guanghulab/video-ai-system/tools/char004_gen.py

#!/usr/bin/env python3
"""
CHAR-004 诸葛风 · VA-05-001-R2 · 纯文生图执行脚本
分两步：先锁服装 → 再调面部
禁用任何图像参考，全程纯文字prompt
"""
import os, sys, json, time, base64, subprocess
from pathlib import Path

sys.path.insert(0, os.path.dirname(__file__))
from secrets_loader import secret, endpoint

# ── API 配置 ──
API_KEY = secret("SC-001")
BASE_URL = endpoint("EPT-001")
IMAGE_MODEL = "doubao-seedream-4-0-250828"  # 已验证通过

# ── 路径配置 ──
REPO_ROOT = Path(__file__).resolve().parents[1]
JZAO_BASE = Path("/Volumes/JZAO/铸渊-ICE-GL-ZY001/OUT-输出/图片/zai-fu-fei-xiu-xian/ep01")
CANDIDATES_DIR = REPO_ROOT / "assets/candidates/D157-VA-05-001-R2" / "CHAR-004-ZhugeFeng"

# ── R2规范 · 全文字prompt ──
CHAR004_PROMPT = """
3D动漫卡通渲染，国风玄幻动态漫风格，画面质感与光影水平统一。

男性，16岁，清瘦挺拔，全身立绘，正面站立姿势。身高约175cm，肩宽偏窄，身形单薄但有力量感，无佝偻萎靡感。

面部：小麦色肤色，面部轮廓棱角分明，剑眉星目，眼神锐利有韧劲，鼻梁挺直，嘴唇偏薄，下颌线清晰。气质偏冷硬，少年气强，与温润型角色明确区分。

发型：纯黑色短发，高束半马尾，发尾有零散碎发。用一根普通深棕色木簪固定，无华丽发饰。额前散落几缕碎发，整体整洁不凌乱。

上衣：灰蓝色粗布交领短打，右衽，袖口收窄贴合手腕。衣服洗得微微褪色，领口、袖口边缘有轻微磨毛痕迹。无破洞、无污渍、无补丁，干净整洁。腰间系一根深棕色粗麻绳腰带，系法简单。

下装：深灰蓝色同色系粗布长裤，裤脚收紧塞进靴筒，膝盖位置有轻微磨白痕迹。无破损、无污渍。

鞋子：黑色粗布短靴，厚鞋底，鞋头有轻微磨损痕迹，干净无泥污。

配饰：无玉佩、香囊、首饰。仅腰间麻绳上挂一个拳头大的深灰色旧布包，布包平整无破损。

整体：清贫但不落魄，整洁有骨气。纯白背景，全身从头到鞋完整展示，画面干净。

负面：不要破洞，不要污渍，不要补丁，不要乞丐，不要乞丐服，不要麻绳缠身，不要短袖，不要破损衣服，不要无鞋，不要半身。
""".strip().replace("\n", " ")

def call_seedream(prompt_text, seed=None):
    """调用Seedream文生图API"""
    url = f"{BASE_URL}/images/generations"
    headers = {
        "Authorization": f"Bearer {API_KEY}",
        "Content-Type": "application/json"
    }
    payload = {
        "model": IMAGE_MODEL,
        "prompt": prompt_text,
        "size": "1440x2560",
        "n": 1,
    }
    if seed:
        payload["seed"] = seed

    print(f"  → POST {url}")
    print(f"  → model: {IMAGE_MODEL}")
    
    # 使用curl调用（避免requests SSL问题）
    payload_json = json.dumps(payload)
    cmd = [
        "curl", "-s", "-m", "120", "--noproxy", "*",
        "-X", "POST", url,
        "-H", f"Authorization: Bearer {API_KEY}",
        "-H", "Content-Type: application/json",
        "-d", payload_json
    ]
    
    result = subprocess.run(cmd, capture_output=True, text=True, timeout=130)
    if result.returncode != 0:
        print(f"  ✗ curl error: {result.stderr}")
        return None
    
    try:
        data = json.loads(result.stdout)
    except json.JSONDecodeError:
        print(f"  ✗ JSON parse error: {result.stdout[:200]}")
        return None
    
    if "error" in data:
        print(f"  ✗ API error: {data['error']}")
        return None
    
