AI Cloth Transfer · skill publik Open Skill

AI Cloth Transfer

AI Cloth Transfer adalah Open Skill publik di Digen. Buka untuk melihat apa yang dibuat alur ini.

@wwws

TryOn

2 0 2

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-20/dfxk9a8k.jpg

# AI Cloth Transfer

Edits an uploaded photo so the person in it wears a different outfit. The subject's face, hair, pose, expression, body proportions, camera angle and background stay intact; only the clothing changes. This Skill has no structured project state — the inputs are the uploaded assets and what the user says in the conversation. Every output is produced by a single `generate_image` call on the image-to-image channel.

## When to use

- The user uploads a person photo and asks to change, replace, swap or "try on" clothes.
- The user uploads a person photo plus a garment photo (product shot, screenshot, another outfit) and wants the garment transferred onto the person.
- The user uploads a person photo and describes an outfit in words only.
- The user asks for another variant of a transfer already produced in this conversation (different color, different garment, different style).

## Inputs and output

| Input | Required | How it is used |
|---|---|---|
| Person photo (subject) | Yes | First entry of `input_urls`; the identity, pose, camera angle and background all come from this image. |
| Garment reference image | No | Second entry of `input_urls`; the new clothing is read from it (silhouette, fabric, color, hardware, accessories that belong to the outfit). |
| Outfit description in text | No | Written into the prompt as an explicit garment specification when no garment image exists. |

Output: exactly one edited image per call, delivered as a canvas asset. If the user asks for several outfits or variants, run one call per variant.

## Steps

1. **Confirm the inputs.** Identify the person photo and, if present, the garment reference. When it is unclear which uploaded asset is the person and which is the outfit, look at them first with `recognize_image` (short factual prompt, e.g. "describe the person, pose, framing and background" / "describe this garment's cut, fabric, color and hardware") and then decide. Ask at most one short clarifying question if a person photo is genuinely missing; otherwise proceed with what is available.

2. **Read the step playbook.** Call `read_skill_file` for `references/step-details.md` and follow it for input validation, the no-garment-image branch, parameter selection and post-generation quality checks.

3. **Read the visual and prompt rules.** Call `read_skill_file` for `references/style-and-prompt.md` and assemble the English prompt for this specific request using its templates and prompt-writing rules. The prompt must be built from the current inputs and the user's requested outfit — never a generic placeholder sentence.

4. **Generate the transfer.** Call `generate_image` with:
   - `model='i2i.hd.lite'` (image-to-image; this Skill never uses text-to-image or video channels).
   - `prompt`: the English prompt assembled in step 3.
   - `input_urls`: person photo first, garment reference second (omit the second entry when the outfit comes from text only).
   - `orientation`: `portrait` by default; use `landscape` or `square` only when the source photo is clearly that shape, so framing is preserved.
   - `resolution`: `1080P` (garment texture and hardware detail matter); use `720P` only when the source photo is low resolution.
   - `name`: a short label such as `Cloth transfer - black leather set`.
   - `entity_type`: `characters`.
   Do not pass `aspect_ratio`, `batch_size` or any model name.

5. **Check the result against the must-keep list.** Apply the quality checklist in `references/step-details.md` (identity, pose, background, garment fidelity). Retry at most once with the fix described there; do not loop.

6. **Reply briefly.** One or two sentences describing the outfit that was applied and what stayed unchanged. Do not paste URLs. If the user seems to want further variants, call `set_guidance` in the same turn (e.g. a different garment, a different color, or reusing the same outfit on another photo). Use `write_todos` only when the user asks for three or more separate tran