Snow Sculpture Object · skill publik Open Skill
Snow Sculpture Object
Snow Sculpture Object
Manhua
0 0 0
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4150474/2026-10-15/x2t2bmmv.jpg
# Snow Sculpture Object
Turn the object specified by the user into a **pure white snow sculpture** that lies on a snow field and is photographed straight down from directly above, producing one portrait 3:4 image. There is no structured state; one image is generated per run, and the result is added to the canvas automatically.
## When to use
- The user names an object (a heart, a vintage camera, a teapot, a sneaker, headphones, a building, a letter, …) and wants it "made of snow / a snow sculpture / a snow version".
- The user provides a photo of an object and wants a snow sculpture matching the shape in that photo.
- The user is unhappy with the previous result and asks to change the object, change the angle, make the volume fuller, make the shadow more realistic, etc.
## Steps
1. **Determine the object and its structural features**
- Extract from the user's text / reference image: what the object is + 2–4 recognizable structural features (e.g. camera = square body, protruding lens, shutter button; teapot = body, handle, spout).
- If the object is unclear, ask at most one question: "Which object would you like as a snow sculpture?" Do not silently substitute a more ordinary object.
2. **Decide the generation mode**
- User has **no** reference image → text-to-image: `model='t2i'`, do not pass `input_urls`.
- User **has** a reference image (an `asset_id` generated earlier in this conversation, an `s3://` URI, or an https image) → image-to-image: `model='i2i'`, put that image in `input_urls` (the first image is authoritative). The reference image is used **only to lock the shape, proportions and structure**; its colors and materials are always replaced by snow, so do not carry over the original palette. For higher fidelity you may use `model='i2i.hd'` (`.hd*` does not support NSFW, which is irrelevant for this use case).
3. **Assemble the English prompt** (neither the reference image nor the user's language changes the prompt language)
Template without a reference image (`t2i`):
```
{object}, reimagined as a sculpture carved entirely from pure white snow:
its recognizable silhouette and key structural features preserved
({structure features}); full, voluminous, rounded mass with softly
rounded edges and finely packed snow grain. The sculpture lies slightly
sunken into a smooth field of fresh white snow, photographed from
directly above in a perfectly top-down bird's-eye view; the whole
outline of the sculpture and the surrounding snow contact surface stay
inside the frame. Soft diffuse winter daylight, crisp specular
highlights on the snow, realistic soft contact shadow and ambient
occlusion where the sculpture meets the snow. Monochrome pure white;
the object's original colors and materials are fully replaced by snow.
No text, no watermark, no people.
```
With a reference image (`i2i`), prepend this sentence to the template:
```
Use image 1 only for geometry: follow the silhouette, proportions and
structural details of the object in image 1.
```
- Append any extra requirements from the user (tinted snow, footprints in the background, a touch of warm light, etc.) to the end of the template, but **never** restore the object's original colors/materials — replacing them is the whole point.
- Do not let text, labels, watermarks, human hands or a second main subject appear in the image.
4. **Call `generate_image`** with these fixed parameters:
- `prompt=<the English prompt from the previous step>`
- `input_urls=[reference image]` (`i2i` only; `asset_id` / `s3://` / https are all fine, the platform presigns them automatically)
- `model='t2i'` or `model='i2i'` (write the channel, never a physical model name)
- `orientation='portrait'` plus `aspect_ratio='3:4'` (the fixed portrait canvas of this skill; `orientation` only covers 16:9/9:16/1:1, so the non-standard 3:4 ratio must be given explicitly via `aspect_ratio`)
- `name=<short