image-suite-studio/server/services/prompts/gpt.py

"""GPT 图像模型（gpt-image-2 / gpt-image-2-vip，RightAPI 中转）提示词。

语义：/v1/images/edits —— 输入图是"被编辑的照片"，prompt 是编辑指令；
与通义/豆包的"主体参考"语义完全不同：参考图不是商品锚，模型会按文字指令
重新渲染整张图。此前与国产模型共用场景提示词，再用文字锚定商品并要求输出
"匹配商品描述"，导致模型把商品改造成营销关键词描述的样子（必现商品被改）。

本模块写法原则：
  1. 商品只由 Image 1 定义；标题/参数仅作识别背景并声明"以图为准"，
     绝不要求输出匹配文字描述（那等于授权模型改商品）；
  2. 指令只说"改什么"（背景/场景/排版/文案），不描述商品外观；
  3. 分节精简、首尾重申保真；不用负面清单（gpt 无 negative_prompt 参数，
     罗列畸形反而往上下文植入概念）；
  4. sku 合集 / 多拼图明确"复制同一件商品，禁止发明新配色或变体"。
"""
from __future__ import annotations

from .common import TEXT_RENDER, requirements_block, resolve_style, selling_point_lines

# 保真锁：商品由 Image 1 唯一定义，其余指令一律不得触碰商品本体
_PRESERVE = (
    "PRESERVE (absolute, overrides every other instruction below): the product shown in Image 1. "
    "Reuse the photographed product exactly as it is — identical shape, silhouette, proportions, "
    "colors, print/pattern, logo and label text, materials, stitching and surface details. "
    "Do not redesign, restyle, recolor, re-pattern, tidy up or substitute the product, "
    "and do not let any style or text instruction below alter it. Image 2 is a secondary "
    "view of the same product for reference only."
)

_STYLE = (
    "SCENE STYLE (applies to background, scene, props and lighting only — never to the product): "
)

_QUALITY = (
    "OUTPUT: photorealistic commercial e-commerce photography, ultra-detailed, "
    "natural light and shadow, professional retouching."
)

_REMINDER = (
    "FINAL CHECK: the product itself must remain exactly as photographed in Image 1 — "
    "only its surroundings, composition, lighting and overlay graphics may differ."
)

# 图内文案：明确是"排版图层"，不落在商品本体上
_TEXT_SCOPE = (
    "TEXT OVERLAY (a graphic layer on the composition, never printed on the product): "
)


def _anchor(ctx: dict) -> str:
    """商品文字锚定：仅供识别，明确声明以图为准。

    只放标题 + 参数、不放营销描述——描述里的卖点词（"卡通""加固""防水"等）
    在 edits 语义下会被执行到商品上；官逆通道（-vip）参考图被弱化时，
    文字锚定用于帮模型认出"是哪件商品"，而不是"长什么样"。
    """
    line = f"CONTEXT (identification only): the product is \"{ctx['title']}\""
    if ctx.get("params_line"):
        line += f" ({ctx['params_line']})"
    return (
        line
        + ". Image 1 — not this text — defines the product's appearance; "
        "if they ever conflict, follow Image 1."
    )


# ── 各图类型的编辑指令（只描述改动，不描述商品）────────────────────────────

def _task_white_bg(ctx: dict, lang: str) -> str:
    return (
        "TASK: Clean up this product photo for a marketplace listing. Replace the entire "
        "background with seamless pure white (RGB 255,255,255); recompose with the product "
        "centered in front view filling about 85% of the frame; keep only a faint natural "
        "contact shadow. No props, no text, no background elements."
    )

def _task_key_features(ctx: dict, lang: str) -> str:
    sp = selling_point_lines(ctx, lang) or ctx["title"]
    return (
        "TASK: Feature infographic on a square canvas. Keep the product unchanged on the left "
        "two-thirds; build the right third as a clean info panel listing 3 feature callouts with "
        f"minimal line icons and thin leader lines pointing at parts of the product. Callout copy: {sp}."
    )

def _task_selling_pt(ctx: dict, lang: str) -> str:
    sp = selling_point_lines(ctx, lang, 1) or ctx["title"]
    return (
        "TASK: Single-selling-point poster. Hero close-up of the unchanged product at a dynamic "
        f"angle, generous negative space, one large bold headline about \"{sp}\", plus one small "
        "magnified circle zooming into an existing detail of the product (zoom only — do not "
        "invent details that are not in the photo)."
    )

