单图生成改为异步(Celery+AITask+前端轮询),与视频同架构

把独立生图从「请求内同步出图」改成异步任务,彻底根治整站 502:
- 后端 enqueue_standalone_images:Web 请求只建任务+预留额度(秒级),慢 ARK 出图派发给
  Celery worker(generate_standalone_image_task)。Web 层不再被 ~30s 请求占住 worker,
  健康探针不会被饿死 → 整站 502 从根上消失。
- run_standalone_image_task:worker 内单张出图,成功落库扣费/失败退费,幂等(只处理 RESERVED,
  重复投递不二次扣费)。
- 提交成功后浏览器关闭/断网,worker 仍把图生成并落库,重开素材库即可见(异步天然抗断连)。
- 视图 POST 返回 202+任务列表,新增 GET ?ids= 轮询状态并带回成图;前端 submit+poll。
- count 上限 4→12,与 UI 一致。

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

core/backend/apps/ai/services.py

@@ -819,16 +819,21 @@ _STANDALONE_TASK_TYPE = {
 }
 
 
-def generate_standalone_image(*, team, user, prompt: str, mode: str = "image", count: int = 1) -> list[Asset]:
+def enqueue_standalone_images(*, team, user, prompt: str, mode: str = "image", count: int = 1) -> list[AITask]:
-    """不绑定项目的独立生图(图片创作 / 模特上身图 / 平台套图)。复用项目内生图链路,AITask.project=None。"""
+    """独立生图(图片创作 / 模特上身图 / 平台套图)改为**异步**:本函数在 Web 请求里只做「建任务 +
+    预留额度」这种秒级的活,真正 ~30s 的 ARK 出图交给 Celery worker(generate_standalone_image_task)。
+
+    这样 Web 层(gunicorn)不会被慢出图请求占住 worker → 健康探针不会被饿死 → 根治"几张图就整站 502"。
+    且任务一旦提交(额度已预留),浏览器关掉 / 断网都不影响——worker 照样把图生成并落库,扣费/退费在
+    worker 内闭环。返回已 RESERVED 的 AITask 列表,前端拿 id 轮询 GET /api/ai/generate-image/?ids=… 取结果。"""
+    from apps.ai.tasks import generate_standalone_image_task
+
     model_config = get_default_model(ModelConfig.Capability.IMAGE)
     if model_config is None:
         raise ValueError("no active image model configured")
-    category = _STANDALONE_CATEGORY.get(mode, Asset.Category.UNCATEGORIZED)
     task_type = _STANDALONE_TASK_TYPE.get(mode, AITask.Type.PRODUCT_IMAGE)
-    count = max(1, min(int(count or 1), 4))
+    count = max(1, min(int(count or 1), 12))
-    provider = get_image_provider(model_config)
+    tasks: list[AITask] = []
-    assets: list[Asset] = []
     for index in range(count):
         cost = estimate_cost(model_config)
         task = AITask.objects.create(
@@ -839,42 +844,62 @@ def generate_standalone_image(*, team, user, prompt: str, mode: str = "image", c
             status=AITask.Status.CREATED,
             model_config=model_config,
             idempotency_key=f"standalone-image:{team.id}:{uuid.uuid4()}",
-            request_payload={"model": model_config.name, "endpoint": model_config.endpoint, "prompt": prompt, "mode": mode},
+            request_payload={"model": model_config.name, "endpoint": model_config.endpoint, "prompt": prompt, "mode": mode, "index": index},
             estimated_cost=cost,
         )
+        # 预留额度若余额不足会抛 ValueError,在同步的 Web 请求里立刻反馈给前端(不会先建半套任务)
         reserve_credit(team=team, user=user, task=task, amount=cost)
         task.status = AITask.Status.RESERVED
         task.save(update_fields=["status", "updated_at"])
-        reservation = task.credit_reservation
+        tasks.append(task)
-        try:
+    # 额度都预留成功后再统一派发,避免"派发了任务但后面某张预留失败"的半成品状态
-            response = provider.image_generation(model=model_config.name, endpoint=model_config.endpoint, prompt=prompt)
+    for task in tasks:
-            media = provider.extract_first_media_url(response)
+        generate_standalone_image_task.delay(str(task.id))
-            with transaction.atomic():
+    return tasks
-                task.status = AITask.Status.SUCCEEDED
+
-                task.response_payload = response
+
-                task.actual_cost = task.estimated_cost
+def run_standalone_image_task(*, task_id: str) -> None:
-                task.completed_at = timezone.now()
+    """Celery worker 内执行**单张**图的慢活:调 ARK → 成功落库扣费 / 失败退费。
-                task.save(update_fields=["status", "response_payload", "actual_cost", "completed_at", "updated_at"])
+    幂等:只处理 RESERVED 状态的任务,重复投递(celery retry / 重启重放)不会二次出图、二次扣费。"""
-                charge_reserved_credit(reservation=reservation, actual_amount=task.actual_cost)
