11

server-node/src/services/customWorldSceneNpcGenerationService.ts

@@ -0,0 +1,586 @@
+import { parseJsonResponseText } from '../../../packages/shared/src/llm/parsers.js';
+import { badRequest } from '../errors.js';
+import type { UpstreamLlmClient } from './llmClient.js';
+
+type SceneNpcGenerationInput = {
+  profile: Record<string, unknown>;
+  landmarkId: string;
+};
+
+type ParsedStoryNpc = {
+  id: string;
+  name: string;
+  title: string;
+  role: string;
+  description: string;
+  personality: string;
+  motivation: string;
+  relationshipHooks: string[];
+  tags: string[];
+};
+
+type ParsedLandmark = {
+  id: string;
+  name: string;
+  description: string;
+  dangerLevel: string;
+  sceneNpcIds: string[];
+};
+
+type ParsedProfile = {
+  name: string;
+  settingText: string;
+  storyNpcs: ParsedStoryNpc[];
+  landmarks: ParsedLandmark[];
+};
+
+type GeneratedNpcDraft = {
+  name: string;
+  title: string;
+  role: string;
+  description: string;
+  backstory: string;
+  personality: string;
+  motivation: string;
+  combatStyle: string;
+  initialAffinity: number;
+  relationshipHooks: string[];
+  tags: string[];
+  publicSummary: string;
+  chapterTeasers: string[];
+  chapterContents: string[];
+  skills: Array<{
+    name: string;
+    summary: string;
+    style: string;
+  }>;
+  initialItems: Array<{
+    name: string;
+    category: string;
+    quantity: number;
+    rarity: 'common' | 'uncommon' | 'rare' | 'epic' | 'legendary';
+    description: string;
+    tags: string[];
+  }>;
+};
+
+function toRecord(value: unknown) {
+  return value && typeof value === 'object' && !Array.isArray(value)
+    ? (value as Record<string, unknown>)
+    : null;
+}
+
+function toText(value: unknown, fallback = '') {
+  return typeof value === 'string' && value.trim() ? value.trim() : fallback;
+}
+
+function toStringArray(value: unknown, maxCount = 12) {
+  if (!Array.isArray(value)) {
+    return [];
+  }
+
+  return value
+    .map((item) => toText(item))
+    .filter(Boolean)
+    .slice(0, maxCount);
+}
+
+function clampText(value: string, maxLength: number) {
+  const normalized = value.replace(/\s+/gu, ' ').trim();
+  if (!normalized) {
+    return '';
+  }
+  if (normalized.length <= maxLength) {
+    return normalized;
+  }
+  return `${normalized.slice(0, Math.max(0, maxLength - 1)).trim()}…`;
+}
+
+function slugify(value: string) {
+  const normalized = value
+    .trim()
+    .toLowerCase()
+    .replace(/[^a-z0-9\u4e00-\u9fa5]+/gu, '-')
+    .replace(/^-+|-+$/gu, '');
+
+  return normalized || 'entry';
+}
+
+function createStableId(prefix: string, label: string, seed: string) {
+  return `${prefix}-${slugify(label || prefix)}-${seed}`;
+}
+
+function dedupeStrings(values: string[], maxCount = 8) {
+  return [...new Set(values.map((value) => value.trim()).filter(Boolean))].slice(
+    0,
+    maxCount,
+  );
+}
+
+function normalizeStoryNpc(value: unknown): ParsedStoryNpc | null {
+  const record = toRecord(value);
+  if (!record) {
+    return null;
+  }
+
+  const id = toText(record.id);
+  const name = toText(record.name);
+  if (!id || !name) {
+    return null;
+  }
+
+  const title = toText(record.title);
+  const role = toText(record.role, title || '场景角色');
+
+  return {
+    id,
+    name,
+    title: title || role || '场景角色',
+    role,
+    description: toText(record.description),
+    personality: toText(record.personality),
+    motivation: toText(record.motivation),
+    relationshipHooks: toStringArray(record.relationshipHooks, 6),
+    tags: toStringArray(record.tags, 8),
+  };
+}
+
+function normalizeLandmark(value: unknown): ParsedLandmark | null {
+  const record = toRecord(value);
+  if (!record) {
+    return null;
+  }
+
+  const id = toText(record.id);
+  const name = toText(record.name);
+  if (!id || !name) {
+    return null;
+  }
+
+  return {
+    id,
+    name,
+    description: toText(record.description),
+    dangerLevel: toText(record.dangerLevel, '中'),
+    sceneNpcIds: toStringArray(record.sceneNpcIds, 12),
+  };
+}
+
+function normalizeProfile(value: unknown): ParsedProfile {
+  const record = toRecord(value);
+  if (!record) {
+    throw badRequest('profile is required');
+  }
+
+  const storyNpcs = Array.isArray(record.storyNpcs)
+    ? record.storyNpcs.map(normalizeStoryNpc).filter((item): item is ParsedStoryNpc => item !== null)
+    : [];
+  const landmarks = Array.isArray(record.landmarks)
+    ? record.landmarks.map(normalizeLandmark).filter((item): item is ParsedLandmark => item !== null)
