From 52c5cdae1d397914cf0ed80e3701811182056b62 Mon Sep 17 00:00:00 2001 From: "Azmat@qq.com" Date: Thu, 27 Aug 2026 18:27:08 +0800 Subject: [PATCH] =?UTF-8?q?=E6=B5=8B=E8=AF=95=E8=A7=86=E9=A2=91=E5=A4=8D?= =?UTF-8?q?=E5=88=BB?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- core/backend/apps/ai/services.py | 24 +-- core/backend/apps/ai/test_video_replace.py | 168 +++++++++++++++++++++ core/backend/apps/ai/video_digest.py | 65 ++++---- core/backend/apps/ai/video_replace.py | 83 ++++++++-- core/frontend/src/pipeline-page.css | 14 -- core/frontend/src/routes/pipeline.tsx | 46 +----- core/frontend/src/routes/video-replace.tsx | 5 +- 7 files changed, 299 insertions(+), 106 deletions(-) diff --git a/core/backend/apps/ai/services.py b/core/backend/apps/ai/services.py index 8b325b7..ee86102 100644 --- a/core/backend/apps/ai/services.py +++ b/core/backend/apps/ai/services.py @@ -901,13 +901,17 @@ NO_EMBEDDED_CAPTIONS_REQUIREMENT = ( "只保留商品包装上参考图中原有的真实物理文字与标识。旁白和环境音可以保留，但不要把声音转成画面文字。" ) -# 故事板图会作为 Seedance 的参考图。导演信息应由页面从结构化脚本渲染， -# 不能烧进参考图，否则边框、时间标签和旁白文字会被视频模型误当成待保留的画面元素。 -STORYBOARD_VIDEO_KEYFRAME_REQUIREMENT = ( - "【视频关键帧硬性规则】这张图将直接作为视频生成参考图：只生成一张干净、全幅、写实的 9:16 关键画面，" - "画面必须完整呈现本段最关键的角色、商品、场景、动作和空间关系。" - "禁止拼图、多宫格、故事板边框、时间轴、机位标注、旁白文字、标题、价格贴片、字幕、UI 或任何叠加文字；" - "只保留商品包装参考图中原有的真实物理文字与标识。" +# 故事板是创作者确认镜头节奏的导演稿，因此必须由 image-2 直接生成完整排版图； +# 视频提示词会明确要求只提取其中的角色、商品、场景、动作和节奏，忽略版式文字。 +STORYBOARD_DIRECTOR_LAYOUT_REQUIREMENT = ( + "【导演故事板成图硬性要求】输出必须是一张完整、正式、可交付的电商短视频导演故事板图片，而不是无说明的拼图或单张海报。" + "整体为竖版 9:16，采用清爽专业的影视分镜排版：顶部 1 行标题区，标出“导演故事板”、本段时长和画幅；" + "主体严格按照本段脚本的秒级分镜，优先整理为 3 个连续时间段（例如 0–5s、5–10s、10–15s；若脚本时长不同则按实际时间划分）。" + "每个时间段必须包含三部分：左侧约 22% 宽的浅色信息栏，清晰写时间、景别、机位/运镜和简短动作；右侧约 78% 的写实主画面；" + "画面底部横条使用扬声器图标与“口播/旁白：”呈现该时间段对应旁白。各段按时间自上而下排列，边框、间距、对齐统一。" + "右侧画面要真实展现该秒级分镜的不同关键动作，人物脸、服装、商品外观、配色和场景连续一致；商品用法必须正确，包装真实文字和 Logo 不得改造。" + "故事板中的中文标注必须工整、简洁、可读，只写脚本已有的时间、机位、动作与旁白，不虚构价格、功效、活动或品牌信息。" + "禁止只给一张静态人物图、禁止无文字的四格拼图、禁止把三段画面重复成同一姿势。" ) @@ -2720,7 +2724,7 @@ def build_storyboard_frame_prompt(project, segment, extra_prompt: str = "") -> s ) cleaned = "\n".join(line for line in rendered.split("\n") if line.strip()) return _apply_storyboard_output_ratio( - f"{cleaned}\n{STORYBOARD_VIDEO_KEYFRAME_REQUIREMENT}", project + f"{cleaned}\n{STORYBOARD_DIRECTOR_LAYOUT_REQUIREMENT}", project ) @@ -2748,6 +2752,8 @@ def build_video_segment_prompt(project, video_segment, scene, refs, user_prompt: default = ( "{设定}{分镜}【脚本】{脚本}\n" "严格按照脚本里的秒级分镜切换景别和动作，商品用法必须真实，不要诡异姿势或错误容器。" + "如果参考图是带时间、机位、旁白标注的导演故事板，只提取其中的角色、商品、场景、动作和时间顺序；" + "不要把边框、标题、时间码、扬声器图标或任何故事板文字生成进视频。" "电商带货短视频，商品露出清晰，节奏有转化感。不要字幕，不要背景音乐，但是要有音效，逼真的音效。" ) rendered = render_prompt( @@ -2933,7 +2939,7 @@ def build_storyboard_frame_prompt_refs(project, segment, refs: list[dict], extra ) cleaned = "\n".join(line for line in rendered.split("\n") if line.strip()) return _apply_storyboard_output_ratio( - f"{cleaned}\n{STORYBOARD_VIDEO_KEYFRAME_REQUIREMENT}", project + f"{cleaned}\n{STORYBOARD_DIRECTOR_LAYOUT_REQUIREMENT}", project ) diff --git a/core/backend/apps/ai/test_video_replace.py b/core/backend/apps/ai/test_video_replace.py index fa13968..ba5ed76 100644 --- a/core/backend/apps/ai/test_video_replace.py +++ b/core/backend/apps/ai/test_video_replace.py @@ -2,6 +2,8 @@ 运行:DB_ENGINE=sqlite python manage.py test apps.ai.test_video_replace --settings=airshelf.settings.test """ +import tempfile +from pathlib import Path from unittest.mock import MagicMock, patch from uuid import uuid4 @@ -170,6 +172,95 @@ class