Street Fighter Chun-Li Style · Open Skill 公開技能

Street Fighter Chun-Li Style

Street Fighter Chun-Li Style 是 Digen 上的公開 Open Skill。開啟後即可了解此工作流能製作什麼。

@JackRhino

TryOn

2 0 2

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284865/2026-10-22/n8zmbpu5.jpg

# Street Fighter Chun-Li Style

Reference-driven image restyle. The user's photo is the subject and the composition; the skill rewrites only the styling so the person reads as a Street Fighter Chun-Li costume render. Identity, pose and structure come from the input image, never from imagination.

## When to use

- User uploads a photo (portrait, full body, cosplay, product/mannequin shot) and wants it turned into Chun-Li / Street Fighter style.
- User asks to change the outfit of an already generated result, switch framing, or push sharpness/color quality further.
- User describes the costume ("blue qipao dress, white sash, white boots") and hands over a reference image.

Not for: free-form illustration with no reference image (ask for a photo first), video, or non-costume edits like background replacement only.

## Inputs and outputs

- Input: one reference image (the subject). Everything about the person — face, pose, angle, body proportions, number of people — is inherited from this image.
- Output: one image, portrait orientation by default, high definition, delivered straight to the canvas.

## Steps

1. **Confirm the reference image.** If no image was attached, ask for one and stop. Explain that the skill restyles an existing photo, not text-only art.
2. **(Optional) Inspect the input.** Call `recognize_image` with the reference image and a short prompt such as "describe the person, pose, camera angle, framing, clothing and background in detail" when the photo is dense, has multiple people, or the user wants a specific region preserved. Skip it for obvious single-subject photos.
3. **Read the style contract.** `read_skill_file` on `references/style-and-prompt.md`, then compose the English `prompt` from the templates there, filling the placeholders with facts observed from the reference image.
4. **Read the pre-flight checks.** `read_skill_file` on `references/step-details.md` and run its checklist (single subject, usable framing, content policy, identity anchors, prompt length) before calling the generator.
5. **Generate.** Call `generate_image` with:
   - `prompt`: the composed English prompt from step 3,
   - `input_urls`: `[reference image]`,
   - `model='i2i.hd.lite'` (reference-to-image restyle; the `.hd` family rejects NSFW content),
   - `orientation='portrait'` unless the input is clearly wide and the user wants to keep it,
   - `resolution='1080P'` by default; use `'2K'` only when the user explicitly asks for higher definition or a tight crop needs the extra detail,
   - `name`: short costume-aware title, e.g. `Chun-Li Style - <subject>`,
   - `entity_type='styled_images'`.
   - If and only if the user insists on a text-only result and provides no image at all, fall back to `model='t2i'` with the same prompt minus the reference clause — say clearly that identity will not be preserved.
6. **Check the result.** Apply the quality checklist in `references/step-details.md`. If it fails, retry at most twice, changing one variable per attempt (usually prompt wording, then `resolution`), never re-rolling blindly.
7. **Reply.** One or two sentences: what was restyled, which framing and resolution were used. Do not paste long URLs. When the user is likely to want a next variation, call `set_guidance` in the same turn with concrete options such as a half-body version, an alternate color of the sash, or a sharper 2K render.

## Notes

- Never invent the costume from memory of the game; the color, materials, trims and proportions live in `references/style-and-prompt.md`.
- Hard-coded channel only. Do not pass physical model names to `generate_image`.
- Keep the subject recognizable: same face, same pose, same body, same number of people. Refuse to swap identity or add a second character unless explicitly requested and possible.
- NSFW, gore and fetishized framing: decline politely without calling the generator. The `.hd` image channels do not support adult content.
- If `generate_image` returns `success=false`, report the ac