+    : [];
+
+  return {
+    name: toText(record.name, '自定义世界'),
+    settingText: toText(record.settingText),
+    storyNpcs,
+    landmarks,
+  };
+}
+
+function ensureUniqueName(name: string, existingNames: string[]) {
+  const normalizedName = name.trim() || '新场景角色';
+  if (!existingNames.includes(normalizedName)) {
+    return normalizedName;
+  }
+
+  let index = 2;
+  let nextName = `${normalizedName}${index}`;
+  while (existingNames.includes(nextName)) {
+    index += 1;
+    nextName = `${normalizedName}${index}`;
+  }
+  return nextName;
+}
+
+function buildFallbackDraft(
+  profile: ParsedProfile,
+  landmark: ParsedLandmark,
+  sceneNpcs: ParsedStoryNpc[],
+): GeneratedNpcDraft {
+  const tags = dedupeStrings([
+    landmark.name,
+    landmark.dangerLevel,
+    ...sceneNpcs.flatMap((npc) => npc.tags),
+  ], 4);
+
+  return {
+    name: `${landmark.name}来客`,
+    title: `${landmark.name}的观察者`,
+    role: `${landmark.name}的观察者`,
+    description: `长期活动于${landmark.name}，熟悉这里的局势与暗线，能为玩家提供新的观察角度。`,
+    backstory: `他在${landmark.name}扎根已久，对这片区域的危险节奏、人物流动与隐藏冲突有自己的判断。`,
+    personality: '谨慎、敏锐，先观察再表态。',
+    motivation: `希望借玩家之手改变${landmark.name}当前逐渐失衡的局面。`,
+    combatStyle: '偏向控场与试探，不轻易暴露底牌。',
+    initialAffinity: 6,
+    relationshipHooks: dedupeStrings([
+      `与${landmark.name}局势深度绑定`,
+      sceneNpcs[0] ? `对${sceneNpcs[0].name}保持长期观察` : '对玩家保持试探',
+      '愿意交换情报，但保留关键秘密',
+    ], 3),
+    tags,
+    publicSummary: `一名活跃于${landmark.name}的关键观察者。`,
+    chapterTeasers: [
+      '他知道这片区域最近正在发生什么。',
+      '他与此地某个旧事件有直接牵连。',
+      '他真正想推动的局面并不只是自保。',
+      '他手里握有改变关系网的最后筹码。',
+    ],
+    chapterContents: [
+      `他常年在${landmark.name}周边活动，对人和事的变化极为敏感。`,
+      `多年前的一次变故把他和${landmark.name}牢牢绑在了一起。`,
+      `他表面克制，实际上一直在寻找扭转局面的机会。`,
+      '他保留着一张只会在局势逼近临界点时才动用的底牌。',
+    ],
+    skills: [
+      {
+        name: '试探起手',
+        summary: '以低风险方式摸清对手意图。',
+        style: '试探压制',
+      },
+      {
+        name: '地形借势',
+        summary: `借助${landmark.name}环境制造主动权。`,
+        style: '环境协同',
+      },
+      {
+        name: '暗线反制',
+        summary: '在关键回合揭示隐藏准备，打乱对方节奏。',
+        style: '后手翻盘',
+      },
+    ],
+    initialItems: [
+      {
+        name: '随身兵装',
+        category: '武器',
+        quantity: 1,
+        rarity: 'rare',
+        description: '常备的近身防护装备。',
+        tags: ['自定义', landmark.name],
+      },
+      {
+        name: '区域通行物',
+        category: '道具',
+        quantity: 1,
+        rarity: 'uncommon',
+        description: `能在${landmark.name}一带快速周转的私人物件。`,
+        tags: ['自定义'],
+      },
+      {
+        name: '情报残页',
+        category: '专属物品',
+        quantity: 1,
+        rarity: 'rare',
+        description: '记录着部分隐藏线索与往事片段。',
+        tags: ['线索'],
+      },
+    ],
+  };
+}
+
+function buildPrompt(
+  profile: ParsedProfile,
+  landmark: ParsedLandmark,
+  sceneNpcs: ParsedStoryNpc[],
+  otherNpcs: ParsedStoryNpc[],
+) {
+  const sceneNpcSummary = sceneNpcs.length
+    ? sceneNpcs
+        .map(
+          (npc, index) =>
+            `${index + 1}. ${npc.name} / ${npc.title || npc.role} / ${npc.description || '无描述'} / 性格：${npc.personality || '未写'} / 动机：${npc.motivation || '未写'}`,
+        )
+        .join('\n')
+    : '当前场景还没有已加入 NPC。';
+
+  const reserveNpcSummary = otherNpcs.length
+    ? otherNpcs
+        .slice(0, 8)
+        .map(
+          (npc, index) =>
+            `${index + 1}. ${npc.name} / ${npc.title || npc.role} / ${npc.description || '无描述'}`,
+        )
+        .join('\n')
+    : '暂无其他场景角色参考。';
+
+  const landmarkSummary = profile.landmarks
+    .slice(0, 10)
+    .map(
+      (entry, index) =>
+        `${index + 1}. ${entry.name} / 危险度：${entry.dangerLevel || '中'} / ${entry.description || '无描述'}`,
+    )
+    .join('\n');
+
+  return [
+    `世界名：${profile.name}`,
+    `世界设定：${profile.settingText || '未提供额外设定文本。'}`,
+    `当前目标场景：${landmark.name}`,
+    `场景描述：${landmark.description || '未填写'}`,
+    `危险度：${landmark.dangerLevel || '中'}`,
+    `当前场景已加入 NPC：\n${sceneNpcSummary}`,
+    `其他可参考 NPC：\n${reserveNpcSummary}`,
+    `世界内其他场景概览：\n${landmarkSummary}`,
+    '请生成 1 名适合加入当前场景的新 NPC。',
+    '要求：',
+    '- 必须与当前场景气质、危险度、已有 NPC 分工互补，不要和已有 NPC 重复。',
+    '- 角色要像真正可落地到游戏里的场景角色，不要写成抽象设定。',
+    '- 关系钩子、技能、初始物品都要可直接进入编辑器。',
+    '- 返回 JSON，不要额外解释。',
+    'JSON 结构：',
+    '{',
+    '  "npc": {',
+    '    "name": "角色名",',
+    '    "title": "头衔",',
+    '    "role": "身份",',
+    '    "description": "一句到两句角色描述",',
+    '    "backstory": "背景",',