SubmitVideoReplaceTests(TestCase): self.assertNotIn("reference_video", roles) # 参考视频不再发给火山 self.assertNotIn("video_url", [item.get("type") for item in content]) + def _fake_source_file(self): + """_FileFromPath 会 stat 真实文件,给个空的临时 mp4 顶替 TOS 下载结果。""" + tmp = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) + tmp.write(b"\x00") + tmp.close() + self.addCleanup(lambda: Path(tmp.name).unlink(missing_ok=True)) + return tmp.name + + def test_digest_asset_video_runs_end_to_end_on_replace_task(self): + """回归:复刻任务的 model_config 是 Seedance(视频),拆解若走 + execute_routed_text_request 会撞「AITask.model_config 必须保持为主模型」直接炸。 + 这条用例不 mock digest_asset_video 本体,只挡住 TOS/ffmpeg/模型,确保真身跑得通。""" + from apps.ai import video_digest + + task = self._submit() + text_model = ModelConfig.objects.filter(capability="text", status="active").first() + self.assertIsNotNone(text_model) + self.assertNotEqual(text_model.id, task.model_config_id) # 就是要两者不同 + with patch.object(video_digest, "_download_source_to_temp", return_value=self._fake_source_file()), \ + patch.object( + video_digest, "digest_input_from_upload", + return_value=(None, [], 8.0, {"width": 720, "height": 1280, "file_name": "参考.mp4"}), + ), \ + patch.object(video_digest, "resolve_digest_model_config", return_value=text_model), \ + patch("apps.ai.services._collect_extract_text", return_value=(DIGEST_SAMPLE, {})): + run_replace_digest(task) + task.refresh_from_db() + self.assertEqual(task.status, AITask.Status.SUBMITTED) + self.assertIn("【镜头 01】", task.request_payload["prompt"]) + self.assertIn("@目标商品", task.request_payload["prompt"]) + + def test_replace_digest_prompt_matches_standalone(self): + """复刻内的提炼必须和「提炼提示词」页一字不差:同一份 SKILL.md、同一段引导语、 + 同样的 temperature / max_tokens,且不许掺商品信息。两处不一致 → 用户在提炼页 + 看到的分镜稿和复刻实际用的对不上,排查会乱。""" + from apps.ai import video_digest + + task = self._submit() + text_model = ModelConfig.objects.filter(capability="text", status="active").first() + seen = {} + + def _capture(provider, model_config, messages, **kwargs): + seen["messages"] = messages + seen["kwargs"] = kwargs + return (DIGEST_SAMPLE, {}) + + with patch.object(video_digest, "_download_source_to_temp", return_value=self._fake_source_file()), \ + patch.object( + video_digest, "digest_input_from_upload", + return_value=(None, [], 8.0, {"width": 720, "height": 1280, "file_name": "参考.mp4"}), + ), \ + patch.object(video_digest, "resolve_digest_model_config", return_value=text_model), \ + patch("apps.ai.services._collect_extract_text", side_effect=_capture): + run_replace_digest(task) + + standalone = video_digest.build_digest_messages( + [], 8.0, product_hint="", video=None, + aspect_ratio="9:16 竖屏", file_title="参考", + ) + self.assertEqual(seen["messages"], standalone) + self.assertEqual(seen["messages"][0]["content"], video_digest.load_digest_skill()) + self.assertEqual(seen["kwargs"].get("temperature"), 0.4) + self.assertEqual( + seen["kwargs"].get("extra_body"), {"max_tokens": video_digest.DIGEST_MAX_TOKENS} + ) + # 提炼阶段绝不能掺进商品名,否则拆出来的就不是「这条参考视频原本的样子」 + self.assertNotIn("净颜精华", str(seen["messages"])) + + def test_digest_retries_once_before_giving_up(self): + """Gemini 偶尔吐废稿。