feat: v0.14.2 推理接入点EP + 参考图片上限9张 + reEdit标签修复

- 推理接入点：model字段优先使用EP接入点ID（ARK_ENDPOINT_SEEDANCE环境变量），无EP降级到模型ID
- 参考图片上限：提交时校验image类型不超过9张，超限返回友好中文提示
- 上传图片标签编号：接着已有素材编号，不再从1重新计数
- 轮询同步assetMentions：polling完成时同时更新references和assetMentions
- reEdit标签修复：用纯文本prompt重建标签，避免blob:URL失效导致图片标签丢失

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

backend/apps/generation/views.py

@@ -8,7 +8,7 @@ from rest_framework.response import Response
 from django.contrib.auth import get_user_model
 from django.utils import timezone
 from django.db import transaction
-from django.db.models import Sum, Q, F, Count, Exists, OuterRef
+from django.db.models import Sum, Q, F, Count, Exists, OuterRef, Case, When, Value
 from django.db.models.functions import TruncDate
 from django.db.utils import OperationalError as DbOperationalError
 from datetime import timedelta
@@ -269,12 +269,51 @@ def video_generate_view(request):
 
         # 构建参考素材
         references = request.data.get('references', [])
+        # 火山限制最多 9 张参考图片
+        image_count = sum(1 for r in references if r.get('type', 'image') == 'image')
+        if image_count > 9:
+            return Response({
+                'error': 'too_many_references',
+                'message': f'参考图片最多 9 张，当前 {image_count} 张，请减少后重试',
+            }, status=status.HTTP_400_BAD_REQUEST)
+
         reference_snapshots = []
         content_items = []
         seen_urls = set()  # 去重：同一个素材只引用一次
+        _asset_cache = {}  # group_id → resolved_url，避免同一素材组重复查询
 
         from .models import Asset as AssetModel
 
+        def _resolve_asset_group(gid, lbl):
+            """查询本地 DB + 必要时刷新火山状态，返回 Asset://xxx 或原始 asset:// URL。"""
+            asset = AssetModel.objects.filter(
+                group_id=gid, status__in=['active', 'processing']
+            ).order_by(
+                Case(When(status='active', then=0), default=1)
+            ).first()
+            if not asset or not asset.remote_asset_id:
+                logger.warning('No asset found for group %s (label=%s)', gid, lbl)
+                return f'asset://group-{gid}'
+            # 本地 processing → 实时查火山刷新
+            if asset.status == 'processing':
+                result, _ = _assets_api_call(assets_client.get_asset, asset.remote_asset_id)
+                if result and result.get('Status') == 'Active':
+                    asset.status = 'active'
+                    asset.url = result.get('Url', asset.url)
+                    asset.save(update_fields=['status', 'url'])
+                    logger.info('Asset %s refreshed to active from Volcano', asset.remote_asset_id)
+                else:
+                    logger.warning('Asset %s still processing on Volcano', asset.remote_asset_id)
+                    return f'asset://group-{gid}'
+            aid = asset.remote_asset_id
+            if aid.startswith('asset-'):
+                aid = 'Asset-' + aid[6:]
+            resolved = f'Asset://{aid}'
+            logger.info('Asset resolved: group=%s -> %s', gid, resolved)
+            return resolved
+
+        from utils import assets_client
+
         for ref in references:
             url = ref.get('url', '')
             original_url = url  # 保留原始 URL 用于 reference_snapshots
@@ -299,27 +338,29 @@ def video_generate_view(request):
             if url.startswith('asset://group-'):
                 try:
                     group_id = int(url.replace('asset://group-', ''))
-                    first_asset = AssetModel.objects.filter(
-                        group_id=group_id, status='active'
-                    ).first()
-                    if first_asset and first_asset.remote_asset_id:
-                        aid = first_asset.remote_asset_id
-                        if aid.startswith('asset-'):
-                            aid = 'Asset-' + aid[6:]
-                        resolved_url = f'Asset://{aid}'
+                    # 跨迭代缓存：同一 group_id 不重复查询/刷新
+                    if group_id in _asset_cache:
+                        resolved_url = _asset_cache[group_id]
                     else:
-                        logger.warning('No active asset found for group %s', group_id)
-                except (ValueError, Exception) as e:
+                        resolved_url = _resolve_asset_group(group_id, label)
+                        _asset_cache[group_id] = resolved_url
+                except Exception as e:
                     logger.warning('Failed to resolve asset group URL %s: %s', url, e)
 
