diff --git a/furniture-3d-system/src/server/services/bom-costing/bom-costing-multimodal-analyzer.test.ts b/furniture-3d-system/src/server/services/bom-costing/bom-costing-multimodal-analyzer.test.ts
index 4cb2e951..f1cf9f6f 100644
--- a/furniture-3d-system/src/server/services/bom-costing/bom-costing-multimodal-analyzer.test.ts
+++ b/furniture-3d-system/src/server/services/bom-costing/bom-costing-multimodal-analyzer.test.ts
@@ -1004,6 +1004,54 @@ describe('BOM 多模态分析器', () => {
     expect(JSON.stringify(compactInput)).not.toContain('不属于当前批次')
   })
 
+  it('高密度详情事实只保留最有成本价值的一组产品级证据', () => {
+    const factSnapshot = createCompleteFactSnapshot(snapshot)
+    const factKinds = [
+      'DIMENSION',
+      'MATERIAL',
+      'STRUCTURE',
+      'MECHANISM',
+      'HARDWARE',
+      'PART',
+      'PROCESS',
+      'PACKAGING',
+      'WEIGHT',
+      'SKU_DIFFERENCE',
+      'TEXT',
+    ] as const
+    factSnapshot.imageFacts.push(...Array.from({ length: 18 }, (_, index) => ({
+      imageId: `DETAIL_DENSE_${index + 1}`,
+      sourceImageUrl: `https://img.example.com/detail-dense-${index + 1}.jpg`,
+      kind: 'DETAIL' as const,
+      label: `高密度详情图 ${index + 1}`,
+      externalSkuIds: [],
+      observedTexts: [],
+      facts: [{
+        factId: `fact_dense_${index + 1}`,
+        kind: factKinds[index % factKinds.length]!,
+        subject: `部件 ${index + 1}`,
+        value: `事实 ${index + 1}`,
+        relatedSkuIds: [],
+        basis: 'IMAGE' as const,
+        confidence: 0.8,
+      }],
+    })))
+
+    const compactInput = buildSynthesisFactInput(
+      factSnapshot,
+      [snapshot.skus[0]!],
+    )
+    const detailImages = compactInput.images.filter(
+      (image) => image.kind === 'DETAIL',
+    )
+    const retainedKinds = new Set(
+      detailImages.flatMap((image) => image.facts.map((fact) => fact.k)),
+    )
+
+    expect(detailImages).toHaveLength(14)
+    expect([...factKinds].every((kind) => retainedKinds.has(kind))).toBe(true)
+  })
+
   it('复用图片事实时，GPT-5 最终汇总在网关截断前结束等待', async () => {
     const timeoutSpy = jest
       .spyOn(AbortSignal, 'timeout')
diff --git a/furniture-3d-system/src/server/services/bom-costing/bom-costing-multimodal-analyzer.ts b/furniture-3d-system/src/server/services/bom-costing/bom-costing-multimodal-analyzer.ts
index 7fe99433..b8419fb8 100644
--- a/furniture-3d-system/src/server/services/bom-costing/bom-costing-multimodal-analyzer.ts
+++ b/furniture-3d-system/src/server/services/bom-costing/bom-costing-multimodal-analyzer.ts
@@ -75,7 +75,7 @@ const SYSTEM_PROMPT = `你是家具产品结构与成本工程师。你的任务
 14. 每条 evidence 必须设置 basis：图片直接可见为 IMAGE；图片标注、详情文字或商品文案为 SOURCE_TEXT；SKU 名称、规格或属性为 SKU_METADATA；由装配或机构功能必需性推断的隐藏部件为 STRUCTURAL_INFERENCE。
 15. STRUCTURAL_INFERENCE 只允许用于实现已确认结构或功能所必需的隐藏五金、连接件或保护包材；必须写明推断原因、数量估算和待确认项，evidence.confidence 与 BOM 行 confidence 都不得高于 0.55。
 16. 安装知识只用于提醒检查，不是当前商品的事实。不得因为知识库列出了某种五金，就在缺少当前商品结构或功能依据时把它写进 BOM。
-17. 逐图事实已经覆盖尺寸标注、局部小图、剖面图、结构示意、折叠前后状态和产品信息图；合并时不得遗漏这些事实。
+17. 系统会从已归档的逐图事实中挑选与 BOM 最相关的产品级证据，并附上当前批次关联的 SKU 图片事实；不得遗漏本次输入中给出的事实。
 18. SKU 名称或属性中的“折叠、升降、推拉、抽屉、脚轮、电动”等功能词必须触发对应机构检查；同时区分收拢态、展开态与整机尺寸。
 19. 对板件之间、玻璃与框架、抽屉与柜体、门板与柜体等装配界面，必须检查连接方式；看不清时可以给出低置信度的必需连接件估算，但不得编造具体型号。
 20. 每个结构组必须至少包含材料类（BOARD、STEEL、ACCESSORY 中至少一条）、PROCESSING、LABOR 和 PACKAGING，确保系统能立即形成一套出厂成本预估；不确定项用低置信度估算并在 evidence 中说明，不得直接省略整个成本组。`
@@ -237,6 +237,20 @@ const DEFAULT_SYNTHESIS_BATCH_CONCURRENCY = 3
 const MAX_EXPLICIT_FACT_SKU_IDS = 20
 const MAX_SYNTHESIS_CONTRACT_ATTEMPTS = 2
 const SYNTHESIS_IMAGE_REFERENCE_PREFIX = 'bom-image:'
+const MAX_PRODUCT_LEVEL_FACT_IMAGES = 14
+const PRODUCT_LEVEL_FACT_KIND_WEIGHTS: Record<string, number> = {
+  DIMENSION: 8,
+  MATERIAL: 8,
+  STRUCTURE: 7,
+  MECHANISM: 7,
+  HARDWARE: 7,
+  PART: 6,
+  PROCESS: 5,
+  PACKAGING: 5,
+  WEIGHT: 5,
+  SKU_DIFFERENCE: 4,
+  TEXT: 1,
+}
 
