diff --git a/backend/internal/service/provider_ark_private_assets.go b/backend/internal/service/provider_ark_private_assets.go
index 87219c13b..d04b959d1 100644
--- a/backend/internal/service/provider_ark_private_assets.go
+++ b/backend/internal/service/provider_ark_private_assets.go
@@ -60,15 +60,19 @@ func (s *Service) prepareArkPrivateAssetReferences(ctx context.Context, userID s
 	if !hasOwnedReference {
 		return nil
 	}
-	if taskID := taskExecutionID(ctx); taskID != "" {
-		_ = s.repo.UpdateTaskProgress(taskID, "同步方舟可信素材", 36)
-	}
 	settingRecord, setting, err := s.readArkPrivateAssetSetting()
 	if err != nil {
 		return err
 	}
-	if !setting.Enabled || setting.AccessKeyID == "" || setting.AccessKeySecret == "" {
-		return errors.New("方舟可信素材库尚未配置，请由管理员在方舟素材库设置中填写启用的 IAM AK/SK")
+	automaticSyncEnabled, err := arkPrivateAssetAutomaticSyncEnabled(setting)
+	if err != nil {
+		return err
+	}
+	if !automaticSyncEnabled {
+		return nil
+	}
+	if taskID := taskExecutionID(ctx); taskID != "" {
+		_ = s.repo.UpdateTaskProgress(taskID, "同步方舟可信素材", 36)
 	}
 	for index := range input.ReferenceImages {
 		reference := &input.ReferenceImages[index]
@@ -106,6 +110,18 @@ func isArkPrivateAssetVideoConfig(config providerConfig) bool {
 	return config.InterfaceType == string(model.ChannelInterfaceVolcengineArkVideo) || isArkPlanVideoConfig(config)
 }
 
+func arkPrivateAssetAutomaticSyncEnabled(setting arkPrivateAssetSettingValue) (bool, error) {
+	// 可信素材同步是可选增强能力。管理员未启用时保持原始参考 URL，
+	// 继续执行常规火山方舟视频请求。
+	if !setting.Enabled {
+		return false, nil
+	}
+	if setting.AccessKeyID == "" || setting.AccessKeySecret == "" {
+		return false, errors.New("方舟可信素材库尚未配置，请由管理员在方舟素材库设置中填写启用的 IAM AK/SK")
+	}
+	return true, nil
+}
+
 func arkPrivateAssetResourceID(reference providerMedia) string {
 	if !strings.HasPrefix(reference.StorageKey, "resource:") {
 		return ""
diff --git a/backend/internal/service/provider_ark_private_assets_test.go b/backend/internal/service/provider_ark_private_assets_test.go
index cbb101c42..3408cad6f 100644
--- a/backend/internal/service/provider_ark_private_assets_test.go
+++ b/backend/internal/service/provider_ark_private_assets_test.go
@@ -96,6 +96,18 @@ func TestArkPrivateAssetControlPlaneDisablesGenerationAnalytics(t *testing.T) {
 	}
 }
 
+func TestArkPrivateAssetAutomaticSyncIsOptionalWhenSettingDisabled(t *testing.T) {
+	enabled, err := arkPrivateAssetAutomaticSyncEnabled(arkPrivateAssetSettingValue{})
+	if err != nil || enabled {
+		t.Fatalf("arkPrivateAssetAutomaticSyncEnabled() = %v, %v; want false, nil", enabled, err)
+	}
+
+	_, err = arkPrivateAssetAutomaticSyncEnabled(arkPrivateAssetSettingValue{Enabled: true})
+	if err == nil || !strings.Contains(err.Error(), "尚未配置") {
+		t.Fatalf("enabled incomplete setting error = %v, want configuration error", err)
+	}
+}
+
 func TestArkPrivateAssetSettingsEncryptSecret(t *testing.T) {
 	svc := &Service{dataDir: t.TempDir()}
 	value, err := arkPrivateAssetSettingFromRequest(ArkPrivateAssetSettingRequest{
diff --git a/web/src/lib/model-selection.ts b/web/src/lib/model-selection.ts
index dbcef2281..8cce7f8ab 100644
--- a/web/src/lib/model-selection.ts
+++ b/web/src/lib/model-selection.ts
@@ -319,9 +319,9 @@ export function modelGroupReferenceLimits(config: AiConfig, selected: string, ca
 