-            # 未解析成功的 asset URL 不发给火山 API（会导致 InvalidParameter）
+            # 未解析成功的 asset URL → 返回明确错误，不再静默跳过
             if resolved_url.startswith('asset://'):
-                logger.warning('Skipping unresolved asset URL: %s', resolved_url)
-                continue
+                logger.error('Unresolved asset URL: %s (label=%s)', resolved_url, label)
+                return Response({
+                    'error': 'asset_not_ready',
+                    'message': f'素材「{label}」尚未就绪，请在素材库中确认状态为"可用"后重试',
+                }, status=status.HTTP_400_BAD_REQUEST)
 
             if ref_type == 'image':
                 item = {'type': 'image_url', 'image_url': {'url': resolved_url}}
-                if role:
+                # API 文档要求：参考图模式下所有图片的 role 必须为 reference_image
+                if mode == 'universal':
+                    item['role'] = 'reference_image'
+                elif role:
                     item['role'] = role
                 content_items.append(item)
             elif ref_type == 'video':
@@ -333,6 +374,8 @@ def video_generate_view(request):
                     item['role'] = role
                 content_items.append(item)
 
+        logger.info('Video generate: %d content_items built (prompt=%s...)', len(content_items), prompt[:60])
+
         # 冻结（不扣余额）
         record = GenerationRecord.objects.create(
             user=user,

backend/config/settings.py

@@ -226,6 +226,9 @@ TOS_CDN_DOMAIN = os.environ.get('TOS_CDN_DOMAIN', 'https://airdrama-media.tos-cn
 # ──────────────────────────────────────────────
 ARK_API_KEY = os.environ.get('ARK_API_KEY', '')
 ARK_BASE_URL = os.environ.get('ARK_BASE_URL', 'https://ark.cn-beijing.volces.com/api/v3')
+# 推理接入点 ID（优先使用，为空时降级到模型 ID）
+ARK_ENDPOINT_SEEDANCE = os.environ.get('ARK_ENDPOINT_SEEDANCE', '')
+ARK_ENDPOINT_SEEDANCE_FAST = os.environ.get('ARK_ENDPOINT_SEEDANCE_FAST', '')
 # Set to True when Seedance model is activated on ARK platform
 SEEDANCE_ENABLED = os.environ.get('SEEDANCE_ENABLED', 'false').lower() == 'true'
 # Set to True to enable the Assets API (virtual avatar library)

backend/utils/airdrama_client.py

@@ -68,6 +68,17 @@ MODEL_MAP = {
     'seedance_2.0_fast': 'doubao-seedance-2-0-fast-260128',
 }
 
+# 推理接入点优先：有 EP 用 EP，没有降级到模型 ID
+def _resolve_model(model):
+    ep_map = {
+        'seedance_2.0': settings.ARK_ENDPOINT_SEEDANCE,
+        'seedance_2.0_fast': settings.ARK_ENDPOINT_SEEDANCE_FAST,
+    }
+    ep = ep_map.get(model, '')
+    if ep:
+        return ep
+    return MODEL_MAP.get(model, model)
+
 
 def _headers():
     return {
@@ -100,7 +111,7 @@ def create_task(prompt, model, content_items, aspect_ratio, duration,
     content.extend(content_items)
 
     payload = {
-        'model': MODEL_MAP.get(model, model),
+        'model': _resolve_model(model),
         'content': content,
         'generate_audio': generate_audio,
         'ratio': aspect_ratio,
@@ -115,6 +126,10 @@ def create_task(prompt, model, content_items, aspect_ratio, duration,
     import logging
     logger = logging.getLogger(__name__)
     logger.info('AirDrama API payload: %s', {k: v for k, v in payload.items() if k != 'content'})
+    # 记录 content 中的非文本项，方便排查素材引用问题
+    media_items = [ci for ci in content if ci.get('type') != 'text']
+    if media_items:
+        logger.info('AirDrama content media items (%d): %s', len(media_items), media_items)
 
