guanghulab/video-ai-system/tools/qc_char004_r2.py

#!/usr/bin/env python3
"""CHAR-004 R2 候选质检 · Qwen-VL · VA-05-001-R2标准"""
import sys, os, json, base64, subprocess

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

API_KEY = secret("SC-004")
EP = endpoint("EPT-002")
FALLBACK_EP = "https://dashscope.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation"

CANDIDATES_DIR = os.path.expanduser("~/guanghulab/video-ai-system/assets/candidates/D157-VA-05-001-R2/CHAR-004-ZhugeFeng")
CANDIDATES = [f"{CANDIDATES_DIR}/CHAR-004-R2-candidate-{i:02d}-1080x1920.jpg" for i in range(1, 5)]

QC_PROMPT = """你是动画资产质量检查员。检查这张CHAR-004诸葛风立绘候选图，按以下标准逐项打分。

## R2规范必查项

1. 服装完整性（一票否决）：
   - 裤子膝盖是否有破洞/破损？是/否
   - 衣服下摆是否有毛边/破损/撕裂？是/否
   - 鞋子是否有泥污/脏污？是/否
   - 服装是否有污渍？是/否
   - 是否有补丁？是/否
   - 整体是否"乞丐化"（像乞丐而不是清贫少年）？

2. 服装部件准确性：
   - 上衣是否为灰蓝色交领短打、右衽？是/否
   - 是否系深棕色粗麻绳腰带？是/否
   - 裤子是否深灰蓝色、裤脚塞进靴子？是/否
   - 鞋子是否黑色粗布短靴？是/否
   - 腰间是否有旧布包？是/否
   - 有无不该有的配饰（玉佩、香囊等）？是/否

3. 面部质量：
   - 五官是否清晰无崩坏？是/否
   - 是否少年感强、气质偏冷硬？是/否
   - 有没有和苏白（温润仙气型）相似？是/否

4. 整体：
   - 人物肢体比例是否正常？是/否
   - 画面是否清晰无噪点？是/否
   - 是否全身从头到鞋完整展示？是/否

输出严格JSON，不要额外文字：
{
  "candidate": N,
  "pass": true/false,
  "score": 0-100,
  "hole_in_pants": true/false,
  "frayed_hem": true/false,
  "dirty_shoes": true/false,
  "stains": true/false,
  "patches": true/false,
  "beggar_look": true/false,
  "correct_top": true/false,
  "correct_belt": true/false,
  "correct_pants": true/false,
  "correct_shoes": true/false,
  "has_bag": true/false,
  "extra_accessories": true/false,
  "face_intact": true/false,
  "cold_temperament": true/false,
  "similar_to_su_bai": true/false,
  "anatomy_ok": true/false,
  "clear_image": true/false,
  "full_body": true/false,
  "summary": "一句话总结"
}"""

def encode_image(path):
    with open(path, "rb") as f:
        b64 = base64.b64encode(f.read()).decode()
    ext = path.rsplit(".", 1)[-1].lower()
    mime = {"jpg": "jpeg", "jpeg": "jpeg", "png": "png"}.get(ext, "jpeg")
    return f"data:image/{mime};base64,{b64}"

def call_qwen(image_b64):
    body = json.dumps({
        "model": "qwen-vl-max",
        "input": {
            "messages": [{
                "role": "user",
                "content": [
                    {"image": image_b64},
                    {"text": QC_PROMPT}
                ]
            }]
        }
    })
    
    endpoints = [EP, FALLBACK_EP]
    for ep in endpoints:
        try:
            proc = subprocess.run(
                ["curl", "-s", "-m", "60", "--noproxy", "*",
                 "-X", "POST", ep,
                 "-H", f"Authorization: Bearer {API_KEY}",
                 "-H", "Content-Type: application/json",
                 "--data-binary", "@-"],
                input=body, capture_output=True, text=True, timeout=65
            )
            if proc.returncode != 0 or not proc.stdout.strip():
                continue
            resp = json.loads(proc.stdout)
            if "output" in resp:
                content = resp["output"]["choices"][0]["message"]["content"][0]["text"]
                # Clean markdown wrappers
                if "```" in content:
                    content = content.split("```")[1]
                    if content.startswith("json"):
                        content = content[4:]
                    content = content.split("```")[0]
                return json.loads(content.strip())
            elif "error" in resp:
                print(f"  [Qwen error] {resp['error']}", file=sys.stderr)
                continue
        except Exception as e:
            print(f"  [fail {ep}] {e}", file=sys.stderr)
            continue
    return None

def main():
    print("=" * 60)
    print("CHAR-004 R2 候选质检 · Qwen-VL · VA-05-001-R2标准")
    print("=" * 60)
    
    results = []
    for i, path in enumerate(CANDIDATES):
        print(f"\n[{i+1}/4] 质检候选 {i+1}...")
        img = encode_image(path)
        qc = call_qwen(img)
        
        if qc:
            print(f"  score={qc.get('score')} pass={qc.get('pass')}")
            print(f"  破洞={qc.get('hole_in_pants')} 毛边={qc.get('frayed_hem')} 脏鞋={qc.get('dirty_shoes')}")
            print(f"  乞丐化={qc.get('beggar_look')} 和苏白像={qc.get('similar_to_su_bai')}")
            print(f"  {qc.get('summary')}")
            results.append(qc)
        else:
            print(f"  ✗ Qwen-VL调用失败")
            results.append({"candidate": i+1, "pass": False, "error": "Qwen-VL call failed"})
    
    # 汇总
    print("\n" + "=" * 60)
    print("质检汇总")
    print("=" * 60)
    
    passed = sum(1 for r in results if r.get("pass"))
    print(f"通过: {passed}/4")
    
    for r in results:
        if not r.get("pass"):
            issues = []
            if r.get("hole_in_pants"): issues.append("裤子破洞")
            if r.get("frayed_hem"): issues.append("下摆毛边")
            if r.get("dirty_shoes"): issues.append("鞋子脏污")
            if r.get("stains"): issues.append("污渍")
            if r.get("patches"): issues.append("补丁")
            if r.get("beggar_look"): issues.append("乞丐化")
            if r.get("extra_accessories"): issues.append("多余配饰")
            print(f"  候选{r.get('candidate')} ✗: {', '.join(issues) if issues else 'other'}")
    
    # 写入JSON
    out_path = os.path.join(CANDIDATES_DIR, "QC-R2-BATCH.json")
    report = {
        "spec": "VA-05-001-R2",
        "asset_id": "CHAR-004-ZhugeFeng",
        "qc_tool": "Qwen-VL (qwen-vl-max)",
        "results": results,
        "summary": {
            "passed": passed,
            "total": 4,
            "verdict": "APPROVED" if passed >= 4 else "REJECTED"
        }
    }
    with open(out_path, "w") as f:
        json.dump(report, f, ensure_ascii=False, indent=2)
    print(f"\nQC报告已保存: {out_path}")
    
    return 0 if passed >= 4 else 1

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