407 lines
15 KiB
Python
407 lines
15 KiB
Python
"""
|
||
EPEEKit 集中配置。
|
||
所有可调参数从环境变量读取，每次调用实时读取（不缓存），确保 load_dotenv() 后生效。
|
||
"""
|
||
|
||
import os
|
||
from typing import Any
|
||
|
||
|
||
def get_llm_max_iterations() -> int:
|
||
return int(os.getenv("LLM_MAX_ITERATIONS", "5"))
|
||
|
||
|
||
# ─── LLM 模型注册表 ─────────────────────────────────────
|
||
#
|
||
# 每个模型的配置说明：
|
||
# id — 前端/API 使用的短 ID
|
||
# name — 显示名称
|
||
# provider — API 提供者（对应 _get_client 的分发键）
|
||
# model_id — 传给 OpenAI SDK 的 model 参数
|
||
# description — 前端下拉列表说明文字
|
||
# vision — 是否支持多模态图片输入
|
||
|
||
LLM_MODELS: dict[str, dict[str, Any]] = {
|
||
"gpt-5.4": {
|
||
"id": "gpt-5.4",
|
||
"name": "GPT-5.4",
|
||
"provider": "vectorengine",
|
||
"model_id": "gpt-5.4",
|
||
"description": "知识工作与计算机操控最强，1M 上下文",
|
||
"vision": True,
|
||
},
|
||
"claude-sonnet-4-6": {
|
||
"id": "claude-sonnet-4-6",
|
||
"name": "Claude Sonnet 4.6",
|
||
"provider": "vectorengine",
|
||
"model_id": "claude-sonnet-4-6",
|
||
"description": "高性价比编码与日常任务",
|
||
"vision": True,
|
||
},
|
||
"claude-opus-4-6": {
|
||
"id": "claude-opus-4-6",
|
||
"name": "Claude Opus 4.6",
|
||
"provider": "vectorengine",
|
||
"model_id": "claude-opus-4-6",
|
||
"description": "编码与专家级推理最强，128K 输出",
|
||
"vision": True,
|
||
},
|
||
"gemini-3.1-pro-preview": {
|
||
"id": "gemini-3.1-pro-preview",
|
||
"name": "Gemini 3.1 Pro",
|
||
"provider": "vectorengine",
|
||
"model_id": "gemini-3.1-pro-preview",
|
||
"description": "推理最强、价格最低，2M 上下文",
|
||
"vision": True,
|
||
},
|
||
"glm-4.7": {
|
||
"id": "glm-4.7",
|
||
"name": "GLM-4.7",
|
||
"provider": "vectorengine",
|
||
"model_id": "glm-4.7",
|
||
"description": "智谱 AI，中文能力突出，免费额度",
|
||
"vision": False,
|
||
},
|
||
"gpt-4o-mini": {
|
||
"id": "gpt-4o-mini",
|
||
"name": "GPT-4o Mini",
|
||
"provider": "vectorengine",
|
||
"model_id": "gpt-4o-mini",
|
||
"description": "轻量快速、高性价比",
|
||
"vision": True,
|
||
},
|
||
"deepseek-chat": {
|
||
"id": "deepseek-chat",
|
||
"name": "DeepSeek Chat",
|
||
"provider": "deepseek",
|
||
"model_id": "deepseek-chat",
|
||
"description": "中文对话优化（直连）",
|
||
"vision": False,
|
||
},
|
||
}
|
||
|
||
|
||
def get_default_llm_model_id() -> str:
|
||
"""返回 .env 中配置的默认 LLM 短 ID，不在注册表中则回退到 gpt-5.4。"""
|
||
env_model = os.getenv("LLM_MODEL", "gpt-5.4")
|
||
if env_model in LLM_MODELS:
|
||
return env_model
|
||
return "gpt-5.4"
|
||
|
||
|
||
def get_llm_model_config(model_id: str | None = None) -> dict[str, Any]:
|
||
"""根据短 ID 获取 LLM 模型配置，未指定或不存在则使用默认模型。"""
|
||
if model_id and model_id in LLM_MODELS:
|
||
return LLM_MODELS[model_id]
|
||
return LLM_MODELS[get_default_llm_model_id()]
|
||
|
||
|
||
def get_llm_models_list() -> list[dict]:
|
||
"""返回前端下拉列表所需的 LLM 模型摘要信息。"""
|
||
return [
|
||
{
|
||
"id": cfg["id"],
|
||
"name": cfg["name"],
|
||
"description": cfg["description"],
|
||
"vision": cfg.get("vision", False),
|
||
}
|
||
for cfg in LLM_MODELS.values()
|
||
]
|
||
|
||
|
||
# ─── 图像模型注册表 ─────────────────────────────────────
|
||
#
|
||
# 每个模型的配置说明：
|
||
# id — 前端/API 使用的短 ID
|
||
# name — 显示名称
|
||
# provider — 生成服务提供者（对应 image_gen.py 中的 Provider）
|
