Evangelion Neon Portrait · Open Skill 公開技能
Evangelion Neon Portrait
Evangelion Neon Portrait 是 Digen 上的公開 Open Skill。開啟後即可了解此工作流能製作什麼。
Roleplay
0 0 3
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284865/2026-10-20/78ev9jkm.jpg
# Evangelion Neon Portrait Recreates the "Neon Genesis Evangelion transformation" image app: one uploaded photo goes in, one neon-lit EVA pilot portrait comes out with the person's face and identity intact. There is no structured project state — the only inputs are the user's reference image, their free-form text, and any assets already on the canvas. ## When to use - The user uploads a photo (selfie, portrait, cosplay shot, character art) and asks for the Evangelion / EVA pilot / plugsuit / NERV version of it. - The user wants clothing or background replacement into an EVA pilot scenario while keeping the original face and body structure. - The user wants to redo or adjust an earlier result: suit color, unit number, hangar vs. entry plug, neon palette, framing, expression, framing width. - The user describes the EVA pilot look in text only and asks for the same style output. ## Inputs and outputs - **Accepts:** a reference image (the core input) plus optional free-form text (color wishes, unit number, mood, pose, extra props). The text never replaces the image; it only steers it. - **Outputs:** one generated portrait image plus a short natural-language summary of what changed. Do not paste URLs. ## Steps 1. **Read the request.** Decide which branch applies: (a) photo supplied, (b) photo supplied plus text direction, (c) text only, (d) revision of a previous result. If no usable image and no clear subject is given, ask one short question for the photo before generating. 2. **Read the workflow reference.** `read_skill_file` → `references/workflow-stages.md`. Follow its intake checks, branches, retry limits, and close-out rules; do not improvise the branch logic. 3. **Inspect the reference image (optional but recommended).** If the photo is available as an asset and its content is unclear (group shot, heavy occlusion, unusual angle, illustrated character), call `recognize_image(image_url=<asset reference>, prompt="Describe the main subject's identity-defining features: face shape, hairstyle and hair color, skin tone, glasses, facial hair, distinguishing marks, build, current clothing, and camera angle.")`. Use the answer only to protect identity details; never invent features the recognizer did not confirm. 4. **Read the style reference.** `read_skill_file` → `references/style-and-prompt.md`. It holds the binding Style / Subject / Layout / Must keep / Must avoid rules for this app plus the base English prompt skeleton. Every generation in this skill must follow it. 5. **Read the prompt templates.** `read_skill_file` → `references/prompt-templates.md`. Assemble one concrete English prompt from its fragment library: identity block from the photo, plugsuit block, NERV hangar block, giant EVA block, neon lighting block, framing block, plus any user text translated into English. 6. **Generate the image (image-to-image, primary path).** Call: `generate_image(prompt=<assembled English prompt>, input_urls=[<user reference image asset>], model='i2i.hd.lite', orientation='portrait', resolution='4K', name='EVA Pilot Portrait', entity_type='characters')` - `model='i2i.hd.lite'` is the channel for image-to-image; never pass a physical model name. - `orientation='portrait'` carries the app's vertical frame. Only if the user explicitly insists on the exact 3:4 ratio may you additionally pass `aspect_ratio='3:4'`; otherwise let `orientation` do the work. - Keep `input_urls` to the single clearest reference (add a second only when the user supplied a separate pose or suit reference). 7. **Generate the image (text-only fallback).** Only when the user has no photo and still wants the look: `generate_image(prompt=<assembled English prompt with the subject described from their text>, model='t2i', orientation='portrait', resolution='4K', name='EVA Pilot Portrait', entity_type='characters')`. State clearly in your reply that identity fidelity is limited without a photo. Ask once for a photo first. 8. **Review the result against the