复刻失败要用户从头重选素材,所以同模型再试一次。""" + from apps.ai import video_digest + + task = self._submit() + text_model = ModelConfig.objects.filter(capability="text", status="active").first() + with patch.object(video_digest, "_download_source_to_temp", return_value=self._fake_source_file()), \ + patch.object( + video_digest, "digest_input_from_upload", + return_value=(None, [], 8.0, {"width": 720, "height": 1280, "file_name": "参考.mp4"}), + ), \ + patch.object(video_digest, "resolve_digest_model_config", return_value=text_model), \ + patch( + "apps.ai.services._collect_extract_text", + side_effect=[("", {}), (DIGEST_SAMPLE, {})], + ) as call: + run_replace_digest(task) + self.assertEqual(call.call_count, 2) # 第一次空文废稿,第二次成功 + task.refresh_from_db() + self.assertEqual(task.status, AITask.Status.SUBMITTED) + def test_digest_failure_fails_task_without_charging(self): from apps.ai.video_digest import VideoDigestError @@ -182,9 +273,86 @@ class SubmitVideoReplaceTests(TestCase): task.refresh_from_db() self.assertEqual(task.status, AITask.Status.FAILED) self.assertIn("不完整", task.error_message or "") + # 模型抖动不能报成「内容未通过生成审核」,那会让用户去白换素材 + self.assertEqual(task.error_code, "processing_failed") self.assertFalse(self.provider.create_video_task.called) self.assertEqual(CreditReservation.objects.filter(task=task).count(), 0) + def test_missing_digest_model_is_rejected_at_submit(self): + """提炼模型没配好要当场 400,不能建了任务、等一轮提炼才告诉用户。""" + + + before = AITask.objects.count() + with patch("apps.ai.video_digest.resolve_digest_model_config", return_value=None): + with self.assertRaisesMessage(ValueError, "视频提炼模型未配置"): + self._submit() + self.assertEqual(AITask.objects.count(), before) + self.assertFalse(self.provider.create_video_task.called) + + + def _make_triview(self): + """商品库的三视图是 standalone Asset:metadata.view=three_view + product_id。""" + tri = _asset(self.team, self.user, name="净颜精华·商品三视图", preview="http://tos/triview.png") + tri.metadata = {"product_id": str(self.product.id), "view": "three_view"} + tri.save(update_fields=["metadata"]) + return tri + + def _run_digest(self, task): + from apps.ai import video_digest + + text_model = ModelConfig.objects.filter(capability="text", status="active").first() + with patch.object(video_digest, "_download_source_to_temp", return_value=self._fake_source_file()), \ + patch.object( + video_digest, "digest_input_from_upload", + return_value=(None, [], 8.0, {"width": 720, "height": 1280, "file_name": "参考.mp4"}), + ), \ + patch.object(video_digest, "resolve_digest_model_config", return_value=text_model), \ + patch("apps.ai.services._collect_extract_text", return_value=(DIGEST_SAMPLE, {})): + run_replace_digest(task) + task.refresh_from_db() + return task + + def test_product_triview_is_sent_with_the_video(self): + """商品库选商品时,三视图要跟着一起发给火山,并在提示词里被点名。""" + tri = self._make_triview() + task = self._submit() + labels = [ref["label"] for ref in task.request_payload["references"]] + self.assertIn("目标商品", labels) + self.assertIn("目标商品三视图", labels) + + task = self._run_digest(task) + self.assertEqual(task.status, AITask.Status.SUBMITTED) + prompt = task.request_payload["prompt"] + self.assertIn("@目标商品三视图", prompt) + self.assertIn("白底多角度图", prompt) + + content = self.provider.create_video_task.call_args.kwargs.get("content_items") or [] + image_items = [item for item in content if item.get("type") == "image_url"] + self.assertEqual(len(image_items), 2) # 商品实拍图 + 三视图 + api_prompt = task.request_payload.get("api_prompt") or "" + self.assertNotIn("@目标商品", api_prompt) # @label 已换成火山认的「图片N」指代 + del tri + + def test_prompt_skips_triview_note_when_product_has_none(self): + """没有三视图就别在提示词里提它,免得模型去找一张不存在的图。""" + task = self._run_digest(self._submit()) + prompt = task.request_payload["prompt"] + self.assertNotIn("三视图", prompt) + self.assertIn("@目标商品", prompt) + + def test_triview_yields_budget_so_refs_never_exceed_nine(self): + """三视图占一张名额:商品图要让位,否则火山那边超 9 