feat: enforce english prompt language strategy

2026-05-18 20:07:11 +08:00
parent adf8b2ba0a
commit bc0b010def
4 changed files with 560 additions and 205 deletions
--- a/api/main.py
+++ b/api/main.py
@@ -91,7 +91,7 @@ YTDLP_COOKIES_FILE = os.getenv("YTDLP_COOKIES_FILE", "").strip()
 YTDLP_COOKIES_FROM_BROWSER = os.getenv("YTDLP_COOKIES_FROM_BROWSER", "").strip()
 AUDIO_PRODUCT_BRIEF = os.getenv(
    "AUDIO_PRODUCT_BRIEF",
-    "SKG 智能按摩产品，主打日常肩颈、腰背、眼部、膝盖或足部放松；广告表达要高级、干净、可信，不做医疗疗效承诺。",
+    "SKG smart massage products for everyday neck-and-shoulder, back, eye, knee, or foot relaxation. Ads should feel premium, clean, trustworthy, and must not make medical efficacy claims.",
 ).strip()
 AUDIO_REWRITE_MODEL = gpt_model_env("AUDIO_REWRITE_MODEL", REWRITE_MODEL)
 VOICE_PROVIDER = "azure_openai"
@@ -454,6 +454,7 @@ class CharacterLibraryItem(BaseModel):
    folder: str = ""
    description: str = ""
    prompt_brief: str = ""
+    prompt_brief_zh: str = ""
    primary_image: str = ""
    images: list[CharacterLibraryImage] = Field(default_factory=list)

@@ -480,6 +481,7 @@ class SubjectTemplateItem(BaseModel):
    description: str = ""
    note: str = ""
    prompt_brief: str = ""
+    prompt_brief_zh: str = ""
    source: Literal["database"] = "database"
    source_job_id: str = ""
    source_frame_idx: int = -1
@@ -534,6 +536,7 @@ class KeyElement(BaseModel):
    subject_kind: SubjectKind = "object"
    subject_assets: list[SubjectAsset] = Field(default_factory=list)
    subject_consensus_brief: str = ""
+    subject_consensus_brief_zh: str = ""
    created_at: float = 0.0


@@ -565,6 +568,7 @@ class AudioScript(BaseModel):
    source_text: str = ""
    source_zh: str = ""
    rewritten_text: str = ""
+    rewritten_text_zh: str = ""
    speaker_profile: str = ""
    rhythm_profile: str = ""
    background_audio_profile: str = ""
@@ -2307,7 +2311,7 @@ def _audio_profile_model_sync(wav: Path, segments: list[TranscriptSegment], targ
    fallback = _fallback_audio_profile(segments, target_seconds)
    if not LLM_API_KEY or not wav.exists():
        return fallback
-    transcript = _transcript_join(segments, "en") or _transcript_join(segments, "zh") or "No reliable transcript."
+    transcript = _ensure_english(_transcript_join(segments, "en") or _transcript_join(segments, "zh") or "No reliable transcript.")
    try:
        audio_b64 = base64.b64encode(wav.read_bytes()).decode("ascii")
    except Exception:
@@ -2373,12 +2377,15 @@ def _build_audio_intake_sync(job_id: str, wav: Path, segments: list[TranscriptSe
    )