    # Seedream4 响应结构：data[0].url
    if "data" in data and len(data["data"]) > 0:
        img_url = data["data"][0].get("url", "")
        print(f"  ✓ 生成成功: {img_url[:80]}...")
        return img_url
    
    print(f"  ✗ 未找到图片URL. 响应: {json.dumps(data)[:300]}")
    return None

def download_image(url, filepath):
    """下载图片到指定路径"""
    Path(filepath).parent.mkdir(parents=True, exist_ok=True)
    print(f"  ↓ 下载到 {filepath}")
    cmd = ["curl", "-s", "-m", "60", "-L", "--noproxy", "*", "-o", filepath, url]
    result = subprocess.run(cmd, capture_output=True, text=True, timeout=65)
    if result.returncode != 0:
        print(f"  ✗ 下载失败: {result.stderr}")
        return False
    size = os.path.getsize(filepath) if os.path.exists(filepath) else 0
    print(f"  ✓ 已保存 {size/1024:.0f}KB")
    return size > 1024  # 至少1KB

def scale_to_1080(src_path, dst_path):
    """1440x2560 → 1080x1920 等比缩放"""
    print(f"  ⇄ 缩放 {src_path} → {dst_path}")
    cmd = [
        "sips", "-z", "1920", "1080",
        str(src_path),
        "--out", str(dst_path)
    ]
    result = subprocess.run(cmd, capture_output=True, text=True)
    return result.returncode == 0

def main():
    print("=" * 60)
    print("CHAR-004 诸葛风 · VA-05-001-R2 · 纯文生图")
    print(f"模型: {IMAGE_MODEL}")
    print(f"分辨率: 1440x2560 → 1080x1920")
    print(f"备选数: 4")
    print(f"策略: 全程纯文字 · 禁用图像参考")
    print("=" * 60)

    if not API_KEY:
        print("✗ 未找到JIMENG_API_KEY，退出")
        sys.exit(1)

    # ── 第一步：生成4张（锁服装身形） ──
    print("\n── 第一步：生成4张候选 ──")
    
    timestamp = time.strftime("%Y%m%d-%H%M%S")
    run_dir = JZAO_BASE / "D157-R2" / timestamp
    run_dir.mkdir(parents=True, exist_ok=True)
    
    candidates_dir = CANDIDATES_DIR
    candidates_dir.mkdir(parents=True, exist_ok=True)
    
    results = []
    seeds = [42, 137, 256, 399]  # 不同seed产生不同变体
    
    for i, seed in enumerate(seeds):
        print(f"\n[{i+1}/4] 生成候选 {i+1} (seed={seed})")
        
        # 微调prompt增加面部变化
        variations = [
            "面部角度：正脸直视镜头",
            "面部角度：微微侧头，目光坚毅",
            "面部角度：略微仰视，眼神向上",
            "面部角度：3/4侧脸，侧视左前方",
        ]
        full_prompt = f"{CHAR004_PROMPT} {variations[i]}"
        
        img_url = call_seedream(full_prompt, seed=seed)
        if not img_url:
            print(f"  ✗ 候选 {i+1} 生成失败")
            results.append({"candidate": i+1, "status": "failed", "error": "API call failed"})
            continue
        
        # 保存原始1440×2560
        raw_name = f"CHAR-004-R2-candidate-{i+1:02d}-1440x2560.jpg"
        raw_path = run_dir / raw_name
        if not download_image(img_url, raw_path):
            results.append({"candidate": i+1, "status": "failed", "error": "download failed"})
            continue
        
        # 缩放至1080×1920
        scaled_name = f"CHAR-004-R2-candidate-{i+1:02d}-1080x1920.jpg"
        scaled_path = candidates_dir / scaled_name
        if not scale_to_1080(raw_path, scaled_path):
            results.append({"candidate": i+1, "status": "failed", "error": "scale failed"})
            continue
        
        results.append({
            "candidate": i+1,
            "status": "generated",
            "raw": str(raw_path),
            "scaled": str(scaled_path),
            "seed": seed
        })
        print(f"  ✓ 候选 {i+1} 完成")
        time.sleep(2)  # 避免请求过快
    
    # ── 汇总 ──
    print("\n── 生成汇总 ──")
    success = sum(1 for r in results if r["status"] == "generated")
    failed = len(results) - success
    print(f"成功: {success}/4")
    print(f"失败: {failed}/4")
    
    if failed > 0:
        for r in results:
            if r["status"] != "generated":
                print(f"  ✗ 候选{r['candidate']}: {r.get('error', 'unknown')}")
    
    # 写入元数据
    meta = {
        "spec": "VA-05-001-R2",
        "asset_id": "CHAR-004-ZhugeFeng",
        "timestamp": timestamp,
        "model": IMAGE_MODEL,
        "resolution_raw": "1440x2560",
        "resolution_final": "1080x1920",
        "strategy": "pure_text_prompt_no_image_reference",
        "results": results
    }
    meta_path = candidates_dir / "generation-meta.json"
    with open(meta_path, "w") as f:
        json.dump(meta, f, ensure_ascii=False, indent=2)
    print(f"\n元数据已保存: {meta_path}")
    
    return 0 if success >= 4 else 1

if __name__ == "__main__":
    sys.exit(main())