Merge pull request 'feat(asset-catalog): 简化分析工具+新增汇总和AI推理脚本' (#24) from feature/asset-catalog-simplify into main

Reviewed-on: http://micecola.top:8418/starmen1265/ChildHood/pulls/24

tools/asset-catalog/ai_inference.py

@@ -0,0 +1,366 @@
+"""GLM API 批量推理脚本 — 资产语义增强。
+
+读取 asset_database.json，按类别分批调用 GLM API（OpenAI 兼容接口），
+推断每个资产的朝向、功能、交互属性等语义信息。
+
+此脚本不依赖 UE5，可离线运行。
+
+环境变量:
+  GLM_API_KEY   — API Key（必需）
+  GLM_API_BASE  — API Base URL（默认 https://open.bigmodel.cn/api/paas/v4）
+  GLM_MODEL     — 模型名（默认 glm-4-flash）
+
+用法:
+  python ai_inference.py [--input FILE] [--output FILE] [--batch-size N] [--dry-run]
+"""
+
+import sys
+import os
+import json
+import argparse
+import time
+from pathlib import Path
+
+_SCRIPT_DIR = Path(__file__).resolve().parent
+if str(_SCRIPT_DIR) not in sys.path:
+    sys.path.insert(0, str(_SCRIPT_DIR))
+
+from _common import read_json, write_json, iso_now
+
+TAG = "[AIInference]"
+
+# ---- 配置 ----
+DEFAULT_API_BASE = "https://open.bigmodel.cn/api/paas/v4"
+DEFAULT_MODEL = "glm-4-flash"
+DEFAULT_BATCH_SIZE = 20
+
+# ---- Prompt 模板 ----
+SYSTEM_PROMPT = """你是一个 UE5 游戏开发助手，专注于为 3D 模型资产提取语义信息。
+
+你会收到一批 StaticMesh 资产的几何数据（名称、尺寸、锚点、材质等），
+需要推断每个资产的以下语义字段：
+
+1. facing_direction: 主朝向方向
+   - 格式: "+X", "-X", "+Y", "-Y", null
+   - 判断依据: 根据资产类型和包围盒尺寸推断"正面"方向
+   - 对称物体（桌子、球、圆柱等）设为 null
+   - 电视/显示器: 屏幕面方向; 门: 开门方向; 椅子: 坐面方向
+
+2. surface_height_cm: 可用水平表面高度 (cm)
+   - 如桌面高度、台面高度、床面高度
+   - 对于无水平表面的物体（墙体、栏杆）设为 null
+   - 使用锚点类型判断: CENTER_BOTTOM 时直接取 max.z
+
+3. function_description: 一句话功能描述 (中文, ≤50字)
+
+4. placement_notes: 放置注意事项 (中文, ≤80字)
+   - 锚点位置提醒、旋转建议、缩放注意等
+
+5. interaction_hints: 交互提示标签数组
+   - 可选值: can_sit, can_place_items, can_open, can_interact_screen,
+     climbable, electronics, lighting, container, decoration,
+     pass_through, driveable, weapon, vegetation
+   - 可多选
+
+6. plausible_substitutes: BSP 基础体降级方案 (中文, 或 null)
+   - 当原始资产不可用时，用什么 Cube/Cylinder + 缩放 + 材质替代
+
+你必须返回严格的 JSON 数组，每个元素对应输入中的一个资产，顺序一致。
+不要输出任何 JSON 之外的内容。"""
+
+BATCH_PROMPT_TEMPLATE = """以下是 {category} 类别的 {count} 个 StaticMesh 资产数据：
+
+{assets_json}
+
+请为每个资产推断语义信息，返回 JSON 数组。每个元素包含:
+facing_direction, surface_height_cm, function_description,
+placement_notes, interaction_hints, plausible_substitutes"""
+
+
+def _get_api_config():
+    """从环境变量读取 API 配置。"""
+    api_key = os.environ.get("GLM_API_KEY", "")
+    api_base = os.environ.get("GLM_API_BASE", DEFAULT_API_BASE)
+    model = os.environ.get("GLM_MODEL", DEFAULT_MODEL)
+
+    if not api_key:
+        print(f"{TAG} 错误: 未设置 GLM_API_KEY 环境变量")
+        print(f"{TAG}   设置方法: set GLM_API_KEY=your_key_here")
+        return None
+
+    return {
+        "api_key": api_key,
+        "api_base": api_base,
+        "model": model,
+    }
+
+
+def _classify_batch_category(tags: dict) -> str:
+    """根据 tags 推断批次类别名。"""
+    style = tags.get("style", [])
+    if not style:
+        return "未分类"
+    # 取第一个非空 style tag 作为类别
+    return style[0] if style[0] != "unknown" else "未分类"
+
+
+def _prepare_batches(assets: dict, batch_size: int) -> list:
+    """将资产按类别分批。
+
+    Returns:
+        list of (category_name, batch_index, [asset_keys])
+    """
+    # 按类别分组
+    by_category = {}
+    for name, data in assets.items():
+        cat = _classify_batch_category(data.get("tags", {}))
+        by_category.setdefault(cat, []).append(name)
+
+    # 每个类别内按 batch_size 切分
+    batches = []
+    for cat, keys in sorted(by_category.items()):
+        for i in range(0, len(keys), batch_size):
+            batch_keys = keys[i:i + batch_size]
+            batches.append((cat, i // batch_size + 1, batch_keys))
+
+    return batches
+
+
+def _format_asset_for_prompt(name: str, data: dict) -> dict:
