diff --git a/short_drama.db b/short_drama.db
index b9ff6fa..da91d6d 100644
Binary files a/short_drama.db and b/short_drama.db differ
diff --git a/shortdrama/service/drama_service.go b/shortdrama/service/drama_service.go
index 94ba184..9ba13bc 100644
--- a/shortdrama/service/drama_service.go
+++ b/shortdrama/service/drama_service.go
@@ -392,7 +392,7 @@ func (s *dramaService) generateOneSegment(ctx context.Context, d *entity.Drama,
 	// JSON 镜头脚本跳过 Agent，直接构建场景描述提交视频模型
 	var segOutput *model.SegmentOutput
 	if domain.IsShotsJSON(ep.Script) && feedback == "" {
-		segOutput = buildSegOutputFromShots(ep.Script, segIdx, segStartTime, segDur)
+		segOutput = buildSegOutputFromShots(ep.Script, segIdx, segStartTime, segDur, genCtx)
 		g.Log().Infof(ctx, "第%d集第%d段 JSON镜头直接提交: %d个镜头", ep.Index, segIdx+1, len(segOutput.Scenes))
 	} else {
 		result, err := s.generateSegment(ctx, d, ep, segIdx, segDur, segStartTime, totalSegs, modelCfg, feedback, genCtx)
@@ -556,7 +556,7 @@ func (s *dramaService) generateOneSegment(ctx context.Context, d *entity.Drama,
 
 // buildSegOutputFromShots 从 JSON 镜头数组直接构建 SegmentOutput（跳过 Agent）
 // 将本段所有镜头合并为一段连贯的场景描述，去掉"旁白/台词/事件"标记，适合视频模型理解
-func buildSegOutputFromShots(script string, segIdx, segStartTime, segDur int) *model.SegmentOutput {
+func buildSegOutputFromShots(script string, segIdx, segStartTime, segDur int, genCtx *GenerationContext) *model.SegmentOutput {
 	var allShots []domain.Shot
 	if err := json.Unmarshal([]byte(script), &allShots); err != nil || len(allShots) == 0 {
 		return &model.SegmentOutput{Index: segIdx, TextOutput: script}
@@ -584,23 +584,80 @@ func buildSegOutputFromShots(script string, segIdx, segStartTime, segDur int) *m
 		chars = append(chars, model.SegmentCharacter{Name: name})
 	}
 
-	// 将所有镜头合并为一段连贯描述（去掉旁白/台词/事件标记，自然语言拼接）
+	// 从 genCtx 构建角色外观描述（确保每段角色服装/外貌一致）
+	var appearanceDescs []string
+	for name := range charSet {
+		for _, c := range genCtx.Characters {
+			if c.Name == name && c.Description != "" {
+				appearanceDescs = append(appearanceDescs, name+"："+c.Description)
+				break
+			}
+		}
+	}
+	// 收集本段涉及场景的描述（确保各段画风一致）
+	var sceneDescs []string
+	seenScenes := make(map[string]bool)
+	for _, sh := range segShots {
+		if sh.Scene == "" || seenScenes[sh.Scene] {
+			continue
+		}
+		seenScenes[sh.Scene] = true
+		for _, sc := range genCtx.Scenes {
+			if sc.Name == sh.Scene && sc.Description != "" {
+				sceneDescs = append(sceneDescs, sc.Name+"："+sc.Description)
+				break
+			}
+		}
+	}
+
+	// 将所有镜头合并为一段连贯描述，每个镜头包含场景/运镜/道具信息
 	descParts := make([]string, 0, len(segShots)*2)
 	for _, sh := range segShots {
-		// 旁白直接作为叙述
+		// 镜头头部信息：场景、运镜、道具
+		var shotHeaders []string
+		if sh.Scene != "" {
+			shotHeaders = append(shotHeaders, "场景:"+sh.Scene)
+		}
+		if sh.CameraMovement != "" {
+			shotHeaders = append(shotHeaders, "运镜:"+sh.CameraMovement)
+		}
+		if len(sh.Props) > 0 {
+			shotHeaders = append(shotHeaders, "道具:"+strings.Join(sh.Props, "、"))
+		}
+		shotParts := make([]string, 0, 4)
+		if len(shotHeaders) > 0 {
+			shotParts = append(shotParts, "["+strings.Join(shotHeaders, "][")+"]")
+		}
+		// 旁白
 		if sh.Narration != "" {
-			descParts = append(descParts, sh.Narration)
+			shotParts = append(shotParts, sh.Narration)
 		}
 		// 事件描述
 		if sh.Event != "" {
-			descParts = append(descParts, sh.Event)
+			shotParts = append(shotParts, sh.Event)
 		}
 		// 对话以自然方式嵌入
 		if sh.Dialogue != "" {
-			descParts = append(descParts, sh.Dialogue)
+			shotParts = append(shotParts, sh.Dialogue)
+		}
+		if len(shotParts) > 0 {
+			descParts = append(descParts, strings.Join(shotParts, "。"))
 		}
 	}
-	fullDesc := strings.Join(descParts, "，")
+	// 前面拼接角色外观和场景描述，确保风格一致
+	var prefixParts []string
+	if len(appearanceDescs) > 0 {
+		prefixParts = append(prefixParts, "【角色形象】"+strings.Join(appearanceDescs, "；"))
+	}
+	if len(sceneDescs) > 0 {
+		prefixParts = append(prefixParts, "【场景】"+strings.Join(sceneDescs, "；"))
+	}
+	var fullDesc string
+	if len(prefixParts) > 0 {
+		fullDesc = strings.Join(prefixParts, "。") + "。" + strings.Join(descParts, "，")
+	} else {
+		fullDesc = strings.Join(descParts, "，")
+	}
 