+    task = AITask.objects.select_related("team", "created_by", "model_config").filter(id=task_id).first()
-                fileobj, content_type = VolcanoArkProvider.media_to_bytes(media)
+    if task is None or task.status != AITask.Status.RESERVED:
-                suffix = ".jpg" if "jpeg" in content_type else (".webp" if "webp" in content_type else ".png")
+        return
-                asset_id = uuid.uuid4()
+    team = task.team
-                object_key = f"teams/{team.id}/standalone/{asset_id}{suffix}"
+    user = task.created_by
-                stored = TosStorage().upload_fileobj(fileobj=fileobj, object_key=object_key, content_type=content_type)
+    payload = task.request_payload or {}
-                asset = Asset.objects.create(
+    prompt = str(payload.get("prompt") or "")
-                    id=asset_id, team=team, created_by=user, name=f"AI 生成 · {mode} · {index + 1}",
+    mode = str(payload.get("mode") or "image")
-                    asset_type=Asset.Type.IMAGE, source=Asset.Source.AI_GENERATED, category=category, origin_task=task,
+    index = int(payload.get("index") or 0)
-                )
+    category = _STANDALONE_CATEGORY.get(mode, Asset.Category.UNCATEGORIZED)
-                AssetFile.objects.create(asset=asset, object_key=stored.object_key, bucket=stored.bucket, content_type=stored.content_type, size_bytes=stored.size_bytes, is_primary=True)
+    model_config = task.model_config
-                assets.append(asset)
+    provider = get_image_provider(model_config)
-        except Exception as exc:
+    reservation = task.credit_reservation
-            task.status = AITask.Status.FAILED
+    try:
-            task.error_message = str(exc)
+        response = provider.image_generation(model=model_config.name, endpoint=model_config.endpoint, prompt=prompt)
+        media = provider.extract_first_media_url(response)
+        with transaction.atomic():
+            task.status = AITask.Status.SUCCEEDED
+            task.response_payload = response
+            task.actual_cost = task.estimated_cost
             task.completed_at = timezone.now()
-            task.save(update_fields=["status", "error_message", "completed_at", "updated_at"])
+            task.save(update_fields=["status", "response_payload", "actual_cost", "completed_at", "updated_at"])
-            release_credit(reservation=reservation, reason=str(exc))
+            charge_reserved_credit(reservation=reservation, actual_amount=task.actual_cost)
-            raise
+            fileobj, content_type = VolcanoArkProvider.media_to_bytes(media)
-    return assets
+            suffix = ".jpg" if "jpeg" in content_type else (".webp" if "webp" in content_type else ".png")
+            asset_id = uuid.uuid4()
+            object_key = f"teams/{team.id}/standalone/{asset_id}{suffix}"
+            stored = TosStorage().upload_fileobj(fileobj=fileobj, object_key=object_key, content_type=content_type)
+            asset = Asset.objects.create(
+                id=asset_id, team=team, created_by=user, name=f"AI 生成 · {mode} · {index + 1}",
+                asset_type=Asset.Type.IMAGE, source=Asset.Source.AI_GENERATED, category=category, origin_task=task,
+            )
+            AssetFile.objects.create(asset=asset, object_key=stored.object_key, bucket=stored.bucket, content_type=stored.content_type, size_bytes=stored.size_bytes, is_primary=True)
+    except Exception as exc:  # noqa: BLE001 — 失败要退费并把错误记进 AITask 供前端轮询读取;不向上抛(避免 celery 重试二次扣费)
+        task.status = AITask.Status.FAILED
+        task.error_message = str(exc)
+        task.completed_at = timezone.now()
+        task.save(update_fields=["status", "error_message", "completed_at", "updated_at"])
+        release_credit(reservation=reservation, reason=str(exc))
 