+    """提取 prompt 中需要的精简字段。"""
+    bb = data.get("bounding_box", {})
+    pivot = data.get("pivot", {})
+    tags = data.get("tags", {})
+
+    return {
+        "name": name,
+        "size_cm": bb.get("size"),
+        "pivot_type": pivot.get("type"),
+        "pivot_offset_xy": pivot.get("offset_xy"),
+        "material_count": len(data.get("material_slots", [])),
+        "socket_count": len(data.get("sockets", [])),
+        "tags_style": tags.get("style", []),
+    }
+
+
+def _call_glm_api(config: dict, category: str, assets_json: str) -> list:
+    """调用 GLM API 进行推理。
+
+    Returns:
+        推理结果列表，失败时返回空列表
+    """
+    try:
+        from openai import OpenAI
+    except ImportError:
+        print(f"{TAG} 错误: 需要安装 openai 包")
+        print(f"{TAG}   pip install openai")
+        return []
+
+    client = OpenAI(
+        api_key=config["api_key"],
+        base_url=config["api_base"],
+    )
+
+    user_prompt = BATCH_PROMPT_TEMPLATE.format(
+        category=category,
+        count=assets_json.count('"name"'),
+        assets_json=assets_json,
+    )
+
+    try:
+        response = client.chat.completions.create(
+            model=config["model"],
+            messages=[
+                {"role": "system", "content": SYSTEM_PROMPT},
+                {"role": "user", "content": user_prompt},
+            ],
+            temperature=0.1,
+            response_format={"type": "json_object"},
+        )
+    except Exception as e:
+        print(f"{TAG} API 调用失败: {e}")
+        return []
+
+    content = response.choices[0].message.content.strip()
+    if not content:
+        print(f"{TAG} API 返回空内容")
+        return []
+
+    # 解析 JSON
+    try:
+        result = json.loads(content)
+        # 如果返回的是 {"results": [...]} 形式，提取数组
+        if isinstance(result, dict):
+            for key in ("results", "data", "assets", "items"):
+                if key in result and isinstance(result[key], list):
+                    return result[key]
+            # 如果找不到数组，可能是单条结果被包在 dict 里
+            print(f"{TAG} 警告: 返回 JSON 不是数组，尝试整体作为单条: {list(result.keys())}")
+            return [result]
+        if isinstance(result, list):
+            return result
+    except json.JSONDecodeError as e:
+        print(f"{TAG} JSON 解析失败: {e}")
+        print(f"{TAG} 原始响应前 200 字: {content[:200]}")
+        return []
+
+    return []
+
+
+def run_inference(
+    input_file: Path,
+    output_file: Path,
+    batch_size: int = DEFAULT_BATCH_SIZE,
+    dry_run: bool = False,
+) -> dict:
+    """执行批量 AI 推理。
+
+    Args:
+        input_file: asset_database.json 路径
+        output_file: 输出路径（通常与 input 相同，就地更新）
+        batch_size: 每批资产数量
+        dry_run: 只打印批次信息，不调用 API
+
+    Returns:
+        统计信息 dict
+    """
+    # 读取数据库
+    if not input_file.exists():
+        print(f"{TAG} 错误: {input_file} 不存在，请先运行 build_database")
+        return {"status": "error", "reason": "no input file"}
+
+    db = read_json(input_file)
+    assets = db.get("assets", {})
+    if not assets:
+        print(f"{TAG} 错误: 数据库中无资产")
+        return {"status": "error", "reason": "empty assets"}
+
+    # 过滤: 跳过已推理的
+    pending = {}
+    for name, data in assets.items():
+        if data.get("ai_inference") is not None:
+            continue
+        pending[name] = data
+
+    print(f"{TAG} 数据库共 {len(assets)} 个资产，待推理 {len(pending)} 个")
+
+    if not pending:
+        print(f"{TAG} 全部已推理，无需处理")
+        return {"status": "ok", "total": len(assets), "inferred": 0, "skipped": len(assets)}
+
+    # 分批
+    batches = _prepare_batches(pending, batch_size)
+    print(f"{TAG} 分为 {len(batches)} 批")
+
+    if dry_run:
+        for cat, idx, keys in batches:
