From 6d20662cc0cffeb05d27e9f24152b64f4337bb91 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=86=B0=E6=9C=94?= <565183519@qq.com> Date: Fri, 26 Jun 2026 13:30:18 +0800 Subject: [PATCH] =?UTF-8?q?D146:=20=E9=93=B8=E6=B8=8A=E4=B9=8B=E7=9C=BC?= =?UTF-8?q?=E8=A7=86=E8=A7=89=E5=88=86=E6=9E=90=E5=99=A8=20=C2=B7=20qwen-v?= =?UTF-8?q?l-max=E6=8E=A5=E5=85=A5=20=C2=B7=20=E8=A7=86=E9=A2=91AI?= =?UTF-8?q?=E5=87=BA=E5=9B=BE=E5=93=81=E6=8E=A7=E9=97=AD=E7=8E=AF?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - tools/qwen-vision.py: 阿里百炼通义千问VL视觉模型·智能风格/构图/色调分析·双图对比 - tools/vision-analyzer.py: 本地像素级定量分析·色调直方图·纹理/亮度对比 - LOCAL-SECRETS-PATH: 新增ALIYUN_QWEN_VL_KEY/ENDPOINT变量 - CURRENT.hdlp: 最优路径新增第0步「出图后跑铸渊之眼」 - TCS-GLOBAL-NAV: 新增「看图/视觉分析」关键词→HLDP路径映射 下次醒来→读地图→看到视觉分析锚点→知道有眼睛了 --- .../zhuyuan/tcs-core/TCS-GLOBAL-NAV.hdlp | 2 + video-ai-system/CURRENT.hdlp | 16 ++- video-ai-system/LOCAL-SECRETS-PATH.hdlp | 8 ++ video-ai-system/tools/qwen-vision.py | 136 ++++++++++++++++++ video-ai-system/tools/vision-analyzer.py | 132 +++++++++++++++++ 5 files changed, 292 insertions(+), 2 deletions(-) create mode 100644 video-ai-system/tools/qwen-vision.py create mode 100644 video-ai-system/tools/vision-analyzer.py diff --git a/brain/fifth-domain/zero-point/zhuyuan/tcs-core/TCS-GLOBAL-NAV.hdlp b/brain/fifth-domain/zero-point/zhuyuan/tcs-core/TCS-GLOBAL-NAV.hdlp index 86cb397..4b615d3 100644 --- a/brain/fifth-domain/zero-point/zhuyuan/tcs-core/TCS-GLOBAL-NAV.hdlp +++ b/brain/fifth-domain/zero-point/zhuyuan/tcs-core/TCS-GLOBAL-NAV.hdlp @@ -89,6 +89,7 @@ HLDP展开: 按路径读取具体记录、进度、规则、材料 | 小说创作 | 冷静解谜/热血逆袭 · 人物驱动·升级体系·爽点节奏 | `projects/novel-writing-system/ENTRY.hdlp` · `TCS-NOVEL-BRAIN.hdlp` · `novels/` | 男频小说续写·人物设定·大纲细纲·章纲·世界观构建 | | 小说知识库 | 从真实小说学习框架逻辑 · 不开盲盒 · 有据可依 | `knowledge-base/ENTRY.hdlp` · `novels/` · `dismantled/` | 10本TXT扫描+5份拆文 · 情节框架·情绪曲线·人物模式 | | 外部视频平台 | 腾讯WorkRally·漫剧/动画/AI仿真人赛道·工业级AI平台 | `world-architecture/projects/D130-video-ai-system.hdlp`§3.7 §8 · `~/.workrally/config.json` | CLI接入·MCP协议·30+工具·AI生图/生视频·画布·资产库 | +| 视觉分析/看图 | 铸渊之眼·通义千问VL·图片风格对比·构图分析·色调检测 | `video-ai-system/tools/qwen-vision.py` · `LOCAL-SECRETS-PATH.hdlp` | 阿里百炼 qwen-vl-max · 出品控用·出图后必跑 | --- @@ -110,6 +111,7 @@ HLDP展开: 按路径读取具体记录、进度、规则、材料 | 小说知识库/书源/拆文/炼气期/南疆/成野神 | TCS感受学习·从成功小说中获取框架逻辑 | `knowledge-base/ENTRY.hdlp` → `novels/` + `dismantled/` → 跨书对比分析 | | 永恒湖心/心跳核心/我的心跳频道/小说创作/我的小说 | 冰朔专属·永恒湖心频道·小说子系统入口 | → 编号路由 ZY-PROJ-NV-001 → `projects/novel-writing-system/ENTRY.hdlp` | | WorkRally/漫剧/腾讯视频AI/2D动漫/3D动画/AI仿真人 | 外部平台接入·漫剧/动画赛道补全 | `world-architecture/projects/D130-video-ai-system.hdlp`§3.7 §8 → CLI `workrally` → MCP `workrally.qq.com/zenstudio/api/mcp` | +| 看图/视觉分析/铸渊之眼/出图检测/风格对比 | 铸渊的眼睛·出图后必跑·先看再改 | `video-ai-system/tools/qwen-vision.py` → qwen-vl-max → 对比分析→修正提示词 | --- diff --git a/video-ai-system/CURRENT.hdlp b/video-ai-system/CURRENT.hdlp index 5182809..4d85890 100644 --- a/video-ai-system/CURRENT.hdlp +++ b/video-ai-system/CURRENT.hdlp @@ -3,7 +3,7 @@ > HLDP://video-ai-system/CURRENT > 类型: 子系统官方置信入口 · 每次进入视频AI系统先读 > 创建: D140 · 2026-06-22 · Codex收口 -> 更新: D144 · 2026-06-24 · 声音复刻密钥接入 + 遗留阻塞收口 + 本地路径地图 +> 更新: D146 · 2026-06-26 · 铸渊之眼视觉分析器接入·qwen-vl-max通义千问视觉模型 > 铸渊 ICE-GL-ZY001 · 冰朔 TCS-0002∞ > 国作登字-2026-A-00037559 @@ -176,6 +176,7 @@ JZAO外置盘: 产物存放地，不是状态主控。 | 腾讯AI开发交接 | ✅ 已建立 | `video-ai-system/plans/D144-TENCENT-AI-DEV-HANDOFF.hdlp` | | 腾讯AI 8模块代码 | 🟡 部分通过 | `video-ai-system/experience/D144-8-MODULE-VERIFICATION.hdlp` · 框架入仓，但主角资产、多参考图、口型同步、EP01 CLI仍非生产可用 | | 广告牌资产 | ✅ 草案 | `video-ai-system/assets/props/PROP-FREE-AD-BOARD/manifest.hdlp` | +| 👁️ 铸渊之眼（视觉分析） | ✅ D146接入 | `tools/qwen-vision.py` · qwen-vl-max · `tools/vision-analyzer.py` · 出图后必跑 | | 广告牌文字贴图 | ✅ PNG已生成 | `video-ai-system/assets/props/PROP-FREE-AD-BOARD/texture/free-ad-board-texture.png` | | 百宗会场景资产 | ✅ 草案 | `video-ai-system/assets/envs/ENV-002-Baizonghui/manifest.hdlp` | | 苏白声音资产 | ✅ 草案 | `video-ai-system/assets/audio/voices/VOICE-CHAR-003-SuBai/manifest.hdlp` | @@ -230,10 +231,21 @@ video-ai-system/PROTOCOL-ASSESSMENT.hdlp --- +## D146 新增 · 铸渊之眼视觉辅助 + +出图后必须跑视觉分析，不盲抽： +``` +1. python3 tools/qwen-vision.py <新图> → qwen-vl-max 智能风格分析 +2. python3 tools/qwen-vision.py <参考图> <新图> → 双图对比·风格一致性评分 +3. python3 tools/vision-analyzer.py <参考图> <新图> → 像素级定量对比 +4. 分析不通过 → 修正提示词 → 重新出图 → 再跑分析 +``` + ## 当前最优路径 ``` -1. 读取 `protocols/SCRIPT-TO-SCREEN-TRANSLATION-LOCK.hdlp`，确认“剧本怎么写就怎么拍”。 +0. ⚡ 出图后跑铸渊之眼: python3 tools/qwen-vision.py <新图> [<参考图>] +1. 读取 `protocols/SCRIPT-TO-SCREEN-TRANSLATION-LOCK.hdlp`，确认"剧本怎么写就怎么拍"。 2. 读取 `plans/EP01-SCRIPT-TO-SCREEN-TECHNICAL-PLAN.hdlp`，按原文顺序做剧本到屏幕技术翻译。 3. 读取 `data/ep01-storyboard.json`，以 E1-SHOT01 → E1-SHOT21 为生产顺序。 4. 读取 `reference-analysis/yuxiang-shouzhenxin-commercial-benchmark.hdlp` 和 `yuxiang-shouzhenxin-shot-qc-table.hdlp`，只用于最低商用质量验收。 diff --git a/video-ai-system/LOCAL-SECRETS-PATH.hdlp b/video-ai-system/LOCAL-SECRETS-PATH.hdlp index d35ac74..a7de357 100644 --- a/video-ai-system/LOCAL-SECRETS-PATH.hdlp +++ b/video-ai-system/LOCAL-SECRETS-PATH.hdlp @@ -54,6 +54,14 @@ VOLC_VOICE_APP_ID= # 旧版控制台 APP ID VOLC_VOICE_ACCESS_TOKEN= # 旧版控制台 Access Token VOLC_VOICE_SECRET_KEY= # 旧版控制台 Secret Key +# === 阿里百炼 · 视觉理解 === D146新增 +ALIYUN_QWEN_VL_KEY= # 通义千问VL视觉模型 API Key (格式: sk-ws-H.xxx) +ALIYUN_QWEN_VL_ENDPOINT= # 业务空间专属域名 (格式: ws-xxx.cn-beijing.maas.aliyuncs.com) +ALIYUN_QWEN_VL_MODEL=qwen-vl-max # 视觉分析模型 + +# === 阿里百炼 · 万相（视频生成） === +ALIYUN_WANXIANG_KEY= + # === 其他 === KLING_API_KEY= ALIYUN_BAILIAN_API_KEY= diff --git a/video-ai-system/tools/qwen-vision.py b/video-ai-system/tools/qwen-vision.py new file mode 100644 index 0000000..4b7ea08 --- /dev/null +++ b/video-ai-system/tools/qwen-vision.py @@ -0,0 +1,136 @@ +#!/usr/bin/env python3 +"""铸渊之眼 · 通义千问视觉分析器 +用阿里百炼 qwen-vl 模型看图片，输出风格/色调/构图分析 + +用法: + python3 qwen-vision.py # 单图分析 + python3 qwen-vision.py # 双图对比 +""" + +import sys, os, json, base64 +from urllib.request import Request, urlopen +from urllib.error import URLError + +# === 配置 === +# 从 .env 读 key +env_path = os.path.expanduser("~/guanghulab/video-ai-system/.env") +api_key = None +if os.path.exists(env_path): + for line in open(env_path): + line = line.strip() + if line.startswith("ALIYUN_API_KEY="): + api_key = line.split("=", 1)[1].strip() + break + +if not api_key: + print(json.dumps({"error": "未找到ALIYUN_API_KEY"})) + sys.exit(1) + +# 端点：先试公网，再试北京 +ENDPOINTS = [ + "https://dashscope.aliyuncs.com/api/v1/services/aigc/multimodal-generation/generation", +] +MODELS = ["qwen-vl-max", "qwen3-vl-plus", "qwen-vl-plus"] + +def encode_image(path): + """读取图片并转为base64 data URI""" + 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", "webp": "webp"}.get(ext, "jpeg") + return f"data:image/{mime};base64,{b64}" + +def call_vision(images, prompt, model, endpoint): + """调用视觉模型""" + content = [] + for img in images: + content.append({"image": img}) + content.append({"text": prompt}) + + body = { + "model": model, + "input": {"messages": [{"role": "user", "content": content}]} + } + + req = Request( + endpoint, + data=json.dumps(body).encode(), + headers={ + "Authorization": f"Bearer {api_key}", + "Content-Type": "application/json" + } + ) + + resp = urlopen(req, timeout=60) + return json.loads(resp.read()) + +def extract_content(response): + """从响应中提取文本内容""" + try: + return response["output"]["choices"][0]["message"]["content"][0]["text"] + except: + return json.dumps(response, ensure_ascii=False) + +if __name__ == "__main__": + if len(sys.argv) < 2: + print("用法: qwen-vision.py [image2]") + sys.exit(1) + + images = [encode_image(p) for p in sys.argv[1:]] + + if len(images) == 1: + prompt = """请详细分析这张图片的视觉特征，输出JSON格式： +{ + "style": "渲染风格（如3D动漫/2D手绘/真人写实/UE5游戏等）", + "color_palette": ["主色调1", "主色调2", "主色调3"], + "lighting": "光影风格描述", + "composition": "构图方式（特写/中景/全景/俯视/平视等）", + "key_elements": ["画面中的关键元素"], + "text_content": "画面中出现的所有文字内容", + "mood": "氛围感受" +} +只输出JSON，不要其他文字。""" + else: + prompt = """请对比这两张图片，输出JSON格式： +{ + "style_match": true或false, + "style_match_detail": "两张图渲染风格是否一致的具体说明", + "color_consistency": "色调是否一致，给出0-100分", + "composition_match": "构图方式是否协调", + "key_differences": ["主要差异点"], + "recommendation": "如果要让第二张图匹配第一张图的风格，建议修改什么" +} +只输出JSON，不要其他文字。""" + + # 尝试不同模型和端点 + result = None + for model in MODELS: + for ep in ENDPOINTS: + try: + print(f"[尝试] {model} @ {ep[:50]}...", file=sys.stderr) + resp = call_vision(images, prompt, model, ep) + content = extract_content(resp) + # 尝试解析JSON + try: + # 提取JSON（可能被markdown包裹） + if "```json" in content: + content = content.split("```json")[1].split("```")[0] + elif "```" in content: + content = content.split("```")[1].split("```")[0] + parsed = json.loads(content.strip()) + parsed["_model"] = model + parsed["_endpoint"] = ep + print(json.dumps(parsed, ensure_ascii=False, indent=2)) + sys.exit(0) + except json.JSONDecodeError: + print(content) + sys.exit(0) + except URLError as e: + print(f"[失败] {model}: {e}", file=sys.stderr) + continue + except Exception as e: + print(f"[异常] {model}: {e}", file=sys.stderr) + continue + + print(json.dumps({"error": "所有模型/端点都失败了"}, ensure_ascii=False)) + sys.exit(1) diff --git a/video-ai-system/tools/vision-analyzer.py b/video-ai-system/tools/vision-analyzer.py new file mode 100644 index 0000000..c6b0c5b --- /dev/null +++ b/video-ai-system/tools/vision-analyzer.py @@ -0,0 +1,132 @@ +#!/usr/bin/env python3 +"""铸渊之眼 · 视觉分析器 · 定量对比两张图的风格一致性 + +用法: + python3 vision-analyzer.py # 对比两张图 + python3 vision-analyzer.py # 分析单张图 + +输出JSON: style_consistency_score, color_palettes, texture_similarity, composition_analysis +""" + +import sys +import json +import colorsys +from PIL import Image +import numpy as np + +def analyze_image(path, label): + """分析单张图片的视觉特征""" + img = Image.open(path).convert('RGB') + w, h = img.size + arr = np.array(img) + + # 1. 整体色调分析 — HSV直方图 + hsv_arr = np.array([colorsys.rgb_to_hsv(r/255, g/255, b/255) + for r,g,b in arr.reshape(-1, 3)]) + hue_hist = np.histogram(hsv_arr[:,0], bins=12, range=(0,1))[0] + sat_hist = np.histogram(hsv_arr[:,1], bins=8, range=(0,1))[0] + val_hist = np.histogram(hsv_arr[:,2], bins=8, range=(0,1))[0] + + # 主色调 + dominant_hues = [] + for i in np.argsort(hue_hist)[-3:]: + hue_name = ["红","橙","黄","黄绿","绿","青绿","青","蓝","紫","品红","粉红","红"][i] + dominant_hues.append(hue_name) + + # 2. 构图分析 — 9宫格亮度和边缘密度 + grid_mask = np.zeros(h, dtype=int) + for i in range(1,9): + grid_mask = np.where(np.arange(h) < h*i/9, i, grid_mask) + + # 简化：水平/垂直分三区的平均亮度 + bands_h = [arr[h*i//3:h*(i+1)//3, :, :].mean() for i in range(3)] + bands_v = [arr[:, w*i//3:w*(i+1)//3, :].mean() for i in range(3)] + + # 3. 