张直接拒。""" + from apps.ai.video_replace import MAX_IMAGES + + self._make_triview() + for i in range(MAX_IMAGES + 3): + extra = _asset(self.team, self.user, name=f"图{i}.png", preview=f"http://tos/p{i}.png") + ProductImage.objects.create(product=self.product, asset=extra, sort_order=i + 1) + task = self._submit() + refs = task.request_payload["references"] + self.assertEqual(len(refs), MAX_IMAGES) + self.assertEqual(refs[-1]["label"], "目标商品三视图") + def test_compact_digest_trims_human_only_sections(self): compact = compact_digest_for_video(DIGEST_SAMPLE) self.assertIn("【镜头 02】", compact) diff --git a/core/backend/apps/ai/video_digest.py b/core/backend/apps/ai/video_digest.py index 153c28d..b037835 100644 --- a/core/backend/apps/ai/video_digest.py +++ b/core/backend/apps/ai/video_digest.py @@ -48,6 +48,8 @@ MAX_FRAMES = 36 FRAME_WIDTH = 768 FRAME_QUALITY = 3 DIGEST_MAX_TOKENS = 12288 +# 视频复刻内联拆解的同模型重试次数(Gemini 偶尔吐废稿;复刻失败要用户从头重选素材,值得多试一次) +DIGEST_MAX_ATTEMPTS = 2 _SHOT_MARK = re.compile(r"【(?:镜头\s*\d+|第\s*\d+\s*镜)】") _FFMPEG_TIMEOUT = 60 @@ -728,12 +730,12 @@ def _store_raw_source(*, team, path: str, suffix: str) -> tuple[str, str]: return stored.object_key, storage.public_url(object_key=stored.object_key) -def _download_source_to_temp(object_key: str, suffix: str) -> str: +def _download_source_to_temp(object_key: str, suffix: str, *, bucket: str = "") -> str: from apps.assets.storage import TosStorage ext = suffix if str(suffix).startswith(".") else f".{suffix or 'mp4'}" storage = TosStorage() - body = storage.client.get_object(Bucket=storage.bucket, Key=object_key)["Body"].read() + body = storage.client.get_object(Bucket=bucket or storage.bucket, Key=object_key)["Body"].read() tmp = tempfile.NamedTemporaryFile(suffix=ext.lower(), delete=False) try: tmp.write(body) @@ -1357,7 +1359,7 @@ def digest_asset_video(*, asset, task, model_config_id=None) -> tuple[str, dict] 复刻本来就是一次收费,拆解是它的内部工序。模型调用的审计仍记在传入的复刻 task 上(AIModelAttempt),出问题查得到。 """ - from apps.ai.services import execute_routed_text_request + from apps.ai.services import _collect_extract_text, get_text_provider primary = asset.files.filter(is_primary=True).first() or asset.files.first() if primary is None or not primary.object_key: @@ -1372,7 +1374,7 @@ def digest_asset_video(*, asset, task, model_config_id=None) -> tuple[str, dict] local_path = "" try: - local_path = _download_source_to_temp(primary.object_key, suffix) + local_path = _download_source_to_temp(primary.object_key, suffix, bucket=primary.bucket or "") base_name = (asset.name or "参考视频").rsplit(".", 1)[0] upload = _FileFromPath(local_path, f"{base_name}{suffix}") video, frames, duration, extras = digest_input_from_upload(upload) @@ -1385,6 +1387,11 @@ def digest_asset_video(*, asset, task, model_config_id=None) -> tuple[str, dict] duration, len(frames), ) + # ⚠️ 这里必须和「提炼提示词」页(_digest_video)构造出完全相同的 messages: + # 同一份 SKILL.md 作 system、同一段 user 引导语、同样的 temperature / max_tokens。 + # 绝对不要在这里追加商品信息(product_hint 保持默认空)——两处提炼稿一旦不同, + # 用户在提炼页看到的分镜和复刻实际用的分镜就对不上,排查会乱。 + # test_video_replace.test_replace_digest_prompt_matches_standalone 会守住这一点。 