Halloween Street Party Portrait · skill publik Open Skill

Halloween Street Party Portrait

Halloween Street Party Portrait adalah Open Skill publik di Digen. Buka untuk melihat apa yang dibuat alur ini.

@eggs

Manhua

1 2 2

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284862/2026-10-16/j4kgyh6b.jpg

# Halloween Street Party Portrait

Bottom line: one uploaded character image goes in, one Halloween street-party portrait photo comes out. The person's **identity stays** (same face, same person); only the **costume, pose and scene** are re-dressed for a Halloween street party. This is a raw image-generation app — the only deliverable is a single still portrait image.

## When to use

- The user uploads a picture and asks for a Halloween version / Halloween costume / Halloween street party photo of that person.
- The user asks for a **stronger** ("wilder", "crazier", "more intense", "chaos") or **milder** ("more natural", "toned down") Halloween look — that is the app's chaos switch, see `references/style-and-prompt.md`.
- The user wants an already-produced portrait adjusted (costume, props, scene, lighting) while keeping the same person.

Not for: text-only image requests with no reference image, video output, collages or multi-person group scenes.

## Input and output

| | |
|---|---|
| Input | Reference image(s) uploaded by the user. The **first** image is the identity anchor. Extra images may only inform costume/scene, never override the face. |
| Output | One image asset (a Halloween street-party portrait photo), added to the canvas. |
| Missing input | If there is no reference image, ask the user to upload one. Never invent a person or fall back to text-to-image. |

## Steps

1. **Read the reference image.** Call `recognize_image(image_url=<the uploaded reference image>, prompt="Describe the person's apparent gender, age range, face shape, hairstyle and hair color, eye color, skin tone, ethnicity, distinctive features (moles, freckles, glasses, tattoos), and what she is currently wearing.")`. Write down these **identity anchors** — they are the one thing that must survive into the output.
2. **Set the intensity (chaos switch).** Default is **chaos off**. Treat "make it wilder / crazier / more intense / maximum Halloween / 更夸张" as **chaos on**; "more natural / subtle / toned down" as **chaos off**. If the user asks for a variant of a result you already produced, flip or keep the level they refer to. Do not ask the user to phrase it as "chaos" — map their words to the two levels yourself.
3. **Load the visual conventions.** Call `read_skill_file(name='halloween-street-party-portrait', path='references/style-and-prompt.md')` and follow it for style, subject matter, layout, must-keep, must-avoid and the prompt templates. Do not improvise the visual direction from memory.
4. **Build the prompt** with the template from that reference file, filling the slots: identity anchors from step 1, the costume/props/scene you chose, the intensity level from step 2. Prompts are written in **English**.
5. **Generate the image** with the reference image attached so the identity is preserved:

   ```
   generate_image(
     prompt=<filled template>,
     input_urls=[<the uploaded reference image: its asset_id or s3:// uri>],
     model='i2i.hd.lite',
     orientation='portrait',
     name='Halloween street party portrait',
     entity_type='portraits'
   )
   ```

   - `model='i2i.hd.lite'` is the image-to-image channel for this app. Never write a physical model name, and never switch to `t2i` — without the reference image the person's identity is lost.
   - `orientation='portrait'` (9:16) is the app default: it is a portrait photo of one standing person on a street. Use `square` only if the user explicitly wants a social-post crop; `landscape` only if they ask for a wide party-scene shot.
   - Clarity: leave `resolution` unset (platform default `720P`). Pass `resolution='1080P'` (or `'2K'`) only when the user asks for a higher-detail / print-quality version.
   - `name`: a short canvas title, e.g. `Halloween street party portrait (chaos on)`. `entity_type='portraits'` groups these results on the canvas.
6. **Reply in natural language**: say what was produced (style, costume, scene, intensity level) in one or two sentences. Do **not** paste