#!/usr/bin/env python3 # -*- coding: utf-8 -*- """ ChildHood AI Art Generation Tool 通过 API易 平台调用 AI 生图模型 Usage: python image_gen.py --test # 测试 API 连通性 python image_gen.py --generate "千禧年中国城郊的小卖部门口" # 文生图（自动加风格前缀） python image_gen.py --generate "a red apple" --no-style # 纯文生图（不加风格前缀） python image_gen.py --calibrate # 风格校准（5 个测试物体） python image_gen.py --fuse --style style.png --content photo.jpg # 双参考融合 python image_gen.py --fuse --style style.png --content photo.jpg --prompt "自定义 prompt" Requirements: pip install openai python-dotenv """ import argparse import base64 import io import json import os import sys import urllib.request # Windows 控制台 UTF-8 支持 if sys.platform == "win32": sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding="utf-8", errors="replace") sys.stderr = io.TextIOWrapper(sys.stderr.buffer, encoding="utf-8", errors="replace") import os import sys import urllib.request from datetime import datetime from pathlib import Path # --------------------------------------------------------------------------- # 路径配置 # --------------------------------------------------------------------------- SCRIPT_DIR = Path(__file__).resolve().parent # ai-art-research/tools/ RESEARCH_DIR = SCRIPT_DIR.parent # ai-art-research/ DOTENV_PATH = RESEARCH_DIR / ".env" OUTPUT_BASE = RESEARCH_DIR / "生成结果" # --------------------------------------------------------------------------- # 加载 .env # --------------------------------------------------------------------------- if DOTENV_PATH.exists(): from dotenv import load_dotenv load_dotenv(DOTENV_PATH) API_KEY = os.getenv("APIYI_API_KEY", "") BASE_URL = os.getenv("APIYI_BASE_URL", "https://api.apiyi.com/v1") DEFAULT_MODEL = os.getenv("APIYI_MODEL", "gpt-image-2") # --------------------------------------------------------------------------- # 风格 Prompt 种子（来自工作流方案 v0.1） # --------------------------------------------------------------------------- STYLE_SEED_EN = ( "3D low-poly game art, flat cel-shaded with solid color blocks, " "bright warm-cool balanced saturated palette, soft diffused lighting " "with subtle shadows, simplified geometric forms with clean edges, " "low-angle third-person perspective, nostalgic warm summer atmosphere" ) # 风格校准测试物体（中文名, 英文 prompt） CALIBRATION_OBJECTS = [ ("乌鸦", "a crow standing on the ground"), ("茶杯", "a teacup on a wooden table"), ("老人", "an old man standing in a village street"), ("山湖", "a mountain lake with trees"), ("苹果", "a red apple on the ground"), ] # 双参考融合默认 prompt DEFAULT_FUSION_PROMPT = """\ 图1 是我的风格参考（一张游戏截图）。图2 是我的内容参考（一张现实照片）。 请生成一张新图片，用图1的美术风格来呈现图2中的场景内容。 风格要求： - 高度简化的低多边形几何形体，干净利落的边缘 - 平面色块着色，不用渐变或写实纹理 - 鲜艳明亮的暖冷平衡色调（暖黄、天蓝、翠绿为主） - 柔和散射光，阴影淡而柔和 - 微低角度第三人称视角 - 温暖怀旧的夏日午后氛围 内容要求： - 保留图2中的建筑类型、布局、空间关系 - 但把所有建筑元素替换为千禧年中国城郊的风格 - 色调保持温暖怀旧""" # --------------------------------------------------------------------------- # 工具函数 # --------------------------------------------------------------------------- def get_client() -> "OpenAI": """创建 OpenAI 客户端""" if not API_KEY: print("❌ 未找到 API Key") print(f" 请在 {DOTENV_PATH} 中设置: APIYI_API_KEY=sk-xxx") sys.exit(1) from openai import OpenAI return OpenAI(api_key=API_KEY, base_url=BASE_URL) def _extract_image_bytes(img_obj) -> bytes: """从 API 响应中提取图片字节数据""" # 优先 base64 b64 = getattr(img_obj, "b64_json", None) if b64: return base64.b64decode(b64) # 其次 URL url = getattr(img_obj, "url", None) if url: print(" ⬇️ 正在下载图片...") return urllib.request.urlopen(url).read() raise ValueError("响应中未找到图片数据（无 