diff --git a/pipeline_llm/inference.py b/pipeline_llm/inference.py
index 1470738..a7f2b7b 100644
--- a/pipeline_llm/inference.py
+++ b/pipeline_llm/inference.py
@@ -92,15 +92,42 @@ def _media_ns_extra():
 
 
 def _normalize_media_aliases(upstream):
-    """同类接口对外契约统一（2026-09-05 用户规则）：上传媒体一律 xxx_file 命名。
+    """同类接口对外契约统一（2026-09-06 用户定夺，对齐 sage/llmage）：
+    上传媒体一律三数组参数 image_files / audio_files / video_files，
+    值为字符串或数组均可（模板 Jinja 动态判断两种形态）。
 
-    旧别名 xxx_url 归一到 xxx_file——同能力模板写法一致（如 i2v 统一
-    image_file 进 → video 出），模板不用为同一类能力写多套参数名。
+    旧别名归一：xxx_file / xxx_url（单值）→ xxx_files（数组）；
+    DashScope 形态的 media 数组（[{"type": "reference_image", "url": ...}]
+    或纯 URL 列表）→ 按 type 分组进三数组（r2v 实测调用方直接传 media）。
     """
     for media in ('image', 'video', 'audio'):
-        fk, uk = media + '_file', media + '_url'
-        if not upstream.get(fk) and upstream.get(uk):
-            upstream[fk] = upstream.get(uk)
+        files_k = media + '_files'
+        if not upstream.get(files_k):
+            for alt in (media + '_file', media + '_url'):
+                v = upstream.get(alt)
+                if v:
+                    upstream[files_k] = [v] if isinstance(v, str) else list(v)
+                    break
+    m = upstream.get('media')
+    if m and not any(upstream.get(x + '_files') for x in ('image', 'video', 'audio')):
+        groups = {'image': [], 'video': [], 'audio': []}
+        items = m if isinstance(m, list) else [m]
+        for it in items:
+            if isinstance(it, str) and it.strip():
+                groups['image'].append(it)
+            elif isinstance(it, dict):
+                url = next((v for v in it.values()
+                            if isinstance(v, str) and v.strip().lower().startswith(
+                                ('http://', 'https://', 'data:', 'asset://', '/'))), '')
+                if not url:
+                    continue
+                t = str(it.get('type') or '').lower()
+                g = 'video' if 'video' in t else ('audio' if 'audio' in t else 'image')
+                groups[g].append(url)
+        for media, lst in groups.items():
+            if lst:
+                upstream[media + '_files'] = lst
+        upstream.pop('media', None)
 
 
 def _texts_len(messages):