+    '    "personality": "性格",',
+    '    "motivation": "动机",',
+    '    "combatStyle": "战斗风格",',
+    '    "initialAffinity": 6,',
+    '    "relationshipHooks": ["关系钩子1", "关系钩子2", "关系钩子3"],',
+    '    "tags": ["标签1", "标签2", "标签3"],',
+    '    "publicSummary": "公开背景摘要",',
+    '    "chapterTeasers": ["表层来意", "旧事裂痕", "隐藏执念", "最终底牌"],',
+    '    "chapterContents": ["对应正文1", "对应正文2", "对应正文3", "对应正文4"],',
+    '    "skills": [',
+    '      { "name": "技能1", "summary": "说明", "style": "风格" },',
+    '      { "name": "技能2", "summary": "说明", "style": "风格" },',
+    '      { "name": "技能3", "summary": "说明", "style": "风格" }',
+    '    ],',
+    '    "initialItems": [',
+    '      { "name": "物品1", "category": "分类", "quantity": 1, "rarity": "rare", "description": "说明", "tags": ["标签"] },',
+    '      { "name": "物品2", "category": "分类", "quantity": 1, "rarity": "uncommon", "description": "说明", "tags": ["标签"] },',
+    '      { "name": "物品3", "category": "分类", "quantity": 1, "rarity": "rare", "description": "说明", "tags": ["标签"] }',
+    '    ]',
+    '  }',
+    '}',
+  ].join('\n');
+}
+
+function sanitizeGeneratedNpc(
+  rawValue: unknown,
+  profile: ParsedProfile,
+  landmark: ParsedLandmark,
+  fallbackDraft: GeneratedNpcDraft,
+) {
+  const record = toRecord(rawValue);
+  const existingNames = profile.storyNpcs.map((npc) => npc.name);
+  const seed = Date.now().toString(36);
+  const chapterTitles = ['表层来意', '旧事裂痕', '隐藏执念', '最终底牌'];
+  const chapterThresholds = [6, 12, 18, 24];
+  const relationshipHooks = dedupeStrings(
+    toStringArray(record?.relationshipHooks, 6).concat(
+      fallbackDraft.relationshipHooks,
+    ),
+    4,
+  );
+  const tags = dedupeStrings(
+    toStringArray(record?.tags, 8).concat(fallbackDraft.tags, landmark.name),
+    6,
+  );
+  const chapterTeasers = toStringArray(record?.chapterTeasers, 4);
+  const chapterContents = toStringArray(record?.chapterContents, 4);
+  const skillRecords = Array.isArray(record?.skills) ? record?.skills : [];
+  const itemRecords = Array.isArray(record?.initialItems) ? record?.initialItems : [];
+
+  const draft: GeneratedNpcDraft = {
+    name: ensureUniqueName(
+      toText(record?.name, fallbackDraft.name),
+      existingNames,
+    ),
+    title: toText(record?.title, fallbackDraft.title),
+    role: toText(record?.role, toText(record?.title, fallbackDraft.role)),
+    description: clampText(
+      toText(record?.description, fallbackDraft.description),
+      120,
+    ),
+    backstory: clampText(toText(record?.backstory, fallbackDraft.backstory), 260),
+    personality: clampText(
+      toText(record?.personality, fallbackDraft.personality),
+      100,
+    ),
+    motivation: clampText(
+      toText(record?.motivation, fallbackDraft.motivation),
+      120,
+    ),
+    combatStyle: clampText(
+      toText(record?.combatStyle, fallbackDraft.combatStyle),
+      100,
+    ),
+    initialAffinity:
+      typeof record?.initialAffinity === 'number' &&
+      Number.isFinite(record.initialAffinity)
+        ? Math.min(12, Math.max(1, Math.round(record.initialAffinity)))
+        : fallbackDraft.initialAffinity,
+    relationshipHooks,
+    tags,
+    publicSummary: clampText(
+      toText(record?.publicSummary, fallbackDraft.publicSummary),
+      120,
+    ),
+    chapterTeasers:
+      chapterTeasers.length === 4
+        ? chapterTeasers
+        : fallbackDraft.chapterTeasers.slice(0, 4),
+    chapterContents:
+      chapterContents.length === 4
+        ? chapterContents
+        : fallbackDraft.chapterContents.slice(0, 4),
+    skills:
+      skillRecords.length >= 3
+        ? skillRecords.slice(0, 3).map((skill, index) => {
+            const skillRecord = toRecord(skill);
+            const fallbackSkill =
+              fallbackDraft.skills[index] ?? fallbackDraft.skills[0];
+            return {
+              name: clampText(
+                toText(skillRecord?.name, fallbackSkill?.name || `技能${index + 1}`),
+                20,
+              ),
+              summary: clampText(
+                toText(skillRecord?.summary, fallbackSkill?.summary || ''),
+                60,
+              ),
+              style: clampText(
+                toText(skillRecord?.style, fallbackSkill?.style || ''),
+                20,
+              ),
+            };
+          })
+        : fallbackDraft.skills,
+    initialItems:
+      itemRecords.length >= 3
+        ? itemRecords.slice(0, 3).map((item, index) => {
+            const itemRecord = toRecord(item);