 export function inferVideoOperation(input: ModelInputSummary) {
     const visualInputCount = input.imageCount + input.characterCount;
-    // 音频参考有独立的协议能力；不能先归类为 reference_to_video，
-    // 否则只声明 audio_to_video 的细分模型永远无法被兼容路由选中。
-    if (input.audioCount > 0) return "audio_to_video";
+    // 纯音频参考使用独立能力；音频与图片、角色或视频组合时属于全模态参考，
+    // 不能把组合请求误路由到只支持 audio_to_video 的细分模型。
+    if (input.audioCount > 0) return visualInputCount > 0 || input.videoCount > 0 ? "reference_to_video" : "audio_to_video";
     if (input.videoCount > 0 || visualInputCount > 2) return "reference_to_video";
     if (visualInputCount > 0) return "image_to_video";
     return "text_to_video";
diff --git a/web/src/pages/create/index.tsx b/web/src/pages/create/index.tsx
index f889731b3..c6ae7b88d 100644
--- a/web/src/pages/create/index.tsx
+++ b/web/src/pages/create/index.tsx
@@ -20,7 +20,7 @@ import { useExternalAssetSources } from "@/hooks/use-external-asset-sources";
 import { buildImageResolutionOptions, formatImageResolutionSize, imageRatioForSize, imageResolutionChoices, imageResolutionOption, imageSizeForResolution, supportsImageResolutionPresets, type ImageResolutionChoice } from "@/lib/image-resolution-tiers";
 import { VIDEO_RESOLUTION_OPTIONS } from "@/lib/video-generation-options";
 import { modelCapabilityConfigFor, normalizeImageValue, normalizeVideoValue, videoDurationAllowed, videoDurationOptions, type ImageCapabilityConfig, type VideoCapabilityConfig } from "@/lib/model-capabilities";
-import { resolveCompatibleModel, mergedImageCapabilityConfig, type ModelRequirements } from "@/lib/model-selection";
+import { inferVideoOperation, resolveCompatibleModel, mergedImageCapabilityConfig, type ModelRequirements } from "@/lib/model-selection";
 import { backendModelRuntimeRequired, isGenerationTaskCancelled, runBackendGenerationTask, runBackendGenerationTaskBatch, type BackendGenerationResult } from "@/services/api/generation-task";
 import { requestImageQuestion, type AiTextContentPart } from "@/services/api/image";
 import { listAddedSkills, type Skill } from "@/services/api/skills";
@@ -487,6 +487,13 @@ export default function CreatePage() {
         const references = selectedCreationReferences(text, mentionReferences);
         // 后端对图片和视频使用不同的参考字段；这里先拆分，避免媒体类型在写入任务时被误判。
         const { referenceImages, referenceVideos, referenceAudios } = splitCreationAttachments(attachments);
+        const videoOperation = inferVideoOperation({
+            textCount: text ? 1 : 0,
+            imageCount: referenceImages.length,
+            videoCount: referenceVideos.length,
+            audioCount: referenceAudios.length,
+            characterCount: 0,
+        });
         const expandedPrompt = expandCreationPrompt(text, references, attachments);
         const referenceMetadata = creationReferenceMetadata(references);
         followLatestMessageRef.current = true;
@@ -627,7 +634,7 @@ export default function CreatePage() {
                     referenceVideos,
                     referenceAudios,
                     signal: requestLifecycle.signal,
-                    metadata: { source: "create-page", conversationId: activeConversation.id, messageId: assistantMessage.id, videoEditOperation: referenceAudios.length && !referenceImages.length && !referenceVideos.length ? "audio_to_video" : attachments.length ? "image_to_video" : "text_to_video", ...referenceMetadata },
+                    metadata: { source: "create-page", conversationId: activeConversation.id, messageId: assistantMessage.id, videoEditOperation: videoOperation, ...referenceMetadata },
                     onTaskUpdate: bindTask,
                     ...retryContext,
                 }));
@@ -647,7 +654,7 @@ export default function CreatePage() {
                 return;
             }
             const message = generationErrorMessage(error);
-            updateOriginAssistant((item) => ({ ...item, status: "error", error: message, generationErrorCode: item.generationErrorCode || generationErrorCode(error), generationOperation: item.generationOperation || (mode === "video" ? (attachments.length ? "image_to_video" : "text_to_video") : mode), createdAt: assistantMessage.createdAt, content: "生成失败" }));
+            updateOriginAssistant((item) => ({ ...item, status: "error", error: message, generationErrorCode: item.generationErrorCode || generationErrorCode(error), generationOperation: item.generationOperation || (mode === "video" ? videoOperation : mode), createdAt: assistantMessage.createdAt, content: "生成失败" }));
         } finally {
             requestLifecycle.release();
             releaseRetryLock();
diff --git a/web/test/canvas-model-policy.test.ts b/web/test/canvas-model-policy.test.ts
index 2f2878ef7..5ca703bdc 100644
--- a/web/test/canvas-model-policy.test.ts
+++ b/web/test/canvas-model-policy.test.ts
@@ -3,7 +3,7 @@ import { describe, expect, test } from "bun:test";
 import { canvasConnectionError } from "../src/lib/canvas/canvas-connection-policy";
 import { assertCanvasImageReferenceLimit, buildGenerationConfig, canvasImageReferenceLimitError, resolveCanvasGenerationModel } from "../src/lib/canvas/canvas-project-generation";
 import { defaultModelCapabilityConfig } from "../src/lib/model-capabilities";
-import { groupModelsByDisplayName, modelCompatibilityError, modelGroupReferenceLimits, resolveCompatibleModel, resolveModelGenerationDefaults } from "../src/lib/model-selection";
+import { groupModelsByDisplayName, inferVideoOperation, modelCompatibilityError, modelGroupReferenceLimits, resolveCompatibleModel, resolveModelGenerationDefaults } from "../src/lib/model-selection";
 import { defaultConfig, normalizeModelOptionValue, type AiConfig, type ModelChannel } from "../src/stores/use-config-store";
 import { CanvasNodeType, type CanvasConnection, type CanvasNodeData } from "../src/types/canvas";
 