     resp = requests.post(url, json=payload, headers=_headers(), timeout=60)
     if resp.status_code != 200:

k8s/backend-deployment.yaml

@@ -92,6 +92,8 @@ spec:
             secretKeyRef:
               name: video-backend-secrets
               key: ARK_API_KEY
+        - name: ARK_ENDPOINT_SEEDANCE
+          value: "ep-m-20260315211214-z9dp6"
         - name: SEEDANCE_ENABLED
           value: "true"
         - name: ASSETS_API_ENABLED

k8s/celery-deployment.yaml

@@ -81,6 +81,10 @@ spec:
             secretKeyRef:
               name: video-backend-secrets
               key: ARK_API_KEY
+        - name: ARK_ENDPOINT_SEEDANCE
+          value: "ep-m-20260315211214-z9dp6"
+        - name: SEEDANCE_ENABLED
+          value: "true"
         resources:
           requests:
             memory: "128Mi"

web/src/store/generation.ts

@@ -54,49 +54,46 @@ function mapProgress(backendStatus: string): number {
   return 50;
 }
 
+/** Check if a URL is an asset library reference (case-insensitive protocol). */
+function isAssetUrl(url: string): boolean {
+  return url.startsWith('asset://') || url.startsWith('Asset://');
+}
+
+/** Build ReferenceSnapshot[] from raw reference_urls, excluding asset refs. */
+function buildReferenceSnapshots(
+  refs: Array<Record<string, string>>,
+  taskId: string,
+): ReferenceSnapshot[] {
+  return refs
+    .filter((ref) => {
+      const url = ref.url || '';
+      return !isAssetUrl(url) && url.trim() !== '';
+    })
+    .map((ref, i) => ({
+      id: `ref_${taskId}_${i}`,
+      type: (ref.type || 'image') as 'image' | 'video' | 'audio',
+      previewUrl: ref.url || '',
+      label: ref.label || `素材${i + 1}`,
+      role: ref.role,
+    }));
+}
+
+/** Extract asset mention metadata from raw reference_urls. */
+function buildAssetMentions(refs: Array<Record<string, string>>) {
+  return refs
+    .filter((ref) => isAssetUrl(ref.url || ''))
+    .map((ref) => {
+      const url = ref.url || '';
+      const groupId = url.startsWith('asset://group-') ? url.replace('asset://group-', '') : '';
+      return { groupId, label: ref.label || '', thumbUrl: ref.thumb_url || '' };
+    });
+}
+
 // Convert a BackendTask to a frontend GenerationTask
 function backendToFrontend(bt: BackendTask): GenerationTask {
   const allRefs = bt.reference_urls || [];
-
-  // 普通引用（参考图/视频/音频）— 可显示的缩略图
-  const references: ReferenceSnapshot[] = allRefs
-    .filter((ref) => {
-      const url = ref.url || '';
-      const thumbUrl = (ref as Record<string, string>).thumb_url || '';
-      // 保留有真实 URL 或有缩略图的引用
-      if (thumbUrl) return true;
-      if (!url || url.trim() === '') return false;
-      if (url.startsWith('asset://') || url.startsWith('Asset://')) return false;
-      return true;
-    })
-    .map((ref, i) => {
-      const url = ref.url || '';
-      const thumbUrl = (ref as Record<string, string>).thumb_url || '';
-      // 素材库引用用 thumb_url，普通上传用 url
-      const displayUrl = (url.startsWith('asset://') || url.startsWith('Asset://')) ? thumbUrl : url;
-      return {
-        id: `ref_${bt.task_id}_${i}`,
-        type: (ref.type || 'image') as 'image' | 'video',
-        previewUrl: displayUrl,
-        label: ref.label || `素材${i + 1}`,
-        role: ref.role,
-      };
-    });
-
-  // Asset 引用 — 仅用于 reEdit/regenerate 重建 mention span
-  const assetMentions = allRefs
-    .filter((ref) => {
-      const url = ref.url || '';
-      return url.startsWith('asset://') || url.startsWith('Asset://');
-    })
-    .map((ref) => {
-      const url = ref.url || '';
-      let groupId = '';
-      if (url.startsWith('asset://group-')) {
-        groupId = url.replace('asset://group-', '');
-      }
-      return { groupId, label: ref.label || '', thumbUrl: (ref as Record<string, string>).thumb_url || '' };
-    });
+  const references = buildReferenceSnapshots(allRefs, bt.task_id);
+  const assetMentions = buildAssetMentions(allRefs);
 