+            const fallbackItem =
+              fallbackDraft.initialItems[index] ?? fallbackDraft.initialItems[0];
+            const rarity = toText(itemRecord?.rarity, fallbackItem?.rarity || 'rare');
+            return {
+              name: clampText(
+                toText(itemRecord?.name, fallbackItem?.name || `物品${index + 1}`),
+                20,
+              ),
+              category: clampText(
+                toText(itemRecord?.category, fallbackItem?.category || '道具'),
+                16,
+              ),
+              quantity:
+                typeof itemRecord?.quantity === 'number' &&
+                Number.isFinite(itemRecord.quantity)
+                  ? Math.min(9, Math.max(1, Math.round(itemRecord.quantity)))
+                  : fallbackItem?.quantity || 1,
+              rarity:
+                rarity === 'common' ||
+                rarity === 'uncommon' ||
+                rarity === 'rare' ||
+                rarity === 'epic' ||
+                rarity === 'legendary'
+                  ? rarity
+                  : fallbackItem?.rarity || 'rare',
+              description: clampText(
+                toText(itemRecord?.description, fallbackItem?.description || ''),
+                80,
+              ),
+              tags: dedupeStrings(
+                toStringArray(itemRecord?.tags, 4).concat(
+                  fallbackItem?.tags ?? [],
+                ),
+                4,
+              ),
+            };
+          })
+        : fallbackDraft.initialItems,
+  };
+
+  return {
+    id: createStableId('story-npc', draft.name, seed),
+    name: draft.name,
+    title: draft.title || draft.role,
+    role: draft.role || draft.title,
+    description: draft.description,
+    backstory: draft.backstory,
+    personality: draft.personality,
+    motivation: draft.motivation,
+    combatStyle: draft.combatStyle,
+    initialAffinity: draft.initialAffinity,
+    relationshipHooks: draft.relationshipHooks,
+    relations: [],
+    tags: draft.tags,
+    backstoryReveal: {
+      publicSummary: draft.publicSummary,
+      chapters: chapterTitles.map((title, index) => ({
+        id: ['surface', 'scar', 'hidden', 'final'][index],
+        title,
+        affinityRequired: chapterThresholds[index],
+        teaser:
+          draft.chapterTeasers[index] ?? fallbackDraft.chapterTeasers[index] ?? '',
+        content:
+          draft.chapterContents[index] ??
+          fallbackDraft.chapterContents[index] ??
+          '',
+        contextSnippet: '',
+      })),
+    },
+    skills: draft.skills.map((skill, index) => ({
+      id: createStableId('skill', `${draft.name}-${skill.name}`, `${seed}-${index + 1}`),
+      name: skill.name,
+      summary: skill.summary,
+      style: skill.style,
+    })),
+    initialItems: draft.initialItems.map((item, index) => ({
+      id: createStableId('item', `${draft.name}-${item.name}`, `${seed}-${index + 1}`),
+      name: item.name,
+      category: item.category,
+      quantity: item.quantity,
+      rarity: item.rarity,
+      description: item.description,
+      tags: item.tags,
+    })),
+  };
+}
+
+export async function generateSceneNpcForLandmark(
+  llmClient: UpstreamLlmClient,
+  input: SceneNpcGenerationInput,
+) {
+  const profile = normalizeProfile(input.profile);
+  const landmark = profile.landmarks.find((entry) => entry.id === input.landmarkId);
+  if (!landmark) {
+    throw badRequest('landmark not found');
+  }
+
+  const storyNpcById = new Map(profile.storyNpcs.map((npc) => [npc.id, npc]));
+  const sceneNpcs = landmark.sceneNpcIds
+    .map((npcId) => storyNpcById.get(npcId))
+    .filter((npc): npc is ParsedStoryNpc => Boolean(npc));
+  const otherNpcs = profile.storyNpcs.filter(
+    (npc) => !landmark.sceneNpcIds.includes(npc.id),
+  );
+  const fallbackDraft = buildFallbackDraft(profile, landmark, sceneNpcs);
+
+  try {
+    const content = await llmClient.requestMessageContent({
+      systemPrompt:
+        '你是游戏世界编辑器的角色生成器。你必须只返回可解析 JSON，不要输出解释、前言或 markdown 代码块之外的额外内容。',
+      userPrompt: buildPrompt(profile, landmark, sceneNpcs, otherNpcs),
+      debugLabel: 'custom-world-scene-npc',
+    });
+    const parsed = parseJsonResponseText(content);
+    const parsedRecord = toRecord(parsed);
+    const npcRecord = parsedRecord?.npc ?? parsed;
+    return sanitizeGeneratedNpc(npcRecord, profile, landmark, fallbackDraft);
+  } catch {
+    return sanitizeGeneratedNpc(fallbackDraft, profile, landmark, fallbackDraft);
+  }
+}