messages = build_digest_messages( frames, duration, @@ -1392,28 +1399,34 @@ def digest_asset_video(*, asset, task, model_config_id=None) -> tuple[str, dict] aspect_ratio=ratio_label(int(extras.get("width") or 0), int(extras.get("height") or 0)), file_title=title_from_filename(str(extras.get("file_name") or "")), ) - routed = execute_routed_text_request( - task=task, - primary_model=model_config, - messages=messages, - streaming=True, - structured_output=False, - business_operation="video_digest", - temperature=0.4, - validate_text=lambda text: validate_digest_text( - text, duration=duration, frame_count=len(frames) or (24 if video else 0) - ), - extra_body={"max_tokens": DIGEST_MAX_TOKENS}, - request_summary={ - "duration_seconds": round(duration, 2), - "frame_count": len(frames), - "input": "native_video" if video is not None else "frames", - "for": "video_replace", - }, - allow_retry=False, - allow_fallback=False, - ) - _text, _response, digest = routed.value + # 这里不能走 execute_routed_text_request:它硬性要求 task.model_config 就是本次主模型, + # 而复刻任务的 model_config 是 Seedance(视频模型),传进去必抛 + # 「AITask.model_config 必须保持为用户选择或系统默认的主模型」。 + # 改为直连脚本/提炼同一条流式通道,并自己做「同模型重试」——绝不 fallback 到别的 + # 文本模型,它们看不了视频,换过去必然废稿。 + provider = get_text_provider(model_config) + expected_frames = len(frames) or (24 if video else 0) + digest = "" + last_error: Exception | None = None + for attempt in range(1, DIGEST_MAX_ATTEMPTS + 1): + try: + text, _payload = _collect_extract_text( + provider, + model_config, + messages, + temperature=0.4, + extra_body={"max_tokens": DIGEST_MAX_TOKENS}, + ) + digest = validate_digest_text(text, duration=duration, frame_count=expected_frames) + break + except Exception as exc: # noqa: BLE001 — 空文/废稿/网络抖动都值得再来一次 + last_error = exc + logger.warning( + "replace digest attempt %s/%s failed for task %s: %s", + attempt, DIGEST_MAX_ATTEMPTS, task.id, exc, + ) + if not digest: + raise VideoDigestError(f"参考视频拆解失败：{last_error}") meta = { "digest_model": model_config.name, "digest_input": "native_video" if video is not None else "frames", diff --git a/core/backend/apps/ai/video_replace.py b/core/backend/apps/ai/video_replace.py index 854da87..df0750f 100644 --- a/core/backend/apps/ai/video_replace.py +++ b/core/backend/apps/ai/video_replace.py @@ -40,6 +40,9 @@ FEATURE = "video_replace" REPLACE_MODES = {"product", "character"} MAX_IMAGES = 9 LEGACY_PROMPT_PREFIX = "[视频复刻]" +# 商品三视图在商品库是 standalone Asset(metadata.view=three_view + product_id), +# 和项目内的 BaseAssetGroup 是两套存储,这里直接取商品库那份。 +TRIVIEW_LABEL = "目标商品三视图" REVIEW_UNAVAILABLE = "素材审核服务暂不可用，请稍后重试" REVIEW_FAILED = "参考素材未通过真人合规审核，请更换视频或图片后重试" REVIEW_SUBMIT_FAILED = "素材提交审核失败，请稍后重试" @@ -57,10 +60,17 @@ PRODUCT_DIGEST_TAIL = ( "1. 严格按上面分镜稿的镜头顺序、时间分配、景别、机位、运镜和剪辑节奏还原全片，不要自行增删镜头。\n" "2. 分镜稿里出现的原商品，全部替换成 @目标商品；商品的外形、颜色、材质、包装和标识必须与参考图完全一致，" "不要改造型、不要改配色、不要凭空补细节。\n" + "{triview}" "3. 人物、场景、光线、色调、动作和表情按分镜稿保持不变，只换商品。\n" "4. 台词/旁白按分镜稿逐字念出；原文提到旧商品名称的地方，改说「{subject}」。\n" "5. 分镜稿里的「字幕」栏只是对原片的记录，不要把这些文字画到画面上。" ) +# 商品库里有三视图时才追加这一条(没有就不提,免得模型去找一张不存在的图)。 +PRODUCT_TRIVIEW_NOTE = ( + " · @目标商品三视图 是这件商品的白底多角度图，用它确认商品的立体结构、比例和各个面的细节；" + "镜头转到任何角度，商品都要和三视图对得上。\n" +) + # 拆解期间的占位提示词。真正的提示词在 worker 拆完后回写。 DIGEST_PENDING_PROMPT = "正在提炼参考视频的分镜稿…" @@ -97,7 +107,7 @@ def compact_digest_for_video(digest_text: str) -> str: return "\n".join(line for line in lines if line.strip()).strip() -def build_product_replace_prompt(digest_text: str, subject_name: str) -> str: +def build_product_replace_prompt(digest_text: str, subject_name: str, *, has_triview: bool = False) -> str: """分镜稿 + 商品替换要求 → 交给 Seedance 的完整提示词。 @目标商品 必须与 references 里第一张商品图的 label 对上,否则 build_content_items @@ -109,7 +119,10 @@ def build_product_replace_prompt(digest_text: str, subject_name: str) -> str: if not body: raise ValueError("参考视频拆解结果为空，请重试") subject = (subject_name or "").strip() or "目标商品" - tail = PRODUCT_DIGEST_TAIL.format(subject=subject) + tail = PRODUCT_DIGEST_TAIL.format( + subject=subject, + triview=PRODUCT_TRIVIEW_NOTE if has_triview else "", + ) return enforce_no_embedded_captions(f"{PRODUCT_DIGEST_HEAD}\n\n{body}\n\n{tail}") @@ -210,7 +223,17 @@ def submit_video_replace(*, team, user, params: dict): } if replace_mode == "product": - # 商品复刻:参考视频只喂给提炼模型,不进火山 references(所以也不用送火山审核)。 + # 商品复刻:参考视频只喂给提炼模型,不进火山 references(所以参考视频本身不再送火山审核)。 + # 两个「一定会失败」的前提在这里就查掉,别让用户白等一轮提炼才看到报错: + from .video_digest import resolve_digest_model_config + + if resolve_digest_model_config() is None: + raise ValueError("视频提炼模型未配置，请联系管理员") + # 商品图仍要过审才能当生成参考。提前送审 → 审核和提炼并行跑,明确不过审的直接 400。 + review_state = _ensure_replace_refs_reviewed(team, image_refs) + if review_state == "failed": + raise ValueError(REVIEW_FAILED) + image_refs = _refresh_replace_refs(team, image_refs) # 先秒回一个 CREATED 任务,拆解这种慢活(Gemini 半分钟起)交给 worker。 extra.update({ "digest_source_asset_id": str(video.id), @@ -313,18 +336,23 @@ def run_replace_digest(task) -> None: task.team, uuid.UUID(str(payload.get("digest_source_asset_id") or "")), "参考视频" ) except (ValueError, TypeError): - _fail_reviewing_task(task, "参考视频已失效，请重新上传") + _fail_reviewing_task(task, "参考视频已失效，请重新上传", error_code="asset_unavailable") return try: digest, meta = digest_asset_video(asset=asset, task=task) - prompt = build_product_replace_prompt(digest, str(payload.get("subject_name") or "")) + has_triview = any( + (ref or {}).get("label") == TRIVIEW_LABEL for ref in (payload.get("references") or []) + ) + prompt = build_product_replace_prompt( + digest, str(payload.get("subject_name") or ""), has_triview=has_triview + ) except (VideoDigestError, ValueError) as exc: - _fail_reviewing_task(task, str(exc)) + _fail_reviewing_task(task, str(exc), error_code="processing_failed") return except Exception as exc: # noqa: BLE001 — 拆解任何异常都要把任务收尾,别留 CREATED 僵尸 logger.exception("video replace digest failed for %s", task.id) - _fail_reviewing_task(task, f"参考视频拆解失败：{exc}") + _fail_reviewing_task(task, f"参考视频拆解失败：{exc}", error_code="processing_failed") return with transaction.atomic(): @@ -354,10 +382,20 @@ def _legacy_replace_mode(prompt: str) -> str: return "character" if prompt.startswith("[视频复刻·角色]") else "product" -def _fail_reviewing_task(task, message: str): +def _fail_reviewing_task(task, message: str, *, error_code: str = ""): + """收尾一个还没提交火山的任务。 + + ``_fail_pending_free_video`` 把错误码写死成 content_rejected —— 那是审核不通过的语义, + 前端会渲染成「内容未通过生成审核」且不可重试。拆解失败/素材丢失不是那回事, + 这里按实际原因改码,否则模型抖一下会被误报成合规问题,用户白白去换素材。 + """ from .free_video import _fail_pending_free_video - return _fail_pending_free_video(task, message) + task = _fail_pending_free_video(task, message) + if error_code and task.status == AITask.Status.FAILED and task.error_code != error_code: + task.error_code = error_code + task.save(update_fields=["error_code", "updated_at"]) + return task def _enqueue_replace_review_poll(task): @@ -701,6 +739,21 @@ def _library_image_ref(asset: Asset, *, team, label: str) -> dict: } +def _product_triview_asset(team, product_id: uuid.UUID): + """商品库里这个商品的三视图(白底多角度单张 16:9)。没有就返回 None,不挡生成。""" + return ( + Asset.objects.filter( + team=team, + is_deleted=False, + purged_at__isnull=True, + metadata__product_id=str(product_id), + metadata__view="three_view", + ) + .order_by("-created_at") + .first() + ) + + def _product_library_refs(team, product_id: uuid.UUID) -> tuple[str, list, str]: product = ( Product.objects.filter(id=product_id, team=team, purged_at__isnull=True, status=Product.Status.ACTIVE) @@ -710,21 +763,27 @@ def _product_library_refs(team, product_id: uuid.UUID) -> tuple[str, list, str]: ) if product is None: raise ValueError("商品不存在或已被删除") + triview = _product_triview_asset(team, product_id) + # 三视图也占一张参考图的名额,商品图要给它让位,否则火山那边超 9 张直接拒。 + image_budget = MAX_IMAGES - 1 if triview is not None else MAX_IMAGES assets = [] - seen = set() + seen = {triview.id} if triview is not None else set() for image in product.images.all(): asset = image.asset if asset is None or asset.id in seen or asset.is_deleted: continue seen.add(asset.id) assets.append(asset) - if len(assets) >= MAX_IMAGES: + if len(assets) >= image_budget: break if not assets and product.cover_asset_id and not product.cover_asset.is_deleted: assets.append(product.cover_asset) - if not assets: + if not assets and triview is None: raise ValueError("这个商品还没有可用图片") refs = [_library_image_ref(asset, team=team, label="目标商品" if index == 0 else f"目标商品{index + 1}") for index, asset in enumerate(assets)] + if triview is not None: + # 放最后:@目标商品 仍指向第一张实拍图,三视图作为「各面长什么样」的补充证据。 + refs.append(_library_image_ref(triview, team=team, label=TRIVIEW_LABEL)) return product.title, refs, "library" diff --git a/core/frontend/src/pipeline-page.css b/core/frontend/src/pipeline-page.css index ec8f4d9..b575d32 100644 --- a/core/frontend/src/pipeline-page.css +++ b/core/frontend/src/pipeline-page.css @@ -1231,20 +1231,6 @@ .sb-scene-thumb .placeholder .sb-frame-rv .rv-label { display: none; } .sb-main-img .sb-main-rv { position: absolute; top: 10px; left: 10px; z-index: 4; } - /* 故事板说明层：从已确认脚本直接排版，不把文字烧进给视频模型的关键帧。 */ - .sb-director-board { margin: 0 0 14px; background: var(--background-lighter); border: 1px solid var(--border-faint); border-radius: var(--r-md); overflow: hidden; } - .sb-director-head { display: flex; align-items: center; justify-content: space-between; gap: 10px; padding: 10px 12px; border-bottom: 1px solid var(--border-faint); background: var(--surface); } - .sb-director-head > div { display: flex; align-items: baseline; gap: 8px; min-width: 0; } - .sb-director-head .mono { font-family: var(--font-mono); font-size: 10px; letter-spacing: .05em; color: var(--black-alpha-48); white-space: nowrap; } - .sb-director-head strong { font-size: 12px; font-weight: 500; color: var(--accent-black); white-space: nowrap; overflow: hidden; text-overflow: ellipsis; } - .sb-director-beat { display: grid; grid-template-columns: 54px minmax(0, 1fr); border-bottom: 1px solid var(--border-faint); } - .sb-director-beat:last-child { border-bottom: 0; } - .sb-director-time { display: flex; align-items: flex-start; justify-content: center; padding: 11px 6px; background: var(--surface); color: var(--heat); font-family: var(--font-mono); font-size: 11px; font-variant-numeric: tabular-nums; } - .sb-director-copy { min-width: 0; padding: 10px 12px; } - .sb-director-direction { color: var(--accent-black); font-size: 12px; line-height: 1.55; } - .sb-director-narration { margin-top: 5px; color: var(--black-alpha-56); font-size: 11.5px; line-height: 1.55; } - .sb-director-empty { padding: 14px 12px; color: var(--black-alpha-48); font-size: 12px; line-height: 1.6; } - .sb-rerun-note { display: flex; align-items: flex-start; gap: 10px; padding: 10px 12px; margin-bottom: 14px; background: rgba(180,83,9,.08); border: 1px solid rgba(180,83,9,.20); border-radius: var(--r-md); color: #7C3A05; line-height: 1.55; } .sb-rerun-note .warn-ic { width: 22px; height: 22px; border-radius: var(--r-sm); background: rgba(180,83,9,.12); color: #B45309; display: grid; place-items: center; flex: 0 0 22px; } .sb-rerun-note .warn-ic svg { width: 14px; height: 14px; } diff --git a/core/frontend/src/routes/pipeline.tsx b/core/frontend/src/routes/pipeline.tsx index c2aefbd..3157432 100644 --- a/core/frontend/src/routes/pipeline.tsx +++ b/core/frontend/src/routes/pipeline.tsx @@ -34,26 +34,6 @@ const KIND_LABEL: Record = { product: "商品", person: "角色" // 脚本来源 → 「来源」brief pill 文案 // 入口收敛为「脚本辅助生成 / 上传脚本」两种。