140 lines
5.2 KiB
Python
140 lines
5.2 KiB
Python
"""按生图模型预处理 prompt（与 provider 解耦）。
|
||
|
||
各策略在对应函数中注明依据：厂商文档、Replicate 模型 API 字段说明或社区通用写法。
|
||
统一入口：prepare_image_prompt_for_model。
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
from dataclasses import dataclass
|
||
from typing import Any
|
||
|
||
# ─── SDXL（Stability / Replicate）────────────────────────────
|
||
# Replicate stability-ai/sdxl 提供独立字段 negative_prompt（见模型 API 页）。
|
||
# 下列为 CLIP 系文生图常见「质量/解剖」排除词，作未显式指定时的基线。
|
||
DEFAULT_SDXL_NEGATIVE = (
|
||
"low quality, worst quality, normal quality, lowres, blurry, jpeg artifacts, "
|
||
"watermark, signature, text, logo, deformed, disfigured, bad anatomy, bad hands, "
|
||
"extra fingers, mutated, cropped, poorly drawn face"
|
||
)
|
||
|
||
|
||
@dataclass(frozen=True)
|
||
class PreparedImagePrompt:
|
||
"""下游只读：prompt 必填；negative_prompt 仅部分后端使用（当前为 SDXL）。"""
|
||
|
||
prompt: str
|
||
negative_prompt: str | None = None
|
||
|
||
|
||
def prepare_image_prompt_for_model(
|
||
model_config: dict[str, Any],
|
||
raw_prompt: str,
|
||
*,
|
||
has_reference_images: bool = False,
|
||
) -> PreparedImagePrompt:
|
||
"""根据注册表 id 选择策略。未知 id 时原样透传。"""
|
||
short_id = model_config.get("id", "")
|
||
text = (raw_prompt or "").strip()
|
||
if not text:
|
||
return PreparedImagePrompt(prompt="")
|
||
|
||
dispatch: dict[str, Any] = {
|
||
"gpt-image-1.5": _openai_gpt_image,
|
||
"gemini-3.1-flash-image": _gemini_native_image,
|
||
"flux-schnell": _flux_bfl,
|
||
"flux-dev": _flux_bfl,
|
||
"sdxl": _sdxl_replicate,
|
||
"instant-style": _replicate_ip_adapter_scene,
|
||
"kolors-ipadapter": _replicate_ip_adapter_scene,
|
||
}
|
||
fn = dispatch.get(short_id, _passthrough)
|
||
return fn(text, model_config, has_reference_images)
|
||
|
||
|
||
def _passthrough(text: str, _model_config: dict[str, Any], _has_ref: bool) -> PreparedImagePrompt:
|
||
return PreparedImagePrompt(prompt=text)
|
||
|
||
|
||
def _openai_gpt_image(text: str, _model_config: dict[str, Any], _has_ref: bool) -> PreparedImagePrompt:
|
||
"""OpenAI GPT Image：自然语言指令遵循强，宜为完整、具体的场景描述。
|
||
|
||
参考：https://platform.openai.com/docs/guides/image-generation
|
||
"""
|
||
return PreparedImagePrompt(prompt=_collapse_ws(text))
|
||
|
||
|
||
def _gemini_native_image(text: str, _model_config: dict[str, Any], _has_ref: bool) -> PreparedImagePrompt:
|
||
"""Gemini 原生生图：多模态指令 + 文本；英文描述通常效果稳定。
|
||
|
||
参考：Google AI 文档 generateContent 与图像输出 modality 说明。
|
||
"""
|
||
return PreparedImagePrompt(prompt=_collapse_ws(text))
|
||
|
||
|
||
def _flux_bfl(text: str, _model_config: dict[str, Any], _has_ref: bool) -> PreparedImagePrompt:
|
||
"""Black Forest Labs FLUX：Subject + Action + Style + Context；靠前放置重点；无 negative API。
|
||
|
||
参考：https://docs.bfl.ai/guides/prompting_guide_t2i_fundamentals
|
||
若用户从 SDXL 复制了 Negative 段，尽量剥掉以免干扰文意。
|
||
"""
|
||
cleaned = _strip_pasted_negative_block(text)
|
||
return PreparedImagePrompt(prompt=_collapse_ws(cleaned))
|
||
|
||
|
||
def _sdxl_replicate(text: str, model_config: dict[str, Any], _has_ref: bool) -> PreparedImagePrompt:
|
||
"""Replicate SDXL：prompt + negative_prompt 双字段；支持显式拆分。
|
||
|
||
约定（可选）：正提示与负提示用单独一行分隔符，便于 Agent/用户手写。
|
||
- ---NEGATIVE--- 或 |||NEG|||
|
||
未拆分时使用 default_params.negative_prompt 或模块默认 DEFAULT_SDXL_NEGATIVE。
|
||
"""
|
||
neg_fallback = model_config.get("default_params", {}).get(
|
||
"negative_prompt", DEFAULT_SDXL_NEGATIVE
|
||
)
|
||
if "---NEGATIVE---" in text:
|
||
pos, _, neg = text.partition("---NEGATIVE---")
|
||
pos = pos.strip()
|
||
neg = neg.strip()
|
||
return PreparedImagePrompt(
|
||
prompt=_collapse_ws(pos),
|
||
negative_prompt=neg or neg_fallback,
|
||
)
|
||
if "|||NEG|||" in text:
|
||
pos, _, neg = text.partition("|||NEG|||")
|
||
pos = pos.strip()
|
||
neg = neg.strip()
|
||
return PreparedImagePrompt(
|
||
prompt=_collapse_ws(pos),
|
||
negative_prompt=neg or neg_fallback,
|
||
)
|
||
return PreparedImagePrompt(
|
||
prompt=_collapse_ws(text),
|
||
negative_prompt=neg_fallback,
|
||
)
|
||
|
||
|
||
def _replicate_ip_adapter_scene(
|
||
text: str, _model_config: dict[str, Any], has_reference_images: bool
|
||
) -> PreparedImagePrompt:
|
||
"""IP-Adapter / InstantStyle / Kolors：参考图承担风格与纹理，prompt 侧重场景与内容语义。
|
||
|
||
Replicate 各模型 README 均强调 prompt + 参考图配合；无参考图时由上层拦截。
|
||
有参考图时不额外堆叠长前缀，避免稀释主体描述。
|
||
"""
|
||
_ = has_reference_images
|
||
return PreparedImagePrompt(prompt=_collapse_ws(text))
|
||
|
||
|
||
def _collapse_ws(s: str) -> str:
|
||
return " ".join(s.split())
|
||
|
||
|
||
def _strip_pasted_negative_block(text: str) -> str:
|
||
lower = text.lower()
|
||
for sep in ("\n---negative---\n", "\nnegative prompt:", "\nnegative:"):
|
||
idx = lower.find(sep)
|
||
if idx != -1:
|
||
return text[:idx].strip()
|
||
return text
|