def _task_material(ctx: dict, lang: str) -> str:
    return (
        "TASK: Material close-up. Zoom tightly into an existing area of the product's surface and "
        "show its true texture — fabric weave, surface finish, stitching — exactly as it appears in "
        "Image 1; shallow depth of field, raking light across the surface; small caption label in a corner."
    )

def _task_lifestyle(ctx: dict, lang: str) -> str:
    return (
        "TASK: Lifestyle scene. Place the unchanged product into a realistic everyday environment "
        "where it would naturally be used: human-scale surroundings, soft daylight, authentic candid "
        "mood, matched shadows and color temperature, the product remaining the clear visual focus."
    )

def _task_multi_scene(ctx: dict, lang: str) -> str:
    sp = selling_point_lines(ctx, lang)
    task = (
        "TASK: Triptych showcase. Build three vertical panels separated by thin gutters; every panel "
        "contains an IDENTICAL copy of the product from Image 1 (do not re-render it differently per "
        "panel) placed in a different usage scene (e.g. home interior / outdoor street / office desk), "
        "with consistent color grading across panels."
    )
    if sp:
        task += f" Panel captions: {sp}."
    return task

def _task_ecommerce_detail(ctx: dict, lang: str) -> str:
    sp = selling_point_lines(ctx, lang) or ctx["title"]
    params = ctx["params_line"]
    return (
        "TASK: Detail-page hero section on a square canvas. Top half: hero banner with the unchanged "
        "product at a 3/4 angle. Bottom half: clean spec card with 3 feature rows and line icons"
        + (f" (specs: {params})" if params else "")
        + f", one highlighted row: {sp}."
    )

def _task_size_chart(ctx: dict, lang: str) -> str:
    dims = ctx["params_line"]
    return (
        "TASK: Measurement chart. Show the unchanged product in clean front and side views on a light "
        "background; overlay thin technical annotation lines (arrows) marking length, width and height "
        "with measurement values rendered beside each line"
        + (f" (known specs: {dims})" if dims else "")
        + "; precise technical-drawing aesthetic, small caption row."
    )

def _task_sku_collection(ctx: dict, lang: str) -> str:
    # 关键差异：不允许像国产模型那样展开"全部配色"——edits 语义下那会凭空造出新商品
    return (
        "TASK: Product lineup. Arrange several IDENTICAL copies of the product from Image 1 in a neat "
        "equal grid (2-4 per row) with a small label chip below each copy; identical lighting and scale "
        "across copies. Every copy must show this exact product — do NOT invent other colorways, "
        "variants or versions."
    )

_TASK_BUILDERS = {
    "white_bg": (_task_white_bg, False),
    "key_features": (_task_key_features, True),
    "selling_pt": (_task_selling_pt, True),
    "material": (_task_material, True),
    "lifestyle": (_task_lifestyle, True),
    "multi_scene": (_task_multi_scene, True),
    "ecommerce_detail": (_task_ecommerce_detail, True),
    "size_chart": (_task_size_chart, True),
    "sku_collection": (_task_sku_collection, True),
}


def build_prompt(type_id: str, ctx: dict, style_set: int, lang: str, extra: dict | None = None,
                 style_prompt: str | None = None, requirements: str | None = None) -> str:
    """构造指定图类型的 edits 语义 prompt：要求块 → 编辑指令 → 保真锁 → 锚定 → 风格 → 文案 → 画质 → 终检。"""
    style = resolve_style(style_set, style_prompt)
    extra = extra or {}
    hint = (extra.get("prompt_hint") or "").strip()

    if type_id == "custom":
        purpose = extra.get("title") or ""
        detail = extra.get("detail") or ""
        task = "TASK: Create an e-commerce marketing image featuring the product from Image 1"
        task += f" — {purpose}" if purpose else ""
        task += f": {detail}" if detail else ""
        task += "."
        wants_text = True
    else:
        entry = _TASK_BUILDERS.get(type_id)
        if entry is None:
            raise ValueError(f"未知图类型: {type_id}")
        builder, wants_text = entry
        task = builder(ctx, lang)
    if hint:
        task += f" Composition guidance: {hint}."

    parts = [p for p in (requirements_block(requirements),) if p]
    parts.append(task)
    parts.append(_PRESERVE)
    parts.append(_anchor(ctx))
    parts.append(f"{_STYLE}{style['tone']}.")
    if wants_text:
        parts.append(f"{_TEXT_SCOPE}{TEXT_RENDER[lang]}")
    parts.append(_QUALITY)
    parts.append(_REMINDER)
    return "\n\n".join(parts)