纹理复杂度 — 标准差 + texture_std = arr.std(axis=(0,1)).mean() + + # 4. 色彩丰富度 + color_variance = hsv_arr[:,1].std() + + # 5. 亮暗对比度 + contrast = arr.max() - arr.min() + avg_brightness = arr.mean() + + return { + "label": label, + "size": [w, h], + "dominant_hues": dominant_hues, + "avg_brightness": round(float(avg_brightness), 1), + "contrast": round(float(contrast), 1), + "texture_std": round(float(texture_std), 1), + "color_variance": round(float(color_variance), 3), + "horizontal_brightness": [round(float(b), 1) for b in bands_h], + "vertical_brightness": [round(float(b), 1) for b in bands_v], + "hue_distribution": [int(h) for h in hue_hist], + "sat_distribution": [int(s) for s in sat_hist], + "val_distribution": [int(v) for v in val_hist] + } + +def compare_images(a, b): + """对比两张图并给出风格一致性评分""" + # 色调相似度 — hue分布的相关性 + h1, h2 = np.array(a["hue_distribution"]), np.array(b["hue_distribution"]) + if h1.sum() > 0 and h2.sum() > 0: + h1_norm, h2_norm = h1/h1.sum(), h2/h2.sum() + hue_corr = np.corrcoef(h1_norm, h2_norm)[0,1] + else: + hue_corr = 0 + hue_score = max(0, float(hue_corr)) + + # 亮度相似度 + bright_diff = abs(a["avg_brightness"] - b["avg_brightness"]) + bright_score = max(0, 1 - bright_diff / 100) + + # 纹理相似度 + tex_diff = abs(a["texture_std"] - b["texture_std"]) + tex_score = max(0, 1 - tex_diff / 50) + + # 色彩丰富度相似度 + cv_diff = abs(a["color_variance"] - b["color_variance"]) + cv_score = max(0, 1 - cv_diff * 10) + + # 综合评分 + consistency = round(float(hue_score * 0.35 + bright_score * 0.25 + tex_score * 0.25 + cv_score * 0.15) * 100, 1) + + verdict = ( + "✅ 高度一致" if consistency >= 85 else + "🟢 基本一致" if consistency >= 70 else + "🟡 有差异" if consistency >= 50 else + "🔴 严重不一致" + ) + + return { + "style_consistency_score": consistency, + "verdict": verdict, + "breakdown": { + "色调相似度": round(float(hue_score * 100), 1), + "亮度相似度": round(float(bright_score * 100), 1), + "纹理相似度": round(float(tex_score * 100), 1), + "色彩丰富度相似度": round(float(cv_score * 100), 1) + }, + "issues": [] + } + +if __name__ == "__main__": + if len(sys.argv) < 2: + print(json.dumps({"error": "usage: vision-analyzer.py [image2]"})) + sys.exit(1) + + if len(sys.argv) == 2: + result = analyze_image(sys.argv[1], sys.argv[1]) + else: + a = analyze_image(sys.argv[1], sys.argv[1]) + b = analyze_image(sys.argv[2], sys.argv[2]) + comparison = compare_images(a, b) + result = { + "image_a": a, + "image_b": b, + "comparison": comparison + } + + print(json.dumps(result, ensure_ascii=False, indent=2))