@@ -167,6 +167,12 @@ describe("逻辑模型选择", () => {
         expect(modelCompatibilityError(config, "relay::cinema-image", requirements)).toContain("参考音频");
     });
 
+    test("音频仅在单独参考时归类为音频生视频", () => {
+        expect(inferVideoOperation({ textCount: 1, imageCount: 0, videoCount: 0, audioCount: 1, characterCount: 0 })).toBe("audio_to_video");
+        expect(inferVideoOperation({ textCount: 1, imageCount: 1, videoCount: 0, audioCount: 1, characterCount: 0 })).toBe("reference_to_video");
+        expect(inferVideoOperation({ textCount: 1, imageCount: 0, videoCount: 1, audioCount: 1, characterCount: 0 })).toBe("reference_to_video");
+    });
+
     test("逻辑视频模型将 720 与 720p 视为同一分辨率", () => {
         const model = "cinema-720p";
         const channel: ModelChannel = {
diff --git a/web/test/generation-task-local.test.ts b/web/test/generation-task-local.test.ts
index aa00490c5..3c46954f7 100644
--- a/web/test/generation-task-local.test.ts
+++ b/web/test/generation-task-local.test.ts
@@ -1318,6 +1318,15 @@ test("remote image video and audio references keep Backend parity without Dreami
             result: { mode: "video", video: { dataUrl: "opaque://video", storageKey: "resource:remote-video-output" } },
             operation: "reference_to_video",
         },
+        {
+            mode: "video" as const,
+            references: {
+                referenceImages: [{ id: "remote-image-audio-image-0001", name: "reference.png", type: "image/png", dataUrl: "", storageKey: "resource:remote-image-audio-image-0001" }],
+                referenceAudios: [{ id: "remote-image-audio-audio-0001", name: "reference.mp3", type: "audio/mpeg", url: "", storageKey: "resource:remote-image-audio-audio-0001" }],
+            },
+            result: { mode: "video", video: { dataUrl: "opaque://video", storageKey: "resource:remote-image-audio-output" } },
+            operation: "reference_to_video",
+        },
         {
             mode: "audio" as const,
             references: { referenceAudios: [{ id: "remote-audio-0001", name: "audio.mp3", type: "audio/mpeg", url: "", storageKey: "resource:remote-audio-0001" }] },
@@ -1677,7 +1686,7 @@ test("Create audio upload converts, previews, removes, and submits through the s
         operation?: string;
         input?: { referenceAudios?: Array<Record<string, unknown>> };
     };
-    expect(submitted.operation).toBe("reference_to_video");
+    expect(submitted.operation).toBe("audio_to_video");
     expect(submitted.input?.referenceAudios).toEqual([
         {
             id: attachment.id,