server-node/src/services/sceneImageService.test.ts

@@ -1,12 +1,12 @@
 import assert from 'node:assert/strict';
-import { createServer, type IncomingMessage, type ServerResponse } from 'node:http';
 import fs from 'node:fs';
+import { createServer, type IncomingMessage, type ServerResponse } from 'node:http';
 import os from 'node:os';
 import path from 'node:path';
 import test from 'node:test';
 
-import type { AppContext } from '../context.js';
 import { type AppConfig } from '../config.js';
+import type { AppContext } from '../context.js';
 import { generateSceneImage } from './sceneImageService.js';
 
 const PNG_BUFFER = Buffer.from(
@@ -24,7 +24,7 @@ function createTestConfig(
     dashScope: {
       baseUrl: dashScopeBaseUrl,
       apiKey: 'test-dashscope-key',
-      imageModel: 'wan2.7-image',
+      imageModel: 'wan2.2-t2i-flash',
       requestTimeoutMs: 5_000,
     },
   } as AppConfig;
@@ -92,11 +92,8 @@ async function withHttpServer<T>(
   }
 }
 
-test('generateSceneImage uploads a public reference image as a data url and saves the generated scene', async () => {
+test('generateSceneImage uses wan2.2-t2i-flash text-to-image payload and saves the generated scene', async () => {
   const tempRoot = fs.mkdtempSync(path.join(os.tmpdir(), 'genarrative-scene-image-'));
-  const publicDir = path.join(tempRoot, 'public');
-  fs.mkdirSync(path.join(publicDir, 'scene_bg'), { recursive: true });
-  fs.writeFileSync(path.join(publicDir, 'scene_bg', 'reference-layout.png'), PNG_BUFFER);
 
   const capturedRequests: Array<{
     pathname: string;
@@ -164,7 +161,6 @@ test('generateSceneImage uploads a public reference image as a data url and save
         profileId: 'world-1',
         landmarkName: '旧港灯塔',
         landmarkId: 'landmark-1',
-        referenceImageSrc: '/scene_bg/reference-layout.png',
       });
 
       assert.equal(result.ok, true);
@@ -177,6 +173,7 @@ test('generateSceneImage uploads a public reference image as a data url and save
       assert.ok(createRequest?.bodyText);
 
       const createPayload = JSON.parse(createRequest.bodyText) as {
+        model: string;
         input: {
           messages: Array<{
             content: Array<{ text?: string; image?: string }>;
@@ -188,8 +185,9 @@ test('generateSceneImage uploads a public reference image as a data url and save
       };
 
       const content = createPayload.input.messages[0]?.content ?? [];
+      assert.equal(createPayload.model, 'wan2.2-t2i-flash');
       assert.equal(content[0]?.text, '海雾港口像素风场景');
-      assert.match(content[1]?.image ?? '', /^data:image\/png;base64,/u);
+      assert.equal(content.length, 1);
       assert.equal(createPayload.parameters.negative_prompt, '模糊');
 