 function describeHighCardinalitySkuCatalog(skus: MarketSkuSnapshot[]) {
   const valuesByAttribute = new Map<string, Set<string>>()
@@ -335,11 +349,87 @@ function createSynthesisImageReferenceMap(
   )
 }
 
+function getSynthesisFactSignature(
+  fact: BomCostingSourceFactSnapshot['imageFacts'][number]['facts'][number],
+) {
+  return JSON.stringify([
+    fact.kind,
+    fact.subject,
+    fact.value,
+    fact.numericValue ?? null,
+    fact.unit ?? null,
+    fact.basis,
+  ])
+}
+
+function selectProductLevelSynthesisImageIds(
+  factSnapshot: BomCostingSourceFactSnapshot,
+) {
+  const selectedIds = new Set(
+    factSnapshot.imageFacts
+      .filter((image) => image.kind === 'MAIN')
+      .map((image) => image.imageId),
+  )
+  const remainingDetails = factSnapshot.imageFacts
+    .filter((image) => image.kind === 'DETAIL')
+    .map((image, originalIndex) => ({ image, originalIndex }))
+  const seenFactSignatures = new Set<string>()
+  const seenFactKinds = new Set<string>()
+
+  for (
+    let selectedCount = 0;
+    selectedCount < MAX_PRODUCT_LEVEL_FACT_IMAGES && remainingDetails.length > 0;
+    selectedCount += 1
+  ) {
+    let bestIndex = 0
+    let bestScore = Number.NEGATIVE_INFINITY
+    for (const [index, candidate] of remainingDetails.entries()) {
+      const candidateSignatures = new Set<string>()
+      let score = 0
+      let newKindScore = 0
+      for (const fact of candidate.image.facts) {
+        const signature = getSynthesisFactSignature(fact)
+        if (candidateSignatures.has(signature)) continue
+        candidateSignatures.add(signature)
+        const weight = PRODUCT_LEVEL_FACT_KIND_WEIGHTS[fact.kind] ?? 1
+        if (!seenFactKinds.has(fact.kind)) newKindScore += weight
+        score += seenFactSignatures.has(signature) ? weight : weight * 3
+      }
+      // A detail image with no structured fact can still contain a small,
+      // useful text fragment. Keep it behind all structured evidence.
+      if (candidateSignatures.size === 0 && candidate.image.observedTexts.length > 0) {
+        score = 0.5
+      }
+      // First retain at least one image for every available fact category.
+      // This prevents a large number of dimensions or parts from crowding out
+      // a lower-volume but cost-relevant category such as packaging.
+      if (newKindScore > 0) score += 10_000 + newKindScore
+      if (
+        score > bestScore ||
+        (score === bestScore && candidate.originalIndex < remainingDetails[bestIndex]!.originalIndex)
+      ) {
+        bestIndex = index
+        bestScore = score
+      }
+    }
+    const [selected] = remainingDetails.splice(bestIndex, 1)
+    if (!selected) break
+    selectedIds.add(selected.image.imageId)
+    selected.image.facts.forEach((fact) => {
+      seenFactSignatures.add(getSynthesisFactSignature(fact))
+      seenFactKinds.add(fact.kind)
+    })
+  }
+
+  return selectedIds
+}
+
 export function buildSynthesisFactInput(
   factSnapshot: BomCostingSourceFactSnapshot,
   skus: MarketSkuSnapshot[],
 ) {
   const batchSkuIds = new Set(skus.map((sku) => sku.externalSkuId))
+  const productLevelImageIds = selectProductLevelSynthesisImageIds(factSnapshot)
   return {
     schemaVersion: factSnapshot.schemaVersion,
     fieldReference: {
@@ -348,8 +438,10 @@ export function buildSynthesisFactInput(
     },
     images: factSnapshot.imageFacts
       .filter((image) => {
-        if (image.kind !== 'SKU') return true
-        return image.externalSkuIds.some((skuId) => batchSkuIds.has(skuId))
+        if (image.kind === 'SKU') {
+          return image.externalSkuIds.some((skuId) => batchSkuIds.has(skuId))
+        }
+        return productLevelImageIds.has(image.imageId)
       })
       .map((image) => {
         const batchImageSkuIds = image.externalSkuIds.filter((skuId) =>