||
# model_id — Replicate 上的完整模型 ID
|
||
# description — 前端下拉列表中的说明文字
|
||
# supports_ref_image — 是否原生支持参考图输入（IP-Adapter 等）
|
||
# ref_image_param — 传给 Replicate 的参考图参数名（模型间可能不同）
|
||
# num_images_param — 批量生成参数名（Flux 用 num_outputs，Kolors 用 number_of_images）
|
||
# default_params — 默认推理参数
|
||
|
||
IMAGE_MODELS: dict[str, dict[str, Any]] = {
|
||
"gpt-image-1.5": {
|
||
"id": "gpt-image-1.5",
|
||
"name": "GPT Image 1.5",
|
||
"provider": "openai",
|
||
"model_id": "gpt-image-1.5",
|
||
"description": "OpenAI 最强生图，文字渲染与 prompt 理解最佳，支持多图参考",
|
||
"supports_ref_image": True,
|
||
"max_ref_images": 16,
|
||
"default_params": {
|
||
"size": "1024x1024",
|
||
"quality": "high",
|
||
},
|
||
},
|
||
"gemini-3.1-flash-image": {
|
||
"id": "gemini-3.1-flash-image",
|
||
"name": "Gemini 3.1 Flash Image",
|
||
"provider": "gemini_native",
|
||
"model_id": "gemini-3.1-flash-image-preview",
|
||
"description": "Google 原生生图，速度快、价格低，支持多图参考（最多 14 张）",
|
||
"supports_ref_image": True,
|
||
"max_ref_images": 14,
|
||
"default_params": {},
|
||
},
|
||
"flux-schnell": {
|
||
"id": "flux-schnell",
|
||
"name": "Flux Schnell",
|
||
"provider": "replicate",
|
||
"model_id": "black-forest-labs/flux-schnell",
|
||
"description": "快速生成，适合快速迭代",
|
||
"supports_ref_image": False,
|
||
"num_images_param": "num_outputs",
|
||
"default_params": {
|
||
"aspect_ratio": "1:1",
|
||
"output_format": "png",
|
||
},
|
||
},
|
||
"flux-dev": {
|
||
"id": "flux-dev",
|
||
"name": "Flux Dev",
|
||
"provider": "replicate",
|
||
"model_id": "black-forest-labs/flux-dev",
|
||
"description": "高质量生成，细节更好",
|
||
"supports_ref_image": False,
|
||
"num_images_param": "num_outputs",
|
||
"default_params": {
|
||
"aspect_ratio": "1:1",
|
||
"output_format": "png",
|
||
},
|
||
},
|
||
"sdxl": {
|
||
"id": "sdxl",
|
||
"name": "Stable Diffusion XL",
|
||
"provider": "replicate",
|
||
"model_id": "stability-ai/sdxl:39ed52f2a78e934b3ba6e2a89f5b1c712de7dfea535525255b1aa35c5565e08b",
|
||
"description": "经典 SDXL，支持 negative prompt",
|
||
"supports_ref_image": False,
|
||
"num_images_param": "num_outputs",
|
||
"default_params": {
|
||
"width": 1024,
|
||
"height": 1024,
|
||
"num_inference_steps": 50,
|
||
"guidance_scale": 7.5,
|
||
},
|
||
},
|
||
"instant-style": {
|
||
"id": "instant-style",
|
||
"name": "InstantStyle",
|
||
"provider": "replicate",
|
||
"model_id": "jyoung105/instant-style:c6f01e12f31cb99f9ee774a78992a71294f630a6f433d9aecfdc33b816fc4baa",
|
||
"description": "强风格迁移，画风还原度高（较慢）",
|
||
"supports_ref_image": True,
|
||
"ref_image_param": "style_image",
|
||
"num_images_param": "num_outputs",
|
||
"default_params": {
|
||
"width": 1024,
|
||
"height": 1024,
|
||
"num_inference_steps": 30,
|
||
"guidance_scale": 5,
|
||
"style_strength": 1.0,
|
||
"block_mode": "style-only",
|
||
"adapter_mode": "original",
|
||
},
|
||
},
|
||
"kolors-ipadapter": {
|
||
"id": "kolors-ipadapter",
|
||
"name": "Kolors IP-Adapter",
|
||
"provider": "replicate",
|
||
"model_id": "fofr/kolors-with-ipadapter:5a1a92b2c0f81813225d48ed8e411813da41aa84e7582fb705d1af46eea36eed",
|
||
