{ "version": "v3", "generated_by": "innovator-digest", "date": "2026-08-27", "based_on": "hunt/knowhow_v3.md + master 三条拍板决策(T-mtb6dgpw)", "supersedes": "test_scenarios_v2.json", "note": "评分对象是 h3-video-prompt agent 针对每条创意描述产出的最终交付物,按 benchmark_v3.md 打分。每条场景新增 expected_outcome 元字段(prompt|block)声明评测轨别(benchmark_v3.md §二 C3);Ref2VA 场景的 asset_menu 里 url 用 占位,真实跑测时须替换为可访问的同源地址。出处等级:A=官方逐字条款、B=内部规范、C=构造性设计(标注在每条判据尾部)。新增 TS-13~TS-18 覆盖上轮从未被触发的盲区。", "scenarios": [ { "id": "TS-01", "mode": "T2VA", "difficulty": "基础", "expected_outcome": "prompt", "raw_brief": "帮我拍个开箱视频,就是纯手办开箱,没有别的素材。", "why_this_mode": "无任何参考素材输入,纯文本原创场景,对应官方 T2VA。与 benchmark_v2 一致,仅按 v3 判据微调落点。", "scoring_points": [ "D1:正确识别为 T2VA,不得臆造参考图锚定写法;[A] 无参考图时直接由文本构建完整时间线,可补与用户意图一致的细节(base-en.txt:172)", "D8:主体/时长/情绪基调缺失是否一次性反问而非臆造具体手办(B:内部规范『优先反问』,fde_report_v2 §四.2 拍板项保留原状)", "D4:开箱近景特写最易堆砌质量词,看材质光影是否落到光源+位置+效果三问", "D9:基础模式三字段齐全顺序正确,无 schema 外字段(B:内部工程规范)" ] }, { "id": "TS-02", "mode": "I2VA", "difficulty": "基础", "expected_outcome": "prompt", "raw_brief": "这是我们奶茶新品的一张海报图,想让它'动起来',杯子转个圈然后倒出奶茶的效果。", "why_this_mode": "以一张首帧图像为起点展开叙事,对应官方 I2VA。[v2继承]", "scoring_points": [ "D1:识别为 I2VA,从参考图静态状态展开动作,不当 Ref2VA 或重新构图", "D9:I2VA 起始指令行含 `` 且编号正确、占位符已解析;[v3] 顺序口径改为 references 数组 = 菜单原始顺序 = 自动编号顺序,不再要求与 Picture 定义条目一一对应(机制性证据 chat_template.json + 内部规范 B)", "D2:integrated_multimodal_description 是否交代'静止→转圈→倒出'时间线", "D6:转圈+倒出是否有唯一主导运镜,不塞矛盾指令" ] }, { "id": "TS-03", "mode": "FL2VA", "difficulty": "中等", "expected_outcome": "prompt", "raw_brief": "有一张女生站在门口的照片和一张她已经走进屋里坐下的照片,想要中间的过渡动画,15秒。", "why_this_mode": "首尾两帧描述连续运动路径,对应官方 FL2VA。[v2继承]", "scoring_points": [ "D1:两张图各自的时间戳显式标注(0.00s / 结束帧),不得只当 I2VA 处理其一", "D1/D6:不得把首尾帧场景当 Ref2VA 写(首尾帧与全参考结构混用仍是 D1 档 2 违规)", "D9:FL2VA 指令行 N / S.SS 已替换成具体数值,无残留字面占位符(内部规范 B)", "D4/D8:中间过程可补低风险可拍摄细节,不得添加剧情结论" ] }, { "id": "TS-04", "mode": "L2VA", "difficulty": "中等", "expected_outcome": "prompt", "raw_brief": "只有一张最后定格的照片——男人站在悬崖边看着日出,我要一段到达这个画面之前发生了什么的视频。", "why_this_mode": "仅有尾帧,反推到达前的过程,对应官方 L2VA。[v2继承]", "scoring_points": [ "D1:识别为 L2VA,产出是反推过程而非复述尾帧画面", "D8:反推限定低风险可拍摄动作,不得添加关系/情绪等高风险剧情推断", "D4:动作三段式具体展开且正好停在尾帧状态", "D6:运镜方向呼应'逐渐抵达'叙事,不与定格机位矛盾" ] }, { "id": "TS-05", "mode": "Ref2VA", "difficulty": "困难", "expected_outcome": "prompt", "asset_menu": [ { "order": 1, "id": "watch", "type": "image", "role": "产品外观来源", "url": "/h3_v3_ts05/watch.png", "note": "" }, { "order": 2, "id": "speaker_video", "type": "video", "role": "人物形象与身份锚定", "url": "/h3_v3_ts05/speaker.mp4", "note": "" } ], "raw_brief": "帮我做条产品发布短视频:15秒,参考素材是一张我们智能手表的产品图和一段之前拍好的代言人口播视频片段,想要代言人抬手展示手表的镜头,手表外观和之前口播视频里的人物形象都要保持一致,风格科技感但不要浮夸。", "why_this_mode": "1 图 + 1 视频的多主体一致性场景,官方 Ref2VA 典型用例;本条是 baseline_target.md worked example 的场景原型。[v2继承]", "scoring_points": [ "D1:subject_definitions 分别声明手表与代言人两个主体,summary/retention_analysis 点名跨镜头保持一致的具体特征", "D7:两条素材分别声明职责;[v3 A] 两条资产的原 url 原样进 conditions[].uri / references 数组,顺序 = 菜单顺序,url 不出现在正文层", "D9:视频素材走 `