-def _rewrite_audio_script_sync(segments: list[TranscriptSegment], target_seconds: float = 12.0) -> tuple[str, str]:
+def _rewrite_audio_script_sync(segments: list[TranscriptSegment], target_seconds: float = 12.0) -> tuple[str, str, str]:
    fallback = _fallback_audio_script(segments, target_seconds)
+    try:
+        fallback_zh = _translate_text_sync(fallback, "zh", max_tokens=300) if LLM_API_KEY else ""
+    except Exception:
+        fallback_zh = ""
    if not LLM_API_KEY:
-        return fallback, "LLM_API_KEY 未配置，使用本地 SKG 模板"
+        return fallback, fallback_zh, "LLM_API_KEY 未配置，使用本地 SKG 模板"
    source_text = _transcript_join(segments, "en")
-    source_zh = _transcript_join(segments, "zh")
    min_words, max_words = _voiceover_target_words(target_seconds)
    prompt = (
        "You are an English short-video voice-over writer for SKG wellness massagers. "
@@ -2392,10 +2399,9 @@ def _rewrite_audio_script_sync(segments: list[TranscriptSegment], target_seconds
        "5. Introduce SKG products directly: smart massage, warmth, rhythm, daily neck/back/eye/knee/foot relaxation.\n"
        "6. Keep it easy for TTS: short sentences, spoken phrasing, no hashtags, no stage directions, no quotation marks.\n"
        "7. If the source transcript is thin, ignore it and write a general SKG product intro.\n"
-        'Return strict JSON only: {"rewritten_text":"..."}.\n\n'
-        f"SKG product context: {AUDIO_PRODUCT_BRIEF}\n\n"
-        f"English transcript:\n{source_text or 'None'}\n\n"
-        f"Chinese translation for reference:\n{source_zh or 'None'}"
+        'Return strict JSON only: {"rewritten_text":"English VO","rewritten_text_zh":"Simplified Chinese mirror for team review"}.\n\n'
+        f"SKG product context: {_ensure_english(AUDIO_PRODUCT_BRIEF)}\n\n"
+        f"English transcript:\n{source_text or 'None'}"
    )
    try:
        resp = llm().chat.completions.create(
@@ -2415,9 +2421,12 @@ def _rewrite_audio_script_sync(segments: list[TranscriptSegment], target_seconds
            raw = match.group(0) if match else raw
        data = json.loads(raw)
        text = str(data.get("rewritten_text", "")).strip()
-        return (text or fallback), ""
+        text_zh = str(data.get("rewritten_text_zh", "")).strip()
+        if text and not text_zh:
+            text_zh = _translate_text_sync(text, "zh", max_tokens=300)
+        return (text or fallback), (text_zh or fallback_zh), ""
    except Exception as e:
-        return fallback, f"改写失败，使用本地模板：{e}"
+        return fallback, fallback_zh, f"改写失败，使用本地模板：{e}"


 def _choose_azure_voice_id() -> str:
@@ -2521,7 +2530,7 @@ def _build_audio_script_sync(job_id: str, segments: list[TranscriptSegment], tar
    source_text = _transcript_join(segments, "en")
    source_zh = _transcript_join(segments, "zh")
    duration = max(float(target_seconds or 0), _segment_duration(segments), 4.0)
-    rewritten, rewrite_error = _rewrite_audio_script_sync(segments, duration)
+    rewritten, rewritten_zh, rewrite_error = _rewrite_audio_script_sync(segments, duration)
    selected_voice_id = _choose_tts_voice_id()
    speaker_profile, rhythm_profile = _audio_delivery_profile(segments, duration, selected_voice_id)
    voice_url = ""
@@ -2539,6 +2548,7 @@ def _build_audio_script_sync(job_id: str, segments: list[TranscriptSegment], tar
        source_text=source_text,
        source_zh=source_zh,
        rewritten_text=rewritten,
+        rewritten_text_zh=rewritten_zh,
        speaker_profile=speaker_profile,
        rhythm_profile=rhythm_profile,
        product_brief=AUDIO_PRODUCT_BRIEF,
@@ -3055,6 +3065,55 @@ class RewriteStoryboardScriptReq(BaseModel):
    segments: list[ScriptRewriteSegmentReq] = Field(default_factory=list)


+_TRANSLATION_CACHE: dict[str, str] = {}
+
+
+def _contains_cjk(text: str) -> bool:
+    return bool(re.search(r"[\u3400-\u9fff]", text or ""))
+
+
+def _translate_text_sync(text: str, target: Literal["en", "zh"] = "en", *, max_tokens: int = 700) -> str:
+    text = (text or "").strip()
+    if not text or not LLM_API_KEY:
+        return text
+    target_label = "English" if target == "en" else "Simplified Chinese"
+    prompt = (
+        f"Translate the following TikTok ad planning text into concise natural {target_label}. "
+        "Preserve concrete product, camera, subject, timing, and structure details. "
+        "Do not add commentary, markdown, quotes, or explanations.\n\n"
+        f"Input:\n{text}"
+    )
+    resp = llm().chat.completions.create(
+        model=TRANSLATE_MODEL,
+        messages=[{"role": "user", "content": prompt}],
+        temperature=0.15,
+        max_tokens=max_tokens,
+    )
+    out = (resp.choices[0].message.content or "").strip()
+    if not out:
+        rc = getattr(resp.choices[0].message, "reasoning_content", "") or ""
+        if rc:
+            out = rc.strip().splitlines()[-1].strip()
+    return re.sub(r'^[\'"「『]+|[\'"」』]+$', "", out).strip() or text
+
+
+def _ensure_english(text: str) -> str:
+    text = (text or "").strip()
+    if not text or not _contains_cjk(text):
+        return text
+    key = hashlib.sha256(("en\0" + text).encode("utf-8")).hexdigest()
+    cached = _TRANSLATION_CACHE.get(key)
+    if cached:
+        return cached
+    try:
+        translated = _translate_text_sync(text, "en", max_tokens=max(700, min(3500, len(text) // 2 + 900)))
+        _TRANSLATION_CACHE[key] = translated
+        return translated
+    except Exception as e:
+        print(f"[ensure english fallback] {e}", flush=True)
+        return text
+
+
@app.post("/translate")
 def translate_text(req: TranslateReq) -> dict:
    """单条文本翻译（给生图自定义提取元素 zh→en 用）"""
@@ -3092,22 +3151,26 @@ def translate_text(req: TranslateReq) -> dict:

 def _fallback_script_rewrite_item(segment: ScriptRewriteSegmentReq, author_intent: str = "") -> dict:
    source = (segment.source or "").strip()
-    intent = (author_intent or "").strip()
+    intent = _ensure_english(author_intent or "")
    role = segment.role or ""
    templates = {
-        "开场钩子": "你有没有发现，低头久了以后，脖子和肩膀会先替你喊累。",
-        "痛点推进": "刷手机、坐电脑、赶通勤叠在一起，肩颈很容易一直绷着放不下来。",
-        "利益证明": "SKG 这种挂脖按摩仪，重点就是贴住肩颈位置，把热敷感和揉按感带到真正紧的地方。",
-        "方案过渡": "这一段可以直接拍拿起、戴上、贴合，让产品自然进入日常放松场景。",
-        "转化收口": "如果你也想把肩颈放松变成每天的小习惯，可以从这台 SKG 开始。",
-        "节奏承接": "顺着原片节奏，把这一句落到一个具体的肩颈使用场景里。",
+        "hook": "Have you noticed that after hours of looking down, your neck and shoulders complain before you do?",
+        "pain": "Phone scrolling, desk work, and commuting can keep your neck and shoulders tight all day.",
+        "proof": "An SKG wearable massager sits around the neck and shoulders, bringing warm, rhythmic comfort to the spots that feel tense.",
+        "solution": "This beat can simply show pick up, wear, fit, and relax, so the product enters a normal daily routine.",
+        "cta": "If you want neck-and-shoulder relaxation to become a daily habit, start with this SKG massager.",
+        "bridge": "Follow the source rhythm, but land this line in one specific neck-and-shoulder use moment.",
    }
-    rewritten = templates.get(role, templates["节奏承接"])
-    if source and role not in {"开场钩子", "转化收口"}:
-        rewritten = f"{rewritten} 原片这一句的节奏可以保留，但内容换成 SKG 的佩戴和放松体验。"
+    rewritten = templates.get(role, templates["bridge"])
+    if source and role not in {"hook", "cta"}:
+        rewritten = f"{rewritten} Keep the source sentence rhythm, but replace the content with SKG wearing and relaxation experience."
    if intent:
-        rewritten = f"{rewritten} 语气按作者想法处理：{intent[:44]}。"
-    return {"index": segment.index, "text": rewritten[:220]}
+        rewritten = f"{rewritten} Adjust the tone based on the creator note: {intent[:90]}."
+    try:
+        zh = _translate_text_sync(rewritten, "zh", max_tokens=260) if LLM_API_KEY else ""
+    except Exception:
+        zh = ""
+    return {"index": segment.index, "text": rewritten[:260], "text_zh": zh}