b64_json 也无 url）") def _sanitize_filename(s: str, max_len: int = 40) -> str: """清理文件名中的非法字符""" result = "".join(c if c.isalnum() or c in "._-" else "_" for c in s) return result[:max_len].rstrip("_") # 模式 → 简短英文文件名 tag # 避免把整段中文 prompt 塞进文件名（CDN URL 过长/编码异常会让图像分析工具解析失败） _MODE_SLUG = { "text_to_image": "gen", "style_calibration": "calib", "multi_reference_fusion": "fusion", "dual_reference_fusion": "fusion", "connection_test": "test", } def save_result(image_bytes: bytes, metadata: dict, subdir: str = "") -> Path: """保存图片和元数据，返回图片路径。 文件名用「时间戳_模型_模式」短英文命名（如 260617-183020_gpt-image-2_fusion.png）； 完整 prompt 与参数存在 .meta.json 里，不丢信息。 """ ts = datetime.now().strftime("%y%m%d-%H%M%S") model_slug = metadata.get("model", "unknown").replace("/", "-").replace(".", "-") mode_slug = _MODE_SLUG.get(metadata.get("mode", ""), "img") out_dir = OUTPUT_BASE / subdir if subdir else OUTPUT_BASE out_dir.mkdir(parents=True, exist_ok=True) filename = f"{ts}_{model_slug}_{mode_slug}.png" img_path = out_dir / filename img_path.write_bytes(image_bytes) meta_path = img_path.with_suffix(".meta.json") meta_path.write_text(json.dumps(metadata, ensure_ascii=False, indent=2), encoding="utf-8") print(f"✅ 已保存: {img_path.name}") return img_path # --------------------------------------------------------------------------- # 功能：测试连通性 # --------------------------------------------------------------------------- def cmd_test(client): """测试 API 易连通性，生成一张简单测试图""" print("🔗 测试 API易 连通性...") print(f" Base URL: {BASE_URL}") print(f" Model: {DEFAULT_MODEL}") print(f" API Key: {API_KEY[:8]}...{API_KEY[-4:]}") print() prompt = f"{STYLE_SEED_EN}, a simple red cube on a white background, test image" print(f" Prompt: {prompt[:80]}...") response = client.images.generate( model=DEFAULT_MODEL, prompt=prompt, size="1024x1024", quality="low", n=1, ) image_bytes = _extract_image_bytes(response.data[0]) save_result(image_bytes, { "model": DEFAULT_MODEL, "prompt": prompt, "mode": "connection_test", "size": "1024x1024", "quality": "low", "timestamp": datetime.now().isoformat(), }, subdir="测试") print() print("✅ API 连通性测试通过！") # --------------------------------------------------------------------------- # 功能：文生图 # --------------------------------------------------------------------------- def cmd_generate(client, prompt: str, use_style: bool, size: str, quality: str, model: str, subdir: str): """文生图""" model = model or DEFAULT_MODEL full_prompt = f"{STYLE_SEED_EN}, {prompt}" if use_style else prompt print(f"🎨 文生图") print(f" 模型: {model}") print(f" 尺寸: {size} 质量: {quality}") print(f" 风格前缀: {'是' if use_style else '否'}") print(f" Prompt: {full_prompt[:100]}{'...' if len(full_prompt) > 100 else ''}") print() response = client.images.generate( model=model, prompt=full_prompt, size=size, quality=quality, n=1, ) image_bytes = _extract_image_bytes(response.data[0]) save_result(image_bytes, { "model": model, "prompt": full_prompt, "mode": "text_to_image", "style_prefix_used": use_style, "size": size, "quality": quality, "timestamp": datetime.now().isoformat(), }, subdir=subdir) # --------------------------------------------------------------------------- # 功能：风格校准 # --------------------------------------------------------------------------- def cmd_calibrate(client, model: str): """风格校准 — 5 个测试物体""" model = model or DEFAULT_MODEL subdir = "风格校准" print("🎯 开始风格校准测试") print(f" 模型: {model}") print(f" 风格种子: {STYLE_SEED_EN[:60]}...") print(f" 测试物体: {', '.join(zh for zh, _ in CALIBRATION_OBJECTS)}") print() success = 0 for zh_name, en_prompt in CALIBRATION_OBJECTS: full_prompt = f"{STYLE_SEED_EN}, {en_prompt}" print(f" 🎨 生成: {zh_name} ...") try: response = client.images.generate( model=model, prompt=full_prompt, size="1024x1024", quality="medium", n=1, response_format="b64_json", ) image_bytes = _extract_image_bytes(response.data[0]) save_result(image_bytes, { "model": model, "prompt": full_prompt, "mode": "style_calibration", "object_zh": zh_name, "object_en": en_prompt, "style_seed": STYLE_SEED_EN, "timestamp": datetime.now().isoformat(), }, subdir=subdir) success += 1 except Exception as e: print(f" ❌ 失败: {e}") print() print(f"🎉 风格校准完成！成功 {success}/{len(CALIBRATION_OBJECTS)} 张") print(f" 输出目录: {OUTPUT_BASE / subdir}") print() print("📋 验证方法：打开上述目录，检查 5 张图是否视觉风格统一。") print(" 如果风格不一致，需要调整 STYLE_SEED_EN 后重新运行。") # --------------------------------------------------------------------------- # 功能：双参考融合 # --------------------------------------------------------------------------- def cmd_fuse(client, style_path: str, content_paths, prompt: str, model: str, subdir: str): """参考图融合（1 张风格参考 + 1~N 张内容参考）""" model = model or DEFAULT_MODEL fusion_prompt = prompt or DEFAULT_FUSION_PROMPT subdir = subdir or "双参考融合" style_file = Path(style_path).resolve() if not style_file.exists(): print(f"❌ 风格参考图不存在: {style_file}") sys.exit(1) # content_paths 兼容单个 str/Path 或 list if isinstance(content_paths, (str, Path)): content_paths = [content_paths] content_files = [] for cp in content_paths: cf = Path(cp).resolve() if not cf.exists(): print(f"❌ 内容参考图不存在: {cf}") sys.exit(1) content_files.append(cf) label = "双参考融合" if len(content_files) == 1 else f"多参考融合（1 风格 + {len(content_files)} 内容）" print(f"🔗 {label}") print(f" 模型: {model}") print(f" 风格参考: {style_file.name}") for cf in content_files: print(f" 内容参考: {cf.name}") print(f" Prompt: {fusion_prompt[:80]}...") print() # 确定图片 MIME 类型 ext_map = {".png": "image/png", ".jpg": "image/jpeg", ".jpeg": "image/jpeg", ".webp": "image/webp", ".gif": "image/gif"} # openai 库文件上传格式：tuple = (filename, bytes, content_type) style_tuple = (style_file.name, style_file.read_bytes(), ext_map.get(style_file.suffix.lower(), "image/png")) content_tuples = [ (cf.name, cf.read_bytes(), ext_map.get(cf.suffix.lower(), "image/png")) for cf in content_files ] image_list = [style_tuple] + content_tuples image_bytes = None errors = [] # 方式 1: images.edit 多图（风格参考 + 全部内容参考） print(f" 尝试方式 1: images.edit ({len(image_list)} 图) 融合 ...") try: response = client.images.edit( model=model, image=image_list, prompt=fusion_prompt, size="1024x1024", ) image_bytes = _extract_image_bytes(response.data[0]) print(" ✅ 方式 1 成功") except Exception as e: errors.append(f"images.edit 多图: {e}") print(f" ⚠️ 方式 1 失败: {e}") # 方式 2: images.edit 单图（仅第一张内容参考，风格靠文字）—— 降级 if image_bytes is None: print(" 尝试方式 2: images.edit 单图（风格写入 prompt）...") single_prompt = ( f"{STYLE_SEED_EN}\n\n" "Re-render the uploaded photo into this 3D low-poly cel-shaded game art style. " "Keep the photo's scene layout, building structure and objects, but re-render " "everything in the style described above. Warm nostalgic late-afternoon lighting, " "clean silhouette, game concept art." ) try: response = client.images.edit( model=model, image=content_tuples[0], prompt=single_prompt, size="1024x1024", ) image_bytes = _extract_image_bytes(response.data[0]) print(" ✅ 方式 2 成功（单图 + 文字风格，效果弱于多图）") except Exception as e: errors.append(f"images.edit 单图: {e}") print(f" ⚠️ 方式 2 失败: {e}") if image_bytes is None: print() print("❌ 双参考融合 API 调用失败（双图 edit 与单图 edit 均失败）。") print() print("💡 可能原因与替代方案：") print(" 1. API易 的该模型可能不支持 images.edit → 改用网页端手动上传两张图") print(" 2. 