 
 # ── 旁白配音(TTS):每镜旁白合成一段语音,导出时作为人声轨混在 BGM 之上 ──

core/backend/apps/ai/tasks.py

@@ -10,3 +10,13 @@ def submit_ai_task(self, task_id: str) -> str:
 def poll_ai_task(self, task_id: str) -> str:
     return task_id
 
+
+@app.task(bind=True, max_retries=0)
+def generate_standalone_image_task(self, task_id: str) -> str:
+    """单张独立图的慢活(ARK 出图 ~30s)在 worker 内跑,Web 层不被占住。
+    幂等且失败自退费(见 run_standalone_image_task),故 max_retries=0,不向上抛重试。"""
+    from apps.ai.services import run_standalone_image_task
+
+    run_standalone_image_task(task_id=task_id)
+    return task_id
+

core/backend/apps/ai/views.py

@@ -9,14 +9,18 @@ from apps.common.celery_health import require_worker
 
 from .models import AITask, ModelConfig
 from .serializers import AITaskSerializer, ModelConfigSerializer
-from .services import generate_standalone_image
+from .services import enqueue_standalone_images
 
 
 class GenerateImageView(APIView):
-    """POST /api/ai/generate-image/ — 独立生图(不绑项目)· 图片创作/模特图/平台套图共用。"""
+    """独立生图(不绑项目)· 图片创作/模特图/平台套图共用 —— **异步**。
+
+    POST /api/ai/generate-image/        提交生成,秒级返回 RESERVED 任务列表(慢出图交给 worker)。
+    GET  /api/ai/generate-image/?ids=…  轮询这些任务的状态;成功的任务带回成图 asset。
+    """
 
     def post(self, request):
-        require_worker()
+        require_worker()  # 异步出图依赖 worker 兜底执行,没 worker 直接拒绝(否则任务永远 RESERVED)
         prompt = str(request.data.get("prompt") or "").strip()
         if not prompt:
             return Response({"detail": "prompt 不能为空"}, status=status.HTTP_400_BAD_REQUEST)
@@ -27,12 +31,34 @@ class GenerateImageView(APIView):
             count = 1
         team = get_current_team(request.user)
         try:
-            assets = generate_standalone_image(team=team, user=request.user, prompt=prompt, mode=mode, count=count)
+            tasks = enqueue_standalone_images(team=team, user=request.user, prompt=prompt, mode=mode, count=count)
-        except ValueError as exc:
+        except ValueError as exc:  # 无可用模型 / 余额不足等,立即反馈
             return Response({"detail": str(exc)}, status=status.HTTP_400_BAD_REQUEST)
-        except Exception as exc:  # noqa: BLE001 — 生成失败已回滚额度,返回明确错误给前端
+        return Response(
-            return Response({"detail": f"生成失败: {exc}"}, status=status.HTTP_502_BAD_GATEWAY)
+            {"tasks": [{"id": str(t.id), "status": t.status} for t in tasks]},
-        return Response({"assets": AssetSerializer(assets, many=True).data}, status=status.HTTP_201_CREATED)
+            status=status.HTTP_202_ACCEPTED,
+        )
+
+    def get(self, request):
+        team = get_current_team(request.user)
+        ids = [s for s in str(request.query_params.get("ids") or "").split(",") if s]
+        if not ids:
+            return Response({"tasks": []})
+        tasks = AITask.objects.filter(team=team, id__in=ids).prefetch_related(
+            "generated_assets", "generated_assets__files"
+        )
+        data = [
+            {
+                "id": str(t.id),
+                "status": t.status,
+                "error_message": t.error_message,
+                "assets": AssetSerializer(
+                    [a for a in t.generated_assets.all() if not a.is_deleted], many=True
+                ).data,
+            }
+            for t in tasks
+        ]
+        return Response({"tasks": data})
 
 
 class AITaskViewSet(TeamScopedViewSetMixin, ReadOnlyModelViewSet):