       const savedImagePath = path.join(tempRoot, 'public', result.imageSrc.slice(1));
@@ -197,3 +195,105 @@ test('generateSceneImage uploads a public reference image as a data url and save
     },
   );
 });
+
+test('generateSceneImage uses qwen-image-2.0 edit flow when a reference image is provided', async () => {
+  const tempRoot = fs.mkdtempSync(path.join(os.tmpdir(), 'genarrative-scene-image-'));
+  const publicDir = path.join(tempRoot, 'public');
+  fs.mkdirSync(path.join(publicDir, 'scene_bg'), { recursive: true });
+  fs.writeFileSync(path.join(publicDir, 'scene_bg', 'reference-layout.png'), PNG_BUFFER);
+
+  const capturedRequests: Array<{
+    pathname: string;
+    bodyText?: string;
+  }> = [];
+
+  await withHttpServer(
+    (baseUrl) => async (req, res) => {
+      const url = new URL(req.url || '/', baseUrl);
+      const bodyText =
+        req.method === 'POST' ? (await readRequestBody(req)).toString('utf8') : undefined;
+      capturedRequests.push({
+        pathname: url.pathname,
+        bodyText,
+      });
+
+      if (
+        req.method === 'POST' &&
+        url.pathname === '/api/v1/services/aigc/multimodal-generation/generation'
+      ) {
+        sendJson(res, {
+          output: {
+            choices: [
+              {
+                message: {
+                  content: [
+                    {
+                      image: `${baseUrl}/downloads/reference-scene.png`,
+                    },
+                  ],
+                },
+              },
+            ],
+          },
+        });
+        return;
+      }
+
+      if (req.method === 'GET' && url.pathname === '/downloads/reference-scene.png') {
+        res.statusCode = 200;
+        res.setHeader('Content-Type', 'image/png');
+        res.end(PNG_BUFFER);
+        return;
+      }
+
+      res.statusCode = 404;
+      res.end('not found');
+    },
+    async (dashScopeBaseUrl) => {
+      const context = {
+        config: createTestConfig(tempRoot, `${dashScopeBaseUrl}/api/v1`),
+      } as AppContext;
+
+      const result = await generateSceneImage(context, {
+        prompt: '废墟月台像素风场景',
+        negativePrompt: '模糊',
+        size: '1280*720',
+        worldName: '碎轨边境',
+        profileId: 'world-2',
+        landmarkName: '裂轨月台',
+        landmarkId: 'landmark-2',
+        referenceImageSrc: '/scene_bg/reference-layout.png',
+      });
+
+      assert.equal(result.ok, true);
+      assert.equal(result.model, 'qwen-image-2.0');
+      assert.match(result.taskId, /^scene-edit-/u);
+      assert.equal(
+        capturedRequests.some(
+          (entry) => entry.pathname === '/api/v1/tasks/scene-task-1',
+        ),
+        false,
+      );
+
+      const createRequest = capturedRequests.find(
+        (entry) =>
+          entry.pathname === '/api/v1/services/aigc/multimodal-generation/generation',
+      );
+      assert.ok(createRequest?.bodyText);
+
+      const createPayload = JSON.parse(createRequest.bodyText) as {
+        model: string;
+        input: {
+          messages: Array<{
+            content: Array<{ text?: string; image?: string }>;
+          }>;
+        };
+      };
+
+      const content = createPayload.input.messages[0]?.content ?? [];
+      assert.equal(createPayload.model, 'qwen-image-2.0');
+      assert.match(content[0]?.image ?? '', /^data:image\/png;base64,/u);
+      assert.equal(content[1]?.text, '废墟月台像素风场景');
+    },
+  );
+});

server-node/src/services/sceneImageService.ts

@@ -19,6 +19,8 @@ export const sceneImageSchema = z.object({
   landmarkId: z.string().trim().optional().default(''),
   referenceImageSrc: z.string().trim().optional().default(''),
 });
+const TEXT_TO_IMAGE_SCENE_MODEL = 'wan2.2-t2i-flash';
+const REFERENCE_IMAGE_SCENE_MODEL = 'qwen-image-2.0';
 
 function parseImageDataUrl(source: string) {
   const matched = /^data:(image\/[^;]+);base64,(.+)$/u.exec(source);
@@ -122,40 +124,72 @@ function extractImageUrls(payload: Record<string, unknown>) {
   return [...new Set(urls)];
 }
 