+            print(f"{TAG}   批次 [{cat}] #{idx}: {len(keys)} 个资产")
+            for k in keys[:3]:
+                print(f"{TAG}     - {k}")
+            if len(keys) > 3:
+                print(f"{TAG}     ... 还有 {len(keys) - 3} 个")
+        return {"status": "ok", "dry_run": True, "batches": len(batches)}
+
+    # 获取 API 配置
+    config = _get_api_config()
+    if not config:
+        return {"status": "error", "reason": "no api key"}
+
+    print(f"{TAG} API: {config['api_base']} 模型: {config['model']}")
+
+    # 逐批推理
+    total_inferred = 0
+    total_failed = 0
+
+    for batch_idx, (cat, batch_num, keys) in enumerate(batches):
+        print(f"{TAG} 批次 [{batch_idx + 1}/{len(batches)}] [{cat}] #{batch_num}: {len(keys)} 个资产...")
+
+        # 构造 prompt 数据
+        prompt_assets = []
+        for k in keys:
+            prompt_assets.append(_format_asset_for_prompt(k, pending[k]))
+
+        assets_json = json.dumps(prompt_assets, ensure_ascii=False, indent=2)
+
+        # 调用 API
+        results = _call_glm_api(config, cat, assets_json)
+
+        if not results:
+            print(f"{TAG}   ⚠️ 推理失败，跳过此批")
+            total_failed += len(keys)
+            continue
+
+        # 合并结果
+        matched = 0
+        for i, k in enumerate(keys):
+            if i < len(results):
+                assets[k]["ai_inference"] = results[i]
+                matched += 1
+                total_inferred += 1
+            else:
+                print(f"{TAG}   ⚠️ {k}: API 返回结果不足")
+
+        print(f"{TAG}   ✅ 匹配 {matched}/{len(keys)}")
+
+        # 每批都保存一次（断点续跑）
+        db["assets"] = assets
+        # 更新 metadata
+        if "ai_inference_metadata" not in db["metadata"]:
+            db["metadata"]["ai_inference_metadata"] = {
+                "model": config["model"],
+                "api_base": config["api_base"],
+                "started_at": iso_now(),
+            }
+        db["metadata"]["ai_inference_metadata"]["last_batch"] = batch_idx + 1
+        db["metadata"]["ai_inference_metadata"]["total_batches"] = len(batches)
+        db["metadata"]["ai_inference_metadata"]["inferred"] = total_inferred
+        db["metadata"]["ai_inference_metadata"]["failed"] = total_failed
+        write_json(db, output_file)
+
+        # 限速: 避免触发 API 限流
+        if batch_idx < len(batches) - 1:
+            time.sleep(1)
+
+    db["metadata"]["ai_inference_metadata"]["completed_at"] = iso_now()
+    write_json(db, output_file)
+
+    print(f"{TAG} 推理完成: {total_inferred} 成功, {total_failed} 失败")
+    return {
+        "status": "ok",
+        "total": len(assets),
+        "inferred": total_inferred,
+        "failed": total_failed,
+    }
+
+
+def main():
+    parser = argparse.ArgumentParser(description="GLM API 批量推理 — 资产语义增强")
+    parser.add_argument(
+        "--input", type=str,
+        default=str(_SCRIPT_DIR.parent.parent / "output" / "asset-catalog" / "asset_database.json"),
+        help="输入 asset_database.json 路径",
+    )
+    parser.add_argument(
+        "--output", type=str, default=None,
+        help="输出路径 (默认与 input 相同，就地更新)",
+    )
+    parser.add_argument(
+        "--batch-size", type=int, default=DEFAULT_BATCH_SIZE,
+        help=f"每批资产数量 (默认 {DEFAULT_BATCH_SIZE})",
+    )
+    parser.add_argument(
+        "--dry-run", action="store_true",
+        help="只打印批次信息，不调用 API",
+    )
+    args = parser.parse_args()
+
+    input_file = Path(args.input)
+    output_file = Path(args.output) if args.output else input_file
+
+    result = run_inference(input_file, output_file, args.batch_size, args.dry_run)
+    if result.get("status") != "ok" and not result.get("dry_run"):
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()