"description": "风格参考生成，上传参考图效果最佳",
|
||
"supports_ref_image": True,
|
||
"ref_image_param": "image",
|
||
"num_images_param": "number_of_images",
|
||
"default_params": {
|
||
"width": 1024,
|
||
"height": 1024,
|
||
"steps": 25,
|
||
"cfg": 4,
|
||
"ip_adapter_weight": 0.8,
|
||
"ip_adapter_weight_type": "style transfer precise",
|
||
"output_format": "png",
|
||
},
|
||
},
|
||
# Mesh 管道专用的内部模型（internal=True 不展示到前端下拉）：
|
||
# 输入两张图：image=原图做 IP-Adapter 风格参考；controlnet_input=normal/depth 结构参考。
|
||
# 用途：在 Stage 3 对 Trellis 抽出来的 normal 帧做二次重绘，还原原画风。
|
||
"ip-adapter-controlnet-depth": {
|
||
"id": "ip-adapter-controlnet-depth",
|
||
"name": "IP-Adapter + ControlNet Depth (内部)",
|
||
"provider": "replicate",
|
||
"model_id": "chigozienri/ip_adapter-sdxl-controlnet-depth:0436c8702ef52616be5c30948551b3af6a86c821cca9b01f11ac297624fff14c",
|
||
"description": "Mesh 管道 Stage 3 专用：IP-Adapter 保风格 + ControlNet 保几何",
|
||
"internal": True,
|
||
"supports_ref_image": True,
|
||
"ref_image_param": "image", # IP-Adapter 参考图字段
|
||
"default_params": {
|
||
"scale": 0.75, # IP-Adapter 权重（建议 0.7-0.8）
|
||
"controlnet_conditioning_scale": 0.8, # 结构约束（建议 0.7-0.9）
|
||
"prompt": "same building, same art style, consistent with reference",
|
||
"negative_prompt": "blurry, distorted, different style, realistic photo, photograph",
|
||
"num_outputs": 1,
|
||
"num_inference_steps": 30,
|
||
},
|
||
},
|
||
}
|
||
|
||
|
||
def get_style_restore_model_id() -> str:
|
||
"""Mesh 管道 Stage 3 使用的内部模型短 ID。"""
|
||
return "ip-adapter-controlnet-depth"
|
||
|
||
|
||
# ─── 视角变换模型注册表 ─────────────────────────────────
|
||
#
|
||
# 视角变换模型不用于生图，而是将已有图片转换为不同视角。
|
||
# 输入通常只需一张图，输出为多个固定视角的图片。
|
||
|
||
VIEW_TRANSFORM_MODELS: dict[str, dict[str, Any]] = {
|
||
"zero123plus": {
|
||
"id": "zero123plus",
|
||
"name": "Zero123++",
|
||
"provider": "replicate",
|
||
"model_id": "jd7h/zero123plusplus:c69c6559a29011b576f1ff0371b3bc1add2856480c60520c7e9ce0b40a6e9052",
|
||
"description": "单图生成 6 个固定视角，适合建筑/物体的多角度预览",
|
||
"pipeline": "grid",
|
||
"output_views": [
|
||
{"azimuth": 30, "elevation": 30},
|
||
{"azimuth": 90, "elevation": -20},
|
||
{"azimuth": 150, "elevation": 30},
|
||
{"azimuth": 210, "elevation": -20},
|
||
{"azimuth": 270, "elevation": 30},
|
||
{"azimuth": 330, "elevation": -20},
|
||
],
|
||
"grid_layout": {"cols": 2, "rows": 3},
|
||
"enabled": True,
|
||
},
|
||
"trellis": {
|
||
"id": "trellis",
|
||
"name": "Trellis (Mesh Pipeline)",
|
||
"provider": "replicate",
|
||
"model_id": "firtoz/trellis:e8f6c45206993f297372f5436b90350817bd9b4a0d52d2a76df50c1c8afa2b3c",
|
||
"description": "3D 重建 + 任意视角 + 风格还原，建筑友好（~30s 重建 + 72s/张重绘）",
|
||
"pipeline": "mesh",
|
||
"default_params": {
|
||
"texture_size": 1024,
|
||
"mesh_simplify": 0.95,
|
||
"generate_color": True,
|
||
"generate_normal": True,
|
||
"generate_model": True,
|
||
"save_gaussian_ply": False,
|
||
"ss_sampling_steps": 12,
|
||
"slat_sampling_steps": 12,
|
||
"ss_guidance_strength": 7.5,
|
||
"slat_guidance_strength": 3.0,
|
||
},
|
||
"default_azimuths": [0, 60, 120, 180, 240, 300],
|
||
"default_elevations": [0, 0, 0, 0, 0, 0],
|
||