-function ensurePayload(
-  payload: z.infer<typeof sceneImageSchema>,
-  defaultModel: string,
-) {
-  if (!payload.landmarkName && !payload.landmarkId) {
-    throw badRequest('landmarkName 或 landmarkId 至少要提供一个');
+async function createSceneImageTask(params: {
+  baseUrl: string;
+  apiKey: string;
+  payload: z.infer<typeof sceneImageSchema>;
+}) {
+  const { baseUrl, apiKey, payload } = params;
+  const response = await fetch(`${baseUrl}/services/aigc/image-generation/generation`, {
+    method: 'POST',
+    headers: {
+      Authorization: `Bearer ${apiKey}`,
+      'Content-Type': 'application/json',
+      'X-DashScope-Async': 'enable',
+    },
+    body: JSON.stringify({
+      model: payload.model,
+      input: {
+        messages: [
+          {
+            role: 'user',
+            content: [{ text: payload.prompt }],
+          },
+        ],
+      },
+      parameters: {
+        n: 1,
+        size: payload.size,
+        prompt_extend: true,
+        watermark: false,
+        ...(payload.negativePrompt
+          ? { negative_prompt: payload.negativePrompt }
+          : {}),
+      },
+    }),
+  });
+  const responseText = await response.text();
+
+  if (!response.ok) {
+    return {
+      ok: false as const,
+      errorMessage: extractApiErrorMessage(
+        responseText,
+        '创建场景图片生成任务失败',
+      ),
+    };
   }
 
   return {
-    ...payload,
-    model: payload.model || defaultModel,
+    ok: true as const,
+    payload: JSON.parse(responseText) as Record<string, unknown>,
   };
 }
 
-export async function generateSceneImage(
-  context: AppContext,
-  input: z.infer<typeof sceneImageSchema>,
-) {
-  const payload = ensurePayload(input, context.config.dashScope.imageModel);
-  const baseUrl = context.config.dashScope.baseUrl.replace(/\/+$/u, '');
-  const referenceImage = payload.referenceImageSrc
-    ? await resolveReferenceImageAsDataUrl(
-        context.config.projectRoot,
-        payload.referenceImageSrc,
-      )
-    : '';
-  const createResponse = await fetch(
-    `${baseUrl}/services/aigc/image-generation/generation`,
+async function createSceneImageFromReference(params: {
+  baseUrl: string;
+  apiKey: string;
+  payload: z.infer<typeof sceneImageSchema>;
+  referenceImage: string;
+}) {
+  const { baseUrl, apiKey, payload, referenceImage } = params;
+  const response = await fetch(
+    `${baseUrl}/services/aigc/multimodal-generation/generation`,
     {
       method: 'POST',
       headers: {
-        Authorization: `Bearer ${context.config.dashScope.apiKey}`,
+        Authorization: `Bearer ${apiKey}`,
         'Content-Type': 'application/json',
-        'X-DashScope-Async': 'enable',
       },
       body: JSON.stringify({
         model: payload.model,
@@ -163,10 +197,7 @@ export async function generateSceneImage(
           messages: [
             {
               role: 'user',
-              content: [
-                { text: payload.prompt },
-                ...(referenceImage ? [{ image: referenceImage }] : []),
-              ],
+              content: [{ image: referenceImage }, { text: payload.prompt }],
             },
           ],
         },
@@ -182,56 +213,65 @@ export async function generateSceneImage(
       }),
     },
   );
-  const createText = await createResponse.text();
-  if (!createResponse.ok) {
-    throw badRequest(
-      extractApiErrorMessage(createText, '创建场景图片生成任务失败'),
-    );
-  }
-
-  const createPayload = JSON.parse(createText) as Record<string, unknown>;
-  const taskId = extractTaskId(createPayload);
-  if (!taskId) {
-    throw badRequest('场景图片生成任务未返回 task_id');
-  }
-
-  const deadline = Date.now() + context.config.dashScope.requestTimeoutMs;
-  let imageUrl = '';
-  let actualPrompt = '';
-
-  while (Date.now() < deadline) {
-    const pollResponse = await fetch(`${baseUrl}/tasks/${taskId}`, {
-      headers: {
-        Authorization: `Bearer ${context.config.dashScope.apiKey}`,
-      },
-    });
-    const pollText = await pollResponse.text();
-    if (!pollResponse.ok) {
-      throw badRequest(
-        extractApiErrorMessage(pollText, '查询场景图片任务失败'),
-      );
-    }
-
-    const pollPayload = JSON.parse(pollText) as Record<string, unknown>;
-    const status = findFirstStringByKey(pollPayload, 'task_status').trim();
-    if (status === 'SUCCEEDED') {
-      imageUrl = extractImageUrls(pollPayload)[0] ?? '';
-      actualPrompt = findFirstStringByKey(pollPayload, 'actual_prompt').trim();
-      break;
-    }
-    if (status === 'FAILED' || status === 'UNKNOWN') {
-      throw badRequest(
-        extractApiErrorMessage(pollText, '场景图片生成任务失败'),
-      );
-    }
-
-    await new Promise((resolve) => setTimeout(resolve, 2000));
+  const responseText = await response.text();
+
+  if (!response.ok) {
+    return {
+      ok: false as const,
+      errorMessage: extractApiErrorMessage(
+        responseText,
+        '创建参考图场景编辑任务失败',
+      ),
+    };
   }
 