theme 键保留:历史稿的 source 仍可能是旧的「一句话主题」。 const SOURCE_LABEL: Record = { ai: "脚本辅助生成", theme: "脚本辅助生成", manual: "上传脚本", video: "上传视频提炼" }; -type DirectorBeat = { time: string; direction: string; narration: string }; -const DIRECTOR_BEAT_RE = /(?:^|\n)\s*(\d{1,2})\s*[-–—~到至]\s*(\d{1,2})\s*(?:s|秒)?\s*[：:]*\s*([^\n]+)/g; -function buildDirectorBeats(visual: string, narration: string, duration: number): DirectorBeat[] { - const matches = [...(visual || "").matchAll(DIRECTOR_BEAT_RE)]; - const narrationLines = (narration || "").split(/(?<=[。！？!?])/).map((line) => line.trim()).filter(Boolean); - if (matches.length) { - return matches.slice(0, 5).map((match, index) => ({ - time: `${match[1]}–${match[2]}s`, - direction: match[3].trim(), - narration: narrationLines[index] || (index === 0 ? narration.trim() : ""), - })); - } - const count = Math.min(3, Math.max(1, narrationLines.length || 1)); - const step = duration / count; - return Array.from({ length: count }, (_, index) => ({ - time: `${Math.round(index * step)}–${Math.round((index + 1) * step)}s`, - direction: visual.trim() || "按本镜脚本完成关键动作与商品露出", - narration: narrationLines[index] || (index === 0 ? narration.trim() : ""), - })); -} // 旁白配音音色预设(语音合成经典版试用包实测可用,与后端 VOICEOVER_VOICES 对齐) const VO_VOICES = [ { key: "BV700_streaming", label: "灿灿 · 活力女声" }, @@ -2070,7 +2050,7 @@ export function PipelinePage(props: { document.body.style.userSelect = ""; }; }, [gutterDragging]); - const SB_PROMPT_DEFAULT = "统一商品、人物、场景风格，生成干净、可直接指导视频的关键帧；导演分镜说明按脚本自动生成"; + const SB_PROMPT_DEFAULT = "统一商品、人物、场景风格，生成完整、可直接指导视频的导演故事板"; // 整张风格提示词:项目级(原 StoryboardVersion.prompt 的去处),重跑时生效 const sbSavedPrompt = (project.metadata as Record | undefined)?.storyboard_prompt as string | undefined; const [storyboardPrompt, setStoryboardPrompt] = useState(sbSavedPrompt || SB_PROMPT_DEFAULT); @@ -3871,13 +3851,6 @@ export function PipelinePage(props: { const sbViewedIsAdopted = !sbViewedVer || sbViewedVer.is_adopted || sbViewedVer.id === sbActiveShot?.adopted_version; const mainUrl = sbViewedVer ? (sbViewedVer.asset_url || assetUrl(sbViewedVer.asset)) : shotImg(sbActiveShot); const mainAid = sbViewedVer?.asset || sbActiveShot?.adopted_asset || ""; - const activeScriptShot = sbActiveShot ? shots[sbActiveShot.sort_order] : null; - const directorBeats = buildDirectorBeats( - activeScriptShot?.visual_prompt || "", - activeScriptShot?.narration || "", - activeScriptShot?.duration_seconds || SEGMENT_DURATION_MAX, - ); - const directorRange = sceneTimes[sbSelected] || "—"; return (
@@ -3949,22 +3922,7 @@ export function PipelinePage(props: { {sbActiveShot.error_message}
)} -
每个场生成一张干净视频关键帧；下方导演稿按本镜真实脚本展开，视频只使用关键帧，不会带入文字或边框。
-
-
-
DIRECTOR BOARD场 {sbSelected + 1} · {directorRange}
- 9:16 · 关键帧 -
- {directorBeats.length ? directorBeats.map((beat, index) => ( -
-
{beat.time}
-
-
{beat.direction}
- {beat.narration &&
口播/旁白：{beat.narration}
} -
-
- )) :
脚本生成后，这里会自动展示每段时间、机位、动作和旁白。
} -
+
每个场由 image-2 直接生成一张完整导演故事板图片；可单独「重跑本场」只重出这一张,每场各自留历史版本,互不影响。
整张风格提示词(重跑时生效,可编辑)
({ type: "image" as const })) : (tempFiles.length ? tempFiles : tempAssetRefs)