Watercolor Conversion · সর্বজনীন Open Skill

Watercolor Conversion

Watercolor Conversion

@ExploreFly

Other

0 0 0

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3656543/2026-10-14/a3k27z8s.jpg

# Watercolor Conversion

Turn the user's **uploaded reference image** into a watercolor rendering using **image-to-image** generation. Composition is locked: same crop, aspect ratio, camera angle, and placement of every subject. Only the **painting medium and stylistic treatment** change.

## When to use

- User uploads or points to an image and asks for watercolor / aquarelle / painted look.
- User names a watercolor sub-style (botanical, urban sketch, portrait wash, etc.).
- User wants another variant or higher fidelity while still matching the original layout.

## Style reference (read on demand)

Detailed style definitions, prompt tokens, and selection tables live in:

| File | Read when |
|---|---|
| `references/watercolor-styles.md` | Choosing or explaining a watercolor type; user asks "what styles exist"; first conversion of a complex photo |

Call `read_skill_file(name='watercolor-conversion', path='references/watercolor-styles.md')` once per task unless the user only repeats the same style ID.

## i2i model channels (required)

Always use **`generate_image`** with the user's image in **`input_urls`** (first URL = structure reference). Set **`model`** to exactly one of:

| Channel | When to use |
|---|---|
| `i2i.hd.lite` | Fast preview, simple scenes, user asks for quick draft |
| `i2i.hd` | **Default** balanced quality for most conversions |
| `i2i.hd.gpt` | Portraits, fine detail, busy scenes, small text, or user asks for best quality |

Do **not** use `t2i*` channels for this skill. Do **not** pass physical model names.

If the first result drifts layout, retry once with the **same** `input_urls` and a stricter prompt (see below), stepping up channel (`lite` → `hd` → `i2i.hd.gpt`) if needed.

## Workflow

1. **Confirm input** — Use the user's uploaded image URL / asset_id from the conversation or canvas. If missing, ask for one image before generating.

2. **Optional analysis** — For ambiguous requests, call `recognize_image` on the same input to detect subject type (portrait, landscape, product, etc.) and map to a style ID via `references/watercolor-styles.md` (Quick selection table).

3. **Pick style** — Default style ID: `classic-transparent`. Override when the user names a look or analysis suggests a better match (document the choice briefly in the reply).

4. **Match dimensions** — Preserve the source aspect ratio:
   - Infer orientation from the reference: wide → `orientation='landscape'`; tall → `orientation='portrait'`; roughly equal → `orientation='square'`.
   - If the user explicitly requires a non-standard ratio (e.g. 3:4, 4:3), pass `aspect_ratio='W:H'` **instead of** guessing wrong from orientation alone.
   - Do **not** pass arbitrary pixel sizes. Omit `resolution` unless the user asks for higher output (then `720P` default → `1080P` or `2K` as requested).

5. **Generate** — Single output per request unless the user asks for multiple variants:

```text
generate_image(
  model='<i2i.hd.lite | i2i.hd | i2i.hd.gpt>',
  input_urls=[<user_reference_image>],
  orientation='<landscape|portrait|square>' OR aspect_ratio='<W:H>',
  prompt=<English prompt from template>,
  name='<short title, e.g. Watercolor – Garden>',
  entity_type='skill_generation'
)
```

6. **Reply** — Summarize style chosen, model tier, and that composition matches the upload. Do not paste long URLs; canvas shows the asset.

For multi-image batches, use `write_todos` and process sequentially with the same rules per image.

## Prompt template (English, layout-locked)

Build one string. Insert `[STYLE_ID]` tokens from the reference file's **Prompt tokens** for the chosen style.

**Core (always include):**

> Image-to-image watercolor transformation of the reference photo. **Preserve the exact composition, crop, aspect ratio, camera angle, and perspective.** Keep every person, object, and background element in the **same position, scale, and pose** as the source. Do not pan, zoom, reframe, rotate, or add/remove objects. Apply only watercolor