core/frontend/src/App.tsx

@@ -355,29 +355,30 @@ export function App() {
   }
 
   function generateImages(payload: { prompt: string; mode?: "image" | "model" | "cover"; count?: number }) {
-    // 每张图一条独立 HTTP 请求(不再让一条请求串行出 N 张)——单张约 30s 不会触发后端请求超时。
+    // 异步生图:提交后立刻拿到任务列表,前端轮询直到出图。慢的 ARK 出图在 Celery worker 里跑——
-    // 但不全量并行:同时最多 CONCURRENCY 条,避免 N 条一起砸向单个后端 pod 致内存峰值翻倍 OOM→502。
+    // Web 层不被 ~30s 请求占住 → 健康探针不饿死 → 根治"几张图整站 502";且提交成功后浏览器关掉/断网,
-    // 任一张失败只丢那一张(各自独立扣费/回滚),其余正常返回。
+    // worker 仍会把图生成并落库(扣费/退费在 worker 内闭环),重开素材库即可见。
-    const n = Math.max(1, Math.min(Number(payload.count) || 1, 12));
-    const CONCURRENCY = 2;
     return action(async () => {
-      const results: Array<{ ok: true; assets: Asset[] } | { ok: false; reason: unknown }> = new Array(n);
+      const { tasks } = await api.submitGenerateImage(payload);
-      let next = 0;
+      const pending = new Set(tasks.map((t) => t.id));
-      async function worker() {
+      if (pending.size === 0) throw new Error("未能提交生成任务");
-        for (let i = next++; i < n; i = next++) {
+      const TERMINAL = new Set(["succeeded", "failed", "cancelled", "compensating"]);
-          try {
+      const sleep = (ms: number) => new Promise((resolve) => setTimeout(resolve, ms));
-            const res = await api.generateImage({ prompt: payload.prompt, mode: payload.mode, count: 1 });
+      const assets: Asset[] = [];
-            results[i] = { ok: true, assets: res.assets };
+      let lastErr = "";
-          } catch (reason) {
+      const deadline = Date.now() + 6 * 60 * 1000; // 兜底:6 分钟还没全好就停止轮询(图仍会在后台出完)
-            results[i] = { ok: false, reason };
+      while (pending.size > 0 && Date.now() < deadline) {
-          }
+        await sleep(2500);
+        const res = await api.generateImageStatus([...pending]);
+        for (const t of res.tasks) {
+          if (!TERMINAL.has(t.status)) continue;
+          pending.delete(t.id);
+          if (t.status === "succeeded") assets.push(...(t.assets || []));
+          else if (t.error_message) lastErr = t.error_message;
         }
       }
-      await Promise.all(Array.from({ length: Math.min(CONCURRENCY, n) }, worker));
-      const assets = results.flatMap((r) => (r.ok ? r.assets : []));
       if (assets.length === 0) {
-        const firstErr = results.find((r) => !r.ok) as { ok: false; reason: unknown } | undefined;
+        throw new Error(lastErr || (pending.size > 0 ? "生成超时,图片仍在后台生成,稍后可在素材库查看" : "未生成任何图片"));
-        throw new Error(firstErr ? String((firstErr.reason as Error)?.message || firstErr.reason || "生成失败") : "未生成任何图片");
       }
       return { assets };
     }, "图片已生成");

core/frontend/src/api.ts

@@ -296,8 +296,12 @@ export const api = {
   aiTasks() {
     return request<Paginated<AITask>>("/api/ai/tasks/");
   },
-  generateImage(payload: { prompt: string; mode?: "image" | "model" | "cover"; count?: number }) {
+  // 异步生图:提交后秒级返回任务列表(慢出图在 worker 跑),再用 generateImageStatus 轮询取结果
-    return request<{ assets: Asset[] }>("/api/ai/generate-image/", { method: "POST", body: JSON.stringify(payload) });
+  submitGenerateImage(payload: { prompt: string; mode?: "image" | "model" | "cover"; count?: number }) {
+    return request<{ tasks: { id: string; status: string }[] }>("/api/ai/generate-image/", { method: "POST", body: JSON.stringify(payload) });
+  },
+  generateImageStatus(ids: string[]) {
+    return request<{ tasks: { id: string; status: string; error_message: string; assets: Asset[] }[] }>(`/api/ai/generate-image/?ids=${encodeURIComponent(ids.join(","))}`);
   },
   recharge(payload: { amount: number | string; bonus?: number | string; channel?: string }) {
     return request<RechargeResult>("/api/billing/recharge/", { method: "POST", body: JSON.stringify(payload) });