tools/asset-catalog/analyze_meshes.py

@@ -6,8 +6,6 @@
 - 网格复杂度（顶点数、三角形数）
 - Socket 列表
 - 材质槽位
-- 包围盒 6 面候选
-- 射线检测细节包围盒（spawn + Line Trace）
 - AI 占位字段（key_surfaces, style_tags, scene_tags 等）
 
 输出: output/meshes/<SubfolderName>.json，每个子目录一个文件
@@ -18,7 +16,6 @@
 """
 
 import sys
-import math
 from pathlib import Path
 
 _SCRIPT_DIR = Path(__file__).resolve().parent
@@ -28,7 +25,7 @@ if str(_SCRIPT_DIR) not in sys.path:
 from _common import (
     HAS_UNREAL, get_scan_root, get_output_dir, get_steps,
     write_json, read_json, wait_for_editor, progress_log,
-    iso_now, safe_subfolder_name, classify_pivot, compute_bbox_faces,
+    iso_now, safe_subfolder_name, classify_pivot,
 )
 
 TAG = "[MeshAnalysis]"
@@ -263,12 +260,9 @@ def _analyze_single_mesh(asset_info: dict, subfolder: str = "") -> dict:
         "mesh_complexity": {},
         "sockets": [],
         "material_slots": [],
-        "detail_boxes": [],
-        "bbox_faces": [],
         "analysis_notes": [],
         # AI Pass 2 占位
         "key_surfaces": [],
-        "detail_boxes_annotated": [],
         "ground_sink_estimate": None,
         "style_tags": [],
         "scene_tags": [],
@@ -301,39 +295,20 @@ def _analyze_single_mesh(asset_info: dict, subfolder: str = "") -> dict:
         # ---- 2. 锚点 ----
         result["pivot"] = classify_pivot(bmin, bmax)
 
-        # ---- 3. 包围盒面 ----
-        result["bbox_faces"] = compute_bbox_faces(bmin, bmax)
-
     except Exception as e:
         result["analysis_notes"].append(f"包围盒提取失败: {e}")
         return result
 
-    # ---- 4. 网格复杂度 ----
+    # ---- 3. 网格复杂度 ----
     result["mesh_complexity"] = _try_get_complexity(mesh)
 
-    # ---- 5. Sockets ----
+    # ---- 4. Sockets ----
     result["sockets"] = _try_get_sockets(mesh)
 
-    # ---- 6. 材质槽位 ----
+    # ---- 5. 材质槽位 ----
     result["material_slots"] = _try_get_material_slots(mesh)
 