 def _parse_script_rewrite_items(raw: str, requested: list[ScriptRewriteSegmentReq], author_intent: str = "") -> list[dict]:
@@ -3123,7 +3186,7 @@ def _parse_script_rewrite_items(raw: str, requested: list[ScriptRewriteSegmentRe
    raw_items = data.get("items") if isinstance(data, dict) else data
    if not isinstance(raw_items, list):
        raw_items = []
-    by_index: dict[int, str] = {}
+    by_index: dict[int, tuple[str, str]] = {}
    for item in raw_items:
        if not isinstance(item, dict):
            continue
@@ -3132,19 +3195,27 @@ def _parse_script_rewrite_items(raw: str, requested: list[ScriptRewriteSegmentRe
        except Exception:
            continue
        value = str(item.get("text") or item.get("rewritten_text") or "").strip()
+        value_zh = str(item.get("text_zh") or item.get("rewritten_text_zh") or "").strip()
        if value:
-            by_index[idx] = re.sub(r"\s+", " ", value).strip()[:260]
-    return [
-        {"index": segment.index, "text": by_index.get(segment.index) or _fallback_script_rewrite_item(segment, author_intent)["text"]}
-        for segment in requested
-    ]
+            by_index[idx] = (re.sub(r"\s+", " ", value).strip()[:260], re.sub(r"\s+", " ", value_zh).strip()[:260])
+    items = []
+    for segment in requested:
+        fallback = _fallback_script_rewrite_item(segment, author_intent)
+        text, text_zh = by_index.get(segment.index, ("", ""))
+        if text and not text_zh:
+            try:
+                text_zh = _translate_text_sync(text, "zh", max_tokens=260) if LLM_API_KEY else ""
+            except Exception:
+                text_zh = ""
+        items.append({"index": segment.index, "text": text or fallback["text"], "text_zh": text_zh or fallback.get("text_zh", "")})
+    return items


 def _rewrite_storyboard_script_sync(req: RewriteStoryboardScriptReq) -> list[dict]:
    segments = [segment for segment in req.segments if (segment.source or segment.current_text).strip()]
    if not segments:
        return []
-    author_intent = (req.author_intent or "").strip()
+    author_intent = _ensure_english(req.author_intent or "")
    if not LLM_API_KEY:
        return [_fallback_script_rewrite_item(segment, author_intent) for segment in segments]
    payload = [
@@ -3152,26 +3223,27 @@ def _rewrite_storyboard_script_sync(req: RewriteStoryboardScriptReq) -> list[dic
            "index": segment.index,
            "time": f"{segment.start:.1f}-{segment.end:.1f}s",
            "role": segment.role,
-            "source_reference": segment.source,
-            "current_voiceover": segment.current_text,
+            "source_reference": _ensure_english(segment.source),
+            "current_voiceover": _ensure_english(segment.current_text),
        }
        for segment in segments
    ]
    prompt = (
-        "你是信息流广告脚本文案改写师。任务：基于原参考文案的节奏和信息结构，把每段改写成 SKG 挂脖肩颈按摩仪的新口播文案。\n"
-        "硬规则：\n"
-        "1. 输出中文短视频口播，不要英文，不要舞台说明，不要引号。\n"
-        "2. 不逐字翻译原文，不保留原品牌、价格、优惠码、平台话术；只参考节奏、钩子、痛点、转化结构。\n"
-        "3. 产品固定为套在脖子上的 U 形肩颈按摩仪，表达肩颈紧绷、久坐低头、热敷感、揉按感、佩戴放松和日常使用场景。\n"
-        "4. 避免医疗疗效、治疗、治愈、止痛等强功效承诺。\n"
-        "5. 每段尽量短，适配该段时间；保持自然创作者口吻。\n"
-        "6. mode=all 时，整片要前后连贯；mode=segment 时，只改给定段落但仍要贴合上下文风格。\n"
-        f"作者想法：{author_intent or '没有额外想法，按原片节奏改成自然卖点口播。'}\n"
-        f"改写模式：{req.mode}\n"
-        f"SKG 产品背景：{AUDIO_PRODUCT_BRIEF}\n\n"
-        "输入段落 JSON：\n"
+        "You are an information-feed ad voice-over rewrite specialist. Rewrite each segment into a new ENGLISH SKG neck-and-shoulder massager voice-over line while preserving the source rhythm and information structure.\n"
+        "Hard rules:\n"
+        "1. The main text field must be English short-video VO. No stage directions, no quotes.\n"
+        "2. Do not translate word-for-word. Do not keep the original brand, price, discount code, platform CTA, or exact claims; only reuse rhythm, hook, pain-point, proof, and conversion structure.\n"
+        "3. The product is a U-shaped neck-and-shoulder wearable massager worn around the neck. Express neck/shoulder tension, desk posture, looking down, warmth, kneading-like comfort, wearing, relaxation, and daily use.\n"
+        "4. Avoid medical treatment, cure, pain elimination, clinical, or disease claims.\n"
+        "5. Keep each segment short enough for its time range and natural for a creator voice.\n"
+        "6. If mode=all, make the whole piece coherent; if mode=segment, rewrite only the given segment while matching the broader style.\n"
+        "7. Also return a Simplified Chinese mirror for team review in text_zh; it is not for model prompts.\n"
+        f"Creator note: {author_intent or 'No extra note; follow the source pacing and turn it into natural SKG product VO.'}\n"
+        f"Rewrite mode: {req.mode}\n"
+        f"SKG product context: {_ensure_english(AUDIO_PRODUCT_BRIEF)}\n\n"
+        "Input segments JSON:\n"
        + json.dumps(payload, ensure_ascii=False)
-        + '\n\n只输出严格 JSON：{"items":[{"index":0,"text":"改写后的中文口播"}]}'
+        + '\n\nReturn strict JSON only: {"items":[{"index":0,"text":"rewritten English VO","text_zh":"中文镜像"}]}'
    )
    models = []
    for model in [AUDIO_REWRITE_MODEL, ASR_FALLBACK_MODEL, TRANSLATE_MODEL]:
@@ -3182,7 +3254,7 @@ def _rewrite_storyboard_script_sync(req: RewriteStoryboardScriptReq) -> list[dic
            resp = llm().chat.completions.create(
                model=model,
                messages=[
-                    {"role": "system", "content": "只返回合法 JSON，不要 markdown，不要解释。"},
+                    {"role": "system", "content": "Return valid JSON only. No markdown. No explanation."},
                    {"role": "user", "content": prompt},
                ],
                response_format={"type": "json_object"},
@@ -3950,6 +4022,7 @@ class UpdateElementReq(BaseModel):
    name_en: str | None = None
    position: str | None = None
    subject_consensus_brief: str | None = None
+    subject_consensus_brief_zh: str | None = None