换支持参考图的模型重试: -m gemini-3-1-flash-image-preview") print(" 3. 改走纯文生图: python tools/image_gen.py -g \"<建筑内容描述>\"") print() print(" 所有错误:") for err in errors: print(f" - {err}") sys.exit(1) save_result(image_bytes, { "model": model, "prompt": fusion_prompt, "mode": "multi_reference_fusion", "style_reference": str(style_file), "content_references": [str(cf) for cf in content_files], "size": "1024x1024", "timestamp": datetime.now().isoformat(), }, subdir=subdir) # --------------------------------------------------------------------------- # CLI 入口 # --------------------------------------------------------------------------- def main(): parser = argparse.ArgumentParser( description="ChildHood AI Art Generation Tool — 通过 API易 生图", formatter_class=argparse.RawDescriptionHelpFormatter, epilog="""\ 示例: python image_gen.py --test # 测试连通性 python image_gen.py -g "千禧年中国城郊的小卖部门口" # 文生图 python image_gen.py -g "a red apple" --no-style # 不加风格前缀 python image_gen.py --calibrate # 风格校准 python image_gen.py --fuse --style ref.png --content photo.jpg # 双参考融合 python image_gen.py --fuse -s style.png -c photo.jpg -p "自定义" # 自定义 prompt 环境变量 (ai-art-research/.env): APIYI_API_KEY API易 令牌 (必填) APIYI_BASE_URL API 地址 (默认 https://api.apiyi.com/v1) APIYI_MODEL 默认模型 (默认 gpt-image-2) """, ) # 互斥的操作模式 mode = parser.add_mutually_exclusive_group() mode.add_argument("--test", action="store_true", help="测试 API 连通性") mode.add_argument("-g", "--generate", type=str, metavar="PROMPT", help="文生图") mode.add_argument("--calibrate", action="store_true", help="风格校准（5 个测试物体）") mode.add_argument("--fuse", action="store_true", help="双参考图融合模式") # 参数 parser.add_argument("-s", "--style", type=str, help="风格参考图路径（配合 --fuse）") parser.add_argument("-c", "--content", action="append", type=str, help="内容参考图路径（配合 --fuse，可重复传多张：首张为整楼，后续为细节特写如晾衣架）") parser.add_argument("-p", "--prompt", type=str, help="自定义 prompt（短）") parser.add_argument("-pf", "--prompt-file", type=str, help="从 .txt 文件读取 prompt（配合 --fuse，适合长 prompt，避免命令行转义噪音）") parser.add_argument("-m", "--model", type=str, help=f"模型 (默认: {DEFAULT_MODEL})") parser.add_argument("--no-style", action="store_true", help="文生图不加风格前缀") parser.add_argument("--size", type=str, default="1024x1024", help="图片尺寸") parser.add_argument("--quality", type=str, default="medium", choices=["low", "medium", "high"], help="质量") parser.add_argument("--subdir", type=str, default="", help="输出子目录名") args = parser.parse_args() # 无参数时显示帮助 if len(sys.argv) == 1: parser.print_help() return client = get_client() if args.test: cmd_test(client) elif args.generate: cmd_generate( client, args.generate, use_style=not args.no_style, size=args.size, quality=args.quality, model=args.model, subdir=args.subdir or "文生图", ) elif args.calibrate: cmd_calibrate(client, model=args.model) elif args.fuse: if not args.style or not args.content: parser.error("双参考融合需要 --style (-s) 和 --content (-c) 参数") if args.prompt_file: pf = Path(args.prompt_file) if not pf.exists(): parser.error(f"prompt 文件不存在: {pf}") fuse_prompt = pf.read_text(encoding="utf-8") print(f" 📄 从文件读取 prompt: {pf.name}") else: fuse_prompt = args.prompt cmd_fuse( client, args.style, args.content, prompt=fuse_prompt, model=args.model, subdir=args.subdir, ) if __name__ == "__main__": main()