-    # ---- 7. 细节包围盒 ----
-    detail_boxes, trace_debug = _try_trace_detail_boxes(mesh, bmin, bmax)
-    if trace_debug:
-        result["analysis_notes"].extend(trace_debug)
-    if detail_boxes:
-        result["detail_boxes"] = detail_boxes
-    else:
-        # 降级: 仅整体包围盒
-        result["detail_boxes"] = [{
-            "label": "overall",
-            "min": _round3(list(bmin)),
-            "max": _round3(list(bmax)),
-            "size": _round3(list(size)),
-            "source": "overall",
-        }]
-
-    # ---- 8. 自动 Tags ----
+    # ---- 6. 自动 Tags ----
     style_tags, scene_tags = _generate_mesh_tags(name, subfolder)
     result["style_tags"] = style_tags
     result["scene_tags"] = scene_tags
@@ -341,198 +316,6 @@ def _analyze_single_mesh(asset_info: dict, subfolder: str = "") -> dict:
     return result
 
 
-# ============================================================
-# Line Trace 细节包围盒
-# ============================================================
-
-# UE5.7 Python correct enum names (discovered via probe)
-_TRACE_CHANNEL = None
-_DRAW_DEBUG = None
-
-def _get_trace_enums():
-    """Lazy-init: get the correct UE5.7 trace enums."""
-    global _TRACE_CHANNEL, _DRAW_DEBUG
-    if _TRACE_CHANNEL is not None:
-        return _TRACE_CHANNEL is not None
-    import unreal
-    try:
-        _TRACE_CHANNEL = unreal.TraceTypeQuery.TRACE_TYPE_QUERY1
-        _DRAW_DEBUG = unreal.DrawDebugTrace.NONE
-        return True
-    except Exception:
-        return False
-
-
-def _try_trace_detail_boxes(mesh, bmin: tuple, bmax: tuple) -> tuple:
-    """通过 Line Trace 获取细节包围盒。
-
-    Returns:
-        (detail_boxes, debug_notes)
-    """
-    import unreal
-
-    detail_boxes = []
-    debug = []
-
-    size = (bmax[0] - bmin[0], bmax[1] - bmin[1], bmax[2] - bmin[2])
-    if max(size) < 10:
-        return [], ["skip_trace: model too small"]
-
-    # Check trace API availability
-    if not _get_trace_enums():
-        return [], ["trace_unavailable: enums not found"]
-
-    # ---- Spawn temporary Actor ----
-    try:
-        editor_subsystem = unreal.get_editor_subsystem(unreal.EditorActorSubsystem)
-        if not editor_subsystem:
-            return [], ["trace_fail: no subsystem"]
-    except Exception as e:
-        return [], [f"trace_fail: {e}"]
-
-    try:
-        actor = editor_subsystem.spawn_actor_from_object(mesh, unreal.Vector(0, 0, 0))
-    except Exception as e:
-        return [], [f"trace_fail: spawn {e}"]
-
-    if not actor:
-        return [], ["trace_fail: spawn None"]
-
-    # Set collision to BlockAll so traces can hit this mesh
-    smc = actor.get_component_by_class(unreal.StaticMeshComponent)
-    if smc:
-        try:
-            smc.set_collision_enabled(unreal.CollisionEnabled.QUERY_AND_PHYSICS)
-            smc.set_collision_profile_name("BlockAll")
-        except Exception as e:
-            debug.append(f"collision_set_warn: {e}")
-
-    try:
-        hit_points = _perform_line_traces(actor, bmin, bmax)
-        debug.append(f"trace_hits: {len(hit_points)}")
-
-        if len(hit_points) >= 4:
-            detail_boxes = _cluster_hit_points(hit_points, bmin, bmax)
-            if detail_boxes:
-                debug.append(f"trace_boxes: {len(detail_boxes)}")
-    finally:
-        try:
-            unreal.EditorLevelLibrary.destroy_actor(actor)
-        except Exception:
-            pass
-
-    return detail_boxes, debug
-
-
-def _perform_line_traces(actor, bmin: tuple, bmax: tuple) -> list:
-    """多方向 Line Trace 收集命中点。"""
-    import unreal
-
-    hit_points = []
-    mn_x, mn_y, mn_z = bmin
-    mx_x, mx_y, mx_z = bmax
-    sx, sy, sz = mx_x - mn_x, mx_y - mn_y, mx_z - mn_z
-
-    step = min(50, max(10, min(sx, sy) / 5))
-    margin = max(20, max(sx, sy, sz) * 0.1)
-    z_step = min(50, max(15, sz / 4))
-
-    trace_ch = _TRACE_CHANNEL
-    draw_dbg = _DRAW_DEBUG
-
-    def _trace(start_v, end_v):
-        try:
-            hit = unreal.SystemLibrary.line_trace_single(
-                actor, start_v, end_v,
-                trace_ch, False, [], draw_dbg,
-            )
-            if hit and hasattr(hit, 'impact_point'):
-                pt = hit.impact_point
-                hit_points.append((pt.x, pt.y, pt.z))
-        except Exception:
-            pass
-
-    try:
-        # Top-down (Z axis)
-        z = mn_z
-        while z <= mx_z + 0.1:
-            x = mn_x + step / 2
-            while x <= mx_x + 0.1:
-                y = mn_y + step / 2
-                while y <= mx_y + 0.1:
-                    _trace(unreal.Vector(x, y, mx_z + margin),
-                           unreal.Vector(x, y, mn_z - margin))
-                    y += step
-                x += step
-            z += z_step
-
-        # Front (Y+)
-        x = mn_x + step / 2
-        while x <= mx_x + 0.1:
-            z = mn_z + step / 2
-            while z <= mx_z + 0.1:
-                _trace(unreal.Vector(x, mx_y + margin, z),
-                       unreal.Vector(x, mn_y - margin, z))
-                z += step
-            x += step
-
-        # Right (X+)
-        y = mn_y + step / 2
-        while y <= mx_y + 0.1:
-            z = mn_z + step / 2
-            while z <= mx_z + 0.1:
-                _trace(unreal.Vector(mx_x + margin, y, z),
-                       unreal.Vector(mn_x - margin, y, z))
-                z += step
-            y += step
-    except Exception:
-        pass
-
-    return hit_points
-
-
-def _cluster_hit_points(hit_points: list, bmin: tuple, bmax: tuple) -> list:
-    """将命中点按 Z 轴聚类为细节包围盒。"""
-    if not hit_points:
-        return []
-
-    sorted_pts = sorted(hit_points, key=lambda p: p[2])
-    z_slices = []
-    current_slice = [sorted_pts[0]]
-    z_gap = 10.0
-
-    for i in range(1, len(sorted_pts)):
-        if sorted_pts[i][2] - current_slice[-1][2] > z_gap:
-            z_slices.append(current_slice)
-            current_slice = [sorted_pts[i]]
-        else:
-            current_slice.append(sorted_pts[i])
-    z_slices.append(current_slice)
-
-    boxes = []
-    for idx, pts in enumerate(z_slices):
-        if len(pts) < 3:
-            continue
-        xs = [p[0] for p in pts]
-        ys = [p[1] for p in pts]
-        zs = [p[2] for p in pts]
-        bmn = [min(xs), min(ys), min(zs)]
-        bmx = [max(xs), max(ys), max(zs)]
-        bsz = [bmx[0]-bmn[0], bmx[1]-bmn[1], bmx[2]-bmn[2]]
-        if bsz[0] < 2 and bsz[1] < 2:
-            continue
-        boxes.append({
-            "label": f"part_{idx}",
-            "min": _round3(bmn), "max": _round3(bmx), "size": _round3(bsz),
-            "source": "trace", "hit_count": len(pts),
-        })
-
-    if len(boxes) > 8:
-        boxes.sort(key=lambda b: b["size"][0]*b["size"][1]*b["size"][2], reverse=True)
-        boxes = boxes[:8]
-    return boxes
-
-
 # ============================================================
 # 主流程
 # ============================================================

