd.lite', name="<Pet name> — <Costume>", entity_type='characters', orientation=<see below>, resolution='1080P')`
   - `model='i2i.hd.lite'` is the image-to-image channel for this app: the reference photo is the input, so always pass `input_urls` with it. Never write a physical model name.
   - `orientation`: default `square` for centered pet portraits and headshots; use `portrait` when the uploaded photo is a vertical full-body shot or the costume silhouette is tall (witch hat, cape, wings); use `landscape` only when the user explicitly asks for a wide scene shot. Keep it as close to the source photo's framing as possible.
   - `resolution`: `1080P` by default so fur and fabric detail survive; drop to `720P` only when the user asks for a fast draft.
   - One call produces one image. For several costumes or several pets, run one call per image and update `write_todos` as you go.
6. **Self-check the result** against the checklist in `references/workflow-st