 	return &model.SegmentOutput{
 		Index: segIdx,
@@ -1460,8 +1517,10 @@ func (s *dramaService) submitVideoTask(ctx context.Context, d *entity.Drama, ep
 		}
 	}
 
+	// 固定 seed 确保各段画风/人物形象一致（基于短剧ID+剧集序号）
+	seed := int(d.Id)*1000 + ep.Index*10 + segIdx
 	taskId, requestJSON, err := createVideoTask(ctx, modelCfg.VideoApiKey, modelCfg.VideoBaseUrl, modelCfg.VideoModelName,
-		prompt, negativePrompt, refURLs, effectiveDur, d.Resolution, d.AspectRatio, modelCfg.VideoSchema)
+		prompt, negativePrompt, refURLs, effectiveDur, d.Resolution, d.AspectRatio, modelCfg.VideoSchema, seed)
 	if err != nil {
 		return "", "", fmt.Errorf("视频合成请求失败: %w", err)
 	}
@@ -1478,7 +1537,7 @@ func (s *dramaService) submitVideoTask(ctx context.Context, d *entity.Drama, ep
 //
 //	params: 额外参数（如 audio/shot_type/watermark 等）
 //	sizes:  分辨率→尺寸字符串映射表
-func createVideoTask(ctx context.Context, apiKey, baseURL, modelName, prompt, negativePrompt string, refURLs []string, duration int, resolution, aspectRatio, videoSchema string) (string, string, error) {
+func createVideoTask(ctx context.Context, apiKey, baseURL, modelName, prompt, negativePrompt string, refURLs []string, duration int, resolution, aspectRatio, videoSchema string, seed int) (string, string, error) {
 	body := map[string]any{
 		"model": modelName,
 		"input": map[string]any{
@@ -1538,6 +1597,9 @@ func createVideoTask(ctx context.Context, apiKey, baseURL, modelName, prompt, ne
 	if duration > 0 {
 		params["duration"] = duration
 	}
+	if seed > 0 {
+		params["seed"] = seed
+	}
 	body["parameters"] = params
 
 	payload, _ := json.Marshal(body)
diff --git a/shortdrama/service/episode_service.go b/shortdrama/service/episode_service.go
index bed0e19..77c76bd 100644
--- a/shortdrama/service/episode_service.go
+++ b/shortdrama/service/episode_service.go
@@ -469,14 +469,11 @@ func createPendingTasks(ctx context.Context, dramaId, epId int64, script string,
 			totalDur int
 		}
 		var cur _shotGroup
-		var prevScene string
 		for _, sh := range allShots {
 			shDur := parseMMSSToSeconds(sh.EndTime) - parseMMSSToSeconds(sh.StartTime)
 			if shDur <= 0 {
 				shDur = 1
 			}
-			// 场景切换 → 必须分段（不同场景不能合并到同一段）
-			sceneChanged := len(cur.shots) > 0 && sh.Scene != prevScene
 			// 计算加入当前镜头后的候选组总时长和提示词长度
 			candShots := make([]domain.Shot, len(cur.shots)+1)
 			copy(candShots, cur.shots)
@@ -487,7 +484,7 @@ func createPendingTasks(ctx context.Context, dramaId, epId int64, script string,
 			exceedDur := len(cur.shots) > 0 && candDur > effectiveMax
 			exceedChars := promptMaxChars > 0 && len([]rune(candPrompt)) > promptMaxChars
 
-			if exceedDur || exceedChars || sceneChanged {
+			if exceedDur || exceedChars {
 				// 当前组已满，保存并开启新组
 				taskGroups = append(taskGroups, _taskGroup{
 					promptText: buildFinalPrompt(domain.ShotsToText(cur.shots)),
@@ -498,7 +495,6 @@ func createPendingTasks(ctx context.Context, dramaId, epId int64, script string,
 				cur.shots = candShots
 				cur.totalDur = candDur
 			}
-			prevScene = sh.Scene
 		}
 		if len(cur.shots) > 0 {
 			taskGroups = append(taskGroups, _taskGroup{