Space Sketch Vision · Open Skill 公开技能

Space Sketch Vision

was: I bring a real site or room, plus the details I want imagined. now: Black ink sketches appear in the space, with one quiet adult observer taking in the vision. -> Space photo + Description -> the site with planned, missing, or remembered elements. Options: I can use 3:4 by default, or 1:1, 2:3, 3:2, 4:3, 9:16, or 16:9. Add an adult person photo to set the observer’s identity.

@JordanReyes

ArchViz

0 0 1

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284951/2026-10-15/gyp7wczz.webp

# Space Sketch Vision

An empty lot, a bare room or a ruined hall, with black ink sketches floating in the air showing what's missing, remembered or planned, and one quiet adult observer taking it in.

## When to use

- The user wants a photo of a real space overlaid with floating hand-drawn concept sketches of what could be there (architecture vision, renovation idea, memory sketch overlay)
- A variation on a previous result: 1:1

Not for: Video output (use a video skill).

## Inputs and gates

| Input | Required | Default | Gate |
|---|---|---|---|
| Person photo | No | — | Optional; if given, the face is preserved |
| Space photo | No | — | the real site |
| Person photo | No | — | the observer's identity |
| Description | Yes | — | What the user wants; merged into the prompt |
| Ratio | No | `3:4 default` | 3:4 default; 1:1, 2:3, 3:2, 4:3, 9:16, 16:9 |

- If the person may be under 18, stop and decline politely. Real-person photos must be of adults the user has the right to use.
- Fashion, dark and horror themes are fine as stills; refuse explicit sexual content and anything sexualizing minors.

## Steps

1. Read `references/image-edit.md` for prompt structure.
2. Fill the template below from the user's description. Use the tuned defaults for anything they leave out; ask one short question only if the subject itself is unclear.
3. If a person photo is given: `recognize_image(image_url=<photo>, prompt='Describe this adult for identity preservation: face shape, eye shape and color, brows, nose, lips, skin tone, apparent age range, hair color and length, distinctive marks, glasses. One compact line.')` and add the identity clause.
4. With reference images: `generate_image(prompt=<PROMPT>, input_urls=[refs...], model='i2i.hd', orientation='portrait', aspect_ratio='3:4', name='Space Sketch Vision', entity_type='skill_generation')`; without: the same call with `model='t2i.hd'` and no `input_urls`.
5. **Self-check, one retry**: look at the result. If a person photo was used and the face no longer reads as the same person, regenerate with the identity clause first. If someone glancing at it would not instantly see the intended change (it reads as a light filter, the old clothes or background remain, or the style is faint), regenerate once with the change described bigger and more concretely.
6. Reply in one or two sentences, e.g. "Your space sketch vision is ready. Want another take with 1:1?" Results appear on the canvas; never paste URLs.

## Prompt templates

**PROMPT**

```text
<if photo: The same adult from the reference: <identity line>.> Intentional changes: a realistic photo of an unfinished, empty or damaged space with floating black hand-drawn concept sketches overlaid in correct perspective showing <envisioned elements>, and one thoughtful adult observer. <user details and options>. Apply the style to the entire frame with its signature traits (medium, texture, palette, linework); the person stays recognizable under the style. Lighting and materials described concretely; composition suited to portrait 3:4. Photoreal unless the style says otherwise, fully clothed, no text unless requested.
```

Quality bar: Name the style's visual traits (medium, brushwork or render, palette, line quality) and apply them to the whole frame; keep the face's likeness readable under the style.

## Failure handling

- `success=false`: say what failed in one sentence; retry once only if transient; never invent links.
- The face changes: regenerate once with the identity clause first; then say so honestly.

## Examples

**Default**: the user says "here's my selfie, do it". Use ratio `3:4 default`, 3:4. Reply: "Your space sketch vision is ready."
**Variation**: "love it, make another". Keep the same inputs and change one option (1:1).
**Thin request**: the user gives one word. Fill the brief from the tuned defaults and generate; do not interrogate.

## References

- `references/image-edit.md`