   return {
     id: `backend_${bt.task_id}`,
@@ -169,28 +166,10 @@ function startPolling(taskId: string, frontendId: string) {
         const { data } = await videoApi.getTaskStatus(taskId);
         const newStatus = mapStatus(data.status);
 
-        // Parse references from polling response (includes thumb_url for asset refs)
-        const pollRefs: ReferenceSnapshot[] = (data.reference_urls || [])
-          .filter((ref: Record<string, string>) => {
-            const url = ref.url || '';
-            const thumbUrl = ref.thumb_url || '';
-            if (thumbUrl) return true;
-            if (!url || url.trim() === '') return false;
-            if (url.startsWith('asset://') || url.startsWith('Asset://')) return false;
-            return true;
-          })
-          .map((ref: Record<string, string>, i: number) => {
-            const url = ref.url || '';
-            const thumbUrl = ref.thumb_url || '';
-            const displayUrl = (url.startsWith('asset://') || url.startsWith('Asset://')) ? thumbUrl : url;
-            return {
-              id: `ref_${taskId}_${i}`,
-              type: (ref.type || 'image') as 'image' | 'video' | 'audio',
-              previewUrl: displayUrl,
-              label: ref.label || `素材${i + 1}`,
-              role: ref.role,
-            };
-          });
+        // Parse references from polling response
+        const pollAllRefs = data.reference_urls || [];
+        const pollRefs = buildReferenceSnapshots(pollAllRefs, taskId);
+        const pollAssetMentions = buildAssetMentions(pollAllRefs);
 
         useGenerationStore.setState((s) => ({
           tasks: s.tasks.map((t) =>
@@ -205,6 +184,7 @@ function startPolling(taskId: string, frontendId: string) {
                   costAmount: data.cost_amount ?? t.costAmount,
                   seed: data.seed ?? t.seed,
                   references: pollRefs.length > 0 ? pollRefs : t.references,
+                  assetMentions: pollAssetMentions.length > 0 ? pollAssetMentions : t.assetMentions,
                 }
               : t
           ),
@@ -596,11 +576,14 @@ export const useGenerationStore = create<GenerationState>((set, get) => ({
         label: r.label,
         tosUrl: r.previewUrl,
       }));
+
+      // Always use plain text prompt for reEdit — let PromptInput's rebuildMentionSpans
+      // reconstruct tags from references + assetMentions (avoids dead blob: URLs)
       const taskSeed = task.seed ?? -1;
       const currentSeedEnabled = useInputBarStore.getState().seedEnabled;
       useInputBarStore.setState({
         prompt: task.prompt,
-        editorHtml: task.editorHtml || task.prompt,
+        editorHtml: task.prompt,
         aspectRatio: task.aspectRatio,
         duration: task.duration,
         references,

web/src/store/inputBar.ts

@@ -135,12 +135,16 @@ export const useInputBarStore = create<InputBarState>((set, get) => ({
 
       fileCounter++;
       const labelPrefix = type === 'video' ? '视频' : type === 'audio' ? '音频' : '图片';
+      // 编号接着已有的同类型素材（包括 @ 引用的 assetMentions）
+      const existingSameType = state.references.filter(r => r.type === type).length
+        + newRefs.filter(r => r.type === type).length
+        + (state.assetMentions || []).filter((m: Record<string, string>) => m.type === type).length;
       newRefs.push({
         id: `ref_${fileCounter}`,
         file,
         type,
         previewUrl: type === 'audio' ? '' : URL.createObjectURL(file),
-        label: `${labelPrefix}${fileCounter}`,
+        label: `${labelPrefix}${existingSameType + 1}`,
       });
       counts[type]++;
     }