360 Secure SkillHub
360 安全技能中心

Phantra AI: Look Matching

从用户自有的角色、道具、场景照片和参考图中提炼光线、质感、构图与色调,写成场景风格描述,并标明每张照片是作为主体还是只借光借色;同时说明可匹配与不可匹配的部分,拒绝搜寻或复制真实人物面孔。

分类:创作
zeropage-onryo
zeropage-onryo
热度 1.62
最近更新 2026-11-03 02:00:19

安装

发给 Agent
帮我安装这个技能:npx secure-skills-cli add zeropage-onryo/phantra-ai-look-matching
兼容性
通用兼容
联网要求
需要联网
数据出网
必须出网
Metadata
name
look-matching
description
Sets the visual look of a Phantra scene from the person's elements and from an image or frame they describe or attach, says what each photo is FOR (its role), and explains what the studio can and cannot match. Use when someone says "make it look like this", attaches a reference image, asks for a mood, light or texture, or wants a photo used only for its light or palette. NOT for writing the whole scene (write-scene), prices (credits-and-limits), or finding or copying a real person's face.

Look matching

Phantra renders a scene against the person's own elements (photos of characters, props, places) (characters, props, places), never against a name or an image somebody else chose for them. A look is written into the scene's style block in cinematography terms; the photos decide identity, product and place.

The rule on people

Never look for, describe for copying, or ask for a specific real person's face. If the scene needs a face, it is the person's own or someone they have the right to use, provided by them as an element. Say this plainly when asked for a celebrity, a public figure or "someone who looks like X", and do not add a picture of one with add_element or ask for one through upload_link.

Working with the person's photos

  • elements lists their characters, props and places with photo refs and a short description. Read it before writing a look: a room they photographed is a room the scene can be set in; a product they photographed is the exact product on screen.
  • If the look needs a photo they do not have (a product, a location, a wardrobe), say which and ask them for it. Besides the studio's own Elements page, two ways in from the chat:
    • add_element files ONE picture as a character, prop or location. Its source is an https address the person gave you in this conversation (an image file -- jpg, png or webp -- or a product's page, whose main picture is taken), or gen:<id> of one of their own renders that project lists. Give it a kind, a name and a few words of notes. A name they already have of that kind gains the picture instead of making a second element. It draws nothing and spends nothing.
    • upload_link when the picture is on their phone or computer: call it with purpose reference and the element_kind + element_name it belongs to, give the person the url, and tell them to open it, pick the file and come back. One file (jpg, png, webp or heic), good for thirty minutes; afterwards elements lists it. Another file needs another link.
  • Never give add_element an address you found or guessed, and never search for a picture on the person's behalf. The photo is theirs to choose.
  • Refs go into write_scene as listed by elements, most important first: the face before the room, the product before the backdrop. The first ref is what the render anchors on.

Say what each photo is for: roles

A photo is attached for WHAT IT SHOWS or for HOW IT LOOKS, and the scene must know which. write_scene and duplicate_scene take roles, {ref: role}:

  • subject, product, place: the picture is OF the thing on screen -- the exact face, the exact product, the exact room. With no role given, a photo is read this way.
  • light, palette, framing, style: ONLY that is taken from the photo. A frame attached for its dusk light lends its light, not its street or the stranger in it; a palette photo lends its colours and nothing else.

Two rules the studio holds to, so hold to them when ordering refs:

  1. Adapt a reference's look, never copy its subject. A look-only photo is translated into the style block; the people, products and places on screen come from the subject photos.
  2. A look-only photo is never the anchor. The subject photo goes first in refs; a photo with a light, palette, framing or style role goes after.

Worked example: Sam's portrait, and a rain-soaked street at dusk the person likes for its light, brought in through upload_link. refs: [sam_ref, street_ref], roles: {street_ref: "light"}. The street's blue-hour glow and wet reflections go into the style block; Sam is the face; the street itself is not on screen unless the beats put it there.

Reading an attached image or video frame

When the person shares an image or a frame, translate it into the terms the scene's style block uses, and offer that back for their yes: light (source, direction, colour temperature, hardness, time of day); surfaces (wet, dusty, matte, worn); framing (how close, what height, lens feel, depth of field); palette and grade (saturated, muted, contrast); motion (what moves, how fast, handheld or locked); finish (grain, softness, sharpness).

Say what the studio will match (light, surfaces, framing, palette) and what it will not (a specific film's or campaign's authorship, a brand's logo, a person's likeness). Do not promise it will "reproduce" the reference.

Writing the look into the scene

The look lives in part 2 of the scene prompt (see write-scene): two or three sentences in cinematography and colour-grade terms, the subject photos winning over the look on anything they show. Realism comes from practical light, believable physics and restraint. Never name a film, a director, a brand or a real person in it.

Reply shape

One short block, readable on a phone: which element photos the scene is held to (by name, ref and role), and the look in the words that will go into the prompt.