tools/asset-catalog/build_database.py

@@ -0,0 +1,147 @@
+"""资产数据库汇总脚本。
+
+读取 asset-catalog 的扫描结果（meshes/*.json），汇总为统一的单文件
+asset_database.json，以资产名为 key 扁平化存储。
+
+此脚本不依赖 UE5，可离线运行。
+
+用法:
+  python build_database.py [--input DIR] [--output FILE]
+
+  --input   扫描输出目录，默认 output/asset-catalog
+  --output  输出文件路径，默认 <input>/asset_database.json
+"""
+
+import sys
+import argparse
+from pathlib import Path
+
+_SCRIPT_DIR = Path(__file__).resolve().parent
+if str(_SCRIPT_DIR) not in sys.path:
+    sys.path.insert(0, str(_SCRIPT_DIR))
+
+from _common import read_json, write_json, iso_now
+
+TAG = "[BuildDB]"
+
+
+def build_database(input_dir: Path, output_file: Path) -> dict:
+    """汇总 meshes/*.json 为统一数据库。
+
+    Args:
+        input_dir: 包含 meshes/ 子目录的扫描输出根目录
+        output_file: 输出的 asset_database.json 路径
+
+    Returns:
+        统计信息 dict
+    """
+    meshes_dir = input_dir / "meshes"
+    if not meshes_dir.exists():
+        print(f"{TAG} 错误: {meshes_dir} 不存在，请先运行 analyze_meshes")
+        return {"status": "error", "reason": "no meshes dir"}
+
+    # 收集所有子目录的扫描文件
+    mesh_files = sorted(meshes_dir.glob("*.json"))
+    if not mesh_files:
+        print(f"{TAG} 错误: {meshes_dir} 下无 JSON 文件")
+        return {"status": "error", "reason": "no mesh files"}
+
+    print(f"{TAG} 发现 {len(mesh_files)} 个扫描文件")
+
+    # 扁平化合并
+    assets = {}
+    total = 0
+    skipped = 0
+
+    for mf in mesh_files:
+        try:
+            data = read_json(mf)
+        except Exception as e:
+            print(f"{TAG} 警告: 读取 {mf.name} 失败: {e}")
+            continue
+
+        subfolder = data.get("subfolder", mf.stem)
+        meshes = data.get("meshes", {})
+        if not meshes:
+            continue
+
+        for name, mesh_data in meshes.items():
+            total += 1
+
+            # 跳过无包围盒的（加载失败的）
+            if mesh_data.get("bounding_box") is None:
+                skipped += 1
+                continue
+
+            # 构建统一条目
+            entry = {
+                "ue_path": mesh_data.get("ue_path", ""),
+                "subfolder": subfolder,
+                "bounding_box": mesh_data.get("bounding_box"),
+                "pivot": mesh_data.get("pivot"),
+                "mesh_complexity": mesh_data.get("mesh_complexity", {}),
+                "material_slots": mesh_data.get("material_slots", []),
+                "sockets": mesh_data.get("sockets", []),
+                "tags": {
+                    "style": mesh_data.get("style_tags", []),
+                    "scene": mesh_data.get("scene_tags", []),
+                },
+                "ai_inference": None,
+            }
+
+            # 如果资产名重复（不同子目录下同名），加后缀
+            key = name
+            if key in assets:
+                safe_sub = subfolder.replace("/", "_").replace("\\", "_")
+                key = f"{name}_{safe_sub}"
+                print(f"{TAG} 注意: 资产名重复 {name}，使用 {key}")
+
+            assets[key] = entry
+
+    # 写入数据库
+    db = {
+        "metadata": {
+            "generated_at": iso_now(),
+            "source": "asset_catalog_v1",
+            "total_assets": len(assets),
+            "skipped_no_bbox": skipped,
+            "input_files": [mf.name for mf in mesh_files],
+        },
+        "assets": assets,
+    }
+
+    write_json(db, output_file)
+    print(f"{TAG} ✅ 汇总完成: {len(assets)} 个资产 → {output_file}")
+    print(f"{TAG}    跳过 {skipped} 个无包围盒资产 / 总计 {total}")
+
+    return {
+        "status": "ok",
+        "total": total,
+        "assets": len(assets),
+        "skipped": skipped,
+    }
+
+
+def main():
+    parser = argparse.ArgumentParser(description="汇总资产扫描数据为统一 JSON 数据库")
+    parser.add_argument(
+        "--input", type=str,
+        default=str(_SCRIPT_DIR.parent.parent / "output" / "asset-catalog"),
+        help="扫描输出目录 (默认: output/asset-catalog)",
+    )
+    parser.add_argument(
+        "--output", type=str, default=None,
+        help="输出文件路径 (默认: <input>/asset_database.json)",
+    )
+    args = parser.parse_args()
+
+    input_dir = Path(args.input)
+    output_file = Path(args.output) if args.output else input_dir / "asset_database.json"
+
+    result = build_database(input_dir, output_file)
+    if result.get("status") != "ok":
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()