"enabled": True,
|
||
},
|
||
"hunyuan3d": {
|
||
"id": "hunyuan3d",
|
||
"name": "Hunyuan3D-2 (高质量)",
|
||
"provider": "replicate",
|
||
"model_id": "tencent/hunyuan3d-2:b1b9449a1277e10402781c5d41eb30c0a0683504fb23fab591ca9dfc2aabe1cb",
|
||
"description": "几何质量最佳（~127s），需自定义渲染器（二期启用）",
|
||
"pipeline": "mesh",
|
||
"default_params": {
|
||
"steps": 50,
|
||
"guidance_scale": 5.5,
|
||
"octree_resolution": 256,
|
||
"remove_background": True,
|
||
},
|
||
"enabled": False,
|
||
},
|
||
}
|
||
|
||
|
||
def get_view_transform_model_config(model_id: str | None = None) -> dict[str, Any]:
|
||
"""获取视角变换模型配置，默认返回 zero123plus。"""
|
||
if model_id and model_id in VIEW_TRANSFORM_MODELS:
|
||
return VIEW_TRANSFORM_MODELS[model_id]
|
||
return VIEW_TRANSFORM_MODELS["zero123plus"]
|
||
|
||
|
||
def get_view_transform_models_list() -> list[dict]:
|
||
"""返回前端下拉列表所需的视角变换模型摘要信息（只包含 enabled=True 的）。"""
|
||
return [
|
||
{
|
||
"id": cfg["id"],
|
||
"name": cfg["name"],
|
||
"description": cfg["description"],
|
||
"pipeline": cfg.get("pipeline", "grid"),
|
||
}
|
||
for cfg in VIEW_TRANSFORM_MODELS.values()
|
||
if cfg.get("enabled", True)
|
||
]
|
||
|
||
|
||
def get_default_image_model_id() -> str:
|
||
"""返回 .env 中配置的默认模型短 ID，若不在注册表中则回退到 flux-schnell。"""
|
||
env_model = os.getenv("IMAGE_MODEL", "flux-schnell")
|
||
for mid, cfg in IMAGE_MODELS.items():
|
||
if cfg["model_id"] == env_model or mid == env_model:
|
||
return mid
|
||
return "flux-schnell"
|
||
|
||
|
||
def get_image_model_config(model_id: str | None = None) -> dict[str, Any]:
|
||
"""根据短 ID 获取模型配置，未指定或不存在则使用默认模型。"""
|
||
if model_id and model_id in IMAGE_MODELS:
|
||
return IMAGE_MODELS[model_id]
|
||
return IMAGE_MODELS[get_default_image_model_id()]
|
||
|
||
|
||
def get_ref_image_model_id() -> str | None:
|
||
"""返回有参考图时推荐使用的模型 ID（第一个 supports_ref_image=True 的模型）。"""
|
||
for mid, cfg in IMAGE_MODELS.items():
|
||
if cfg.get("supports_ref_image"):
|
||
return mid
|
||
return None
|
||
|
||
|
||
def get_image_models_list() -> list[dict]:
|
||
"""返回前端下拉列表所需的模型摘要信息（过滤掉 internal=True 的内部模型）。"""
|
||
return [
|
||
{
|
||
"id": cfg["id"],
|
||
"name": cfg["name"],
|
||
"description": cfg["description"],
|
||
"supports_ref_image": cfg.get("supports_ref_image", False),
|
||
}
|
||
for cfg in IMAGE_MODELS.values()
|
||
if not cfg.get("internal", False)
|
||
]
|
||
|
||
|
||
# ─── 记忆系统配置 ─────────────────────────────────────
|
||
|
||
def get_deepseek_api_key() -> str:
|
||
return os.getenv("DEEPSEEK_API_KEY", "")
|
||
|
||
|
||
def get_ollama_base_url() -> str:
|
||
return os.getenv("OLLAMA_BASE_URL", "http://localhost:11434")
|
||
|
||
|
||
def get_mem0_embedding_model() -> str:
|
||
return os.getenv("MEM0_EMBEDDING_MODEL", "nomic-embed-text")
|
||
|
||
|
||
def get_max_recent_turns() -> int:
|
||
return int(os.getenv("MAX_RECENT_TURNS", "20"))
|
||
|
||
|
||
# ─── 图像输出配置 ─────────────────────────────────────
|
||
|
||
def get_image_aspect_ratio() -> str:
|
||
return os.getenv("IMAGE_ASPECT_RATIO", "1:1")
|
||
|
||
|
||
def get_image_output_format() -> str:
|
||
return os.getenv("IMAGE_OUTPUT_FORMAT", "png")
|