+  const responsePayload = JSON.parse(responseText) as Record<string, unknown>;
+  const imageUrl = extractImageUrls(responsePayload)[0] ?? '';
   if (!imageUrl) {
-    throw badRequest('场景图片生成超时或未返回图片地址');
+    return {
+      ok: false as const,
+      errorMessage: '参考图场景编辑未返回图片地址',
+    };
   }
 
+  return {
+    ok: true as const,
+    imageUrl,
+    actualPrompt: findFirstStringByKey(responsePayload, 'actual_prompt').trim(),
+    taskId: `scene-edit-${Date.now()}`,
+  };
+}
+
+function ensurePayload(
+  payload: z.infer<typeof sceneImageSchema>,
+  _defaultModel: string,
+) {
+  if (!payload.landmarkName && !payload.landmarkId) {
+    throw badRequest('landmarkName 或 landmarkId 至少要提供一个');
+  }
+
+  const referenceImageSrc =
+    typeof payload.referenceImageSrc === 'string'
+      ? payload.referenceImageSrc.trim()
+      : '';
+
+  return {
+    ...payload,
+    referenceImageSrc,
+    model: referenceImageSrc
+      ? REFERENCE_IMAGE_SCENE_MODEL
+      : TEXT_TO_IMAGE_SCENE_MODEL,
+  };
+}
+
+async function saveSceneImageAsset(params: {
+  context: AppContext;
+  payload: z.infer<typeof sceneImageSchema>;
+  imageUrl: string;
+  taskId: string;
+  actualPrompt: string;
+}) {
+  const { context, payload, imageUrl, taskId, actualPrompt } = params;
   const imageResponse = await fetch(imageUrl);
   if (!imageResponse.ok) {
     throw badRequest('下载生成图片失败');
@@ -295,3 +335,99 @@ export async function generateSceneImage(
     actualPrompt,
   };
 }
+
+export async function generateSceneImage(
+  context: AppContext,
+  input: z.infer<typeof sceneImageSchema>,
+) {
+  const payload = ensurePayload(input, context.config.dashScope.imageModel);
+  const baseUrl = context.config.dashScope.baseUrl.replace(/\/+$/u, '');
+  const referenceImage = payload.referenceImageSrc.trim()
+    ? await resolveReferenceImageAsDataUrl(
+        context.config.projectRoot,
+        payload.referenceImageSrc,
+      )
+    : '';
+
+  if (referenceImage) {
+    const referenceResult = await createSceneImageFromReference({
+      baseUrl,
+      apiKey: context.config.dashScope.apiKey,
+      payload,
+      referenceImage,
+    });
+
+    if (!referenceResult.ok) {
+      throw badRequest(referenceResult.errorMessage);
+    }
+
+    return saveSceneImageAsset({
+      context,
+      payload,
+      imageUrl: referenceResult.imageUrl,
+      taskId: referenceResult.taskId,
+      actualPrompt: referenceResult.actualPrompt,
+    });
+  }
+
+  const createTaskResult = await createSceneImageTask({
+    baseUrl,
+    apiKey: context.config.dashScope.apiKey,
+    payload,
+  });
+
+  if (!createTaskResult.ok) {
+    throw badRequest(createTaskResult.errorMessage);
+  }
+
+  const createPayload = createTaskResult.payload;
+  const taskId = extractTaskId(createPayload);
+  if (!taskId) {
+    throw badRequest('场景图片生成任务未返回 task_id');
+  }
+
+  const deadline = Date.now() + context.config.dashScope.requestTimeoutMs;
+  let imageUrl = '';
+  let actualPrompt = '';
+
+  while (Date.now() < deadline) {
+    const pollResponse = await fetch(`${baseUrl}/tasks/${taskId}`, {
+      headers: {
+        Authorization: `Bearer ${context.config.dashScope.apiKey}`,
+      },
+    });
+    const pollText = await pollResponse.text();
+    if (!pollResponse.ok) {
+      throw badRequest(
+        extractApiErrorMessage(pollText, '查询场景图片任务失败'),
+      );
+    }
+
+    const pollPayload = JSON.parse(pollText) as Record<string, unknown>;
+    const status = findFirstStringByKey(pollPayload, 'task_status').trim();
+    if (status === 'SUCCEEDED') {
+      imageUrl = extractImageUrls(pollPayload)[0] ?? '';
+      actualPrompt = findFirstStringByKey(pollPayload, 'actual_prompt').trim();
+      break;
+    }
+    if (status === 'FAILED' || status === 'UNKNOWN') {
+      throw badRequest(
+        extractApiErrorMessage(pollText, '场景图片生成任务失败'),
+      );
+    }
+
+    await new Promise((resolve) => setTimeout(resolve, 2000));
+  }
+
+  if (!imageUrl) {
+    throw badRequest('场景图片生成超时或未返回图片地址');
+  }
+
+  return saveSceneImageAsset({
+    context,
+    payload,
+    imageUrl,
+    taskId,
+    actualPrompt,
+  });
+}