 class GenerateSceneAssetReq(BaseModel):
@@ -3998,8 +4071,8 @@ class GenerateSubjectAssetsReq(BaseModel):
 def _subject_profile_prompt_clause(profile: SubjectProfilePreference | None) -> str:
    if not profile:
        return ""
-    prompt_summary = (profile.prompt_summary or "").strip()
-    resolved_summary = (profile.resolved_summary or "").strip()
+    prompt_summary = _ensure_english(profile.prompt_summary or "")
+    resolved_summary = _ensure_english(profile.resolved_summary or "")
    if prompt_summary:
        body = prompt_summary[:1400]
    else:
@@ -4013,7 +4086,7 @@ def _subject_profile_prompt_clause(profile: SubjectProfilePreference | None) ->
            ("hair style", profile.hair),
            ("commercial mood", profile.mood),
        ]
-        body = "; ".join(f"{name}: {value.strip()}" for name, value in parts if value and value.strip())[:1400]
+        body = "; ".join(f"{name}: {_ensure_english(value.strip())}" for name, value in parts if value and value.strip())[:1400]
    if not body and not resolved_summary:
        return ""
    mode = "random-composed" if profile.mode == "random" else "manually selected"
@@ -4125,7 +4198,9 @@ def update_element(job_id: str, idx: int, element_id: str, req: UpdateElementReq
                    if req.position is not None:
                        e.position = req.position.strip()
                    if req.subject_consensus_brief is not None:
-                        e.subject_consensus_brief = req.subject_consensus_brief.strip()[:2200]
+                        e.subject_consensus_brief = _ensure_english(req.subject_consensus_brief.strip())[:2200]
+                    if req.subject_consensus_brief_zh is not None:
+                        e.subject_consensus_brief_zh = req.subject_consensus_brief_zh.strip()[:2200]
        new_frames.append(f)
    if not found:
        raise HTTPException(404, "element not found")
@@ -4208,7 +4283,7 @@ def generate_scene_asset(job_id: str, idx: int, req: GenerateSceneAssetReq) -> J
        if confirmed_subjects
        else "Remove the main foreground subject from the frame if present. "
    )
-    subject_brief = req.subject_brief.strip()
+    subject_brief = _ensure_english(req.subject_brief.strip())
    subject_brief_clause = (
        f"Subject identity (text only, no image reference): {subject_brief[:1800]}. "
        "Maintain this identity across this and other endpoint frames in the same storyboard. "
@@ -4237,7 +4312,7 @@ def generate_scene_asset(job_id: str, idx: int, req: GenerateSceneAssetReq) -> J
        "warm_lifestyle": "Use a warm lifestyle style: realistic lived-in details, soft natural light, approachable atmosphere.",
        "cinematic": "Use a cinematic style: dramatic but natural lighting, richer depth, filmic contrast, not fantasy.",
    }[req.scene_style]
-    user_prompt = req.prompt.strip()
+    user_prompt = _ensure_english(req.prompt.strip())
    user_prompt_clause = (
        "User scene direction: " + user_prompt[:1200] + " "
        if user_prompt
@@ -4483,6 +4558,7 @@ def generate_subject_assets(job_id: str, idx: int, element_id: str, req: Generat
        brief = template.prompt_brief.strip() or template.note.strip() or template.description.strip()
        if similar_mode and not brief:
            brief = _describe_subject_template_from_images(template.name, template.subject_style, template_paths, template.note)
+        brief = _ensure_english(brief)
        selected_template_brief = brief.strip()
        template_brief_clause = (
            f"Reference character brief from saved database template '{template.name}': {brief}. "
@@ -4496,6 +4572,7 @@ def generate_subject_assets(job_id: str, idx: int, element_id: str, req: Generat
        character_label = character.name
        character_reference_paths.extend(character_library_file(image.filename) for image in character.images[:7])
        brief = character.prompt_brief.strip() or character.description.strip()
+        brief = _ensure_english(brief)
        selected_template_brief = brief.strip()
        template_brief_clause = (
            f"Reference character brief from built-in creative character '{character.name}': {brief}. "
@@ -4558,7 +4635,7 @@ def generate_subject_assets(job_id: str, idx: int, element_id: str, req: Generat
        if req.reconstruction_mode == "similar"
        else "Preserve identity, proportions, silhouette, material, colors, styling, and distinctive details across all generated views. "
    )
-    prompt_extra = req.prompt.strip()
+    prompt_extra = _ensure_english(req.prompt.strip())
    prompt_extra_clause = f"User direction: {prompt_extra[:1200]} " if prompt_extra else ""
    subject_profile_clause = _subject_profile_prompt_clause(req.subject_profile)
    identity_lock_clause = (
@@ -4709,7 +4786,7 @@ def generate_subject_assets(job_id: str, idx: int, element_id: str, req: Generat
                            ]
                            fallback_brief = " ".join(part.strip() for part in fallback_parts if part and part.strip())[:1800]
                            if selected_template_brief:
-                                e.subject_consensus_brief = selected_template_brief[:1800]
+                                e.subject_consensus_brief = _ensure_english(selected_template_brief)[:1800]
                            else:
                                asset_paths = [
                                    job_dir(job_id) / "assets" / f"{asset.id}.jpg"
@@ -4722,9 +4799,14 @@ def generate_subject_assets(job_id: str, idx: int, element_id: str, req: Generat
                                    asset_paths,
                                    fallback_brief,
                                )
-                                e.subject_consensus_brief = brief or current_brief or fallback_brief or (
+                                e.subject_consensus_brief = _ensure_english(brief or current_brief or fallback_brief or (
                                    "Generated SKG ad subject; identity brief unavailable. Keep one consistent commercial subject with clear neck and shoulder placement area."
-                                )
+                                ))[:1800]
+                            if e.subject_consensus_brief and not e.subject_consensus_brief_zh:
+                                try:
+                                    e.subject_consensus_brief_zh = _translate_text_sync(e.subject_consensus_brief, "zh", max_tokens=500)[:1800]
+                                except Exception:
+                                    e.subject_consensus_brief_zh = ""
        new_frames.append(f)
    if generation_errors:
        msg = f"主体资产包部分生成完成 · {el.name_zh} · {len(generated)} 张，失败 {len(generation_errors)} 张"
@@ -5296,18 +5378,23 @@ def save_subject_template(job_id: str, req: SaveSubjectTemplateReq) -> SubjectTe
        raise HTTPException(404, "subject asset files missing")

    primary = next((image.id for image in images if image.view == "front"), images[0].id)
-    prompt_brief = _describe_subject_template_from_images(
+    prompt_brief = _ensure_english(_describe_subject_template_from_images(
        name,
        req.subject_style,
        saved_image_paths,
        req.note.strip(),
-    ) or req.note.strip()
+    ) or req.note.strip())
+    try:
+        prompt_brief_zh = _translate_text_sync(prompt_brief, "zh", max_tokens=500) if prompt_brief else ""
+    except Exception:
+        prompt_brief_zh = ""
    item = SubjectTemplateItem(
        id=template_id,
        name=name,
        description=req.note.strip(),
        note=req.note.strip(),
        prompt_brief=prompt_brief,
+        prompt_brief_zh=prompt_brief_zh,
        source_job_id=job_id,
        source_frame_idx=frame.index,
        source_element_id=element.id,