Task_13 ABO 样品确认页 —— A/B/C 三样本 group 结构

每件 ABO 商品生成一个 group、三条独立 sample：A=small_arc video，B=static video，C=N 张 raw JPG。三条 sample 共用同一 prompt 内容与完全相同的 target bytes，通过 group_id 关联；metadata 共同核心相同，但 sample/input 字段分别记录。

⚠️ 这是新结构样品，尚未写入 final_v1。每个 group 已按 A/B/C 三个 sample 目录落盘；现存旧 final_v1 仍只有 A/B 两条，迁移需另行批准。

N 视角难度分布（本 5 样品 vs 批量目标）

N = 给模型的视角数（绕圈均匀撒开），是难度旋钮：N 越小盲区越大越难。批量按 N=2:25% / N=3:40% / N=4:20% / N=5:15% 分配（同 Task_13）。本 5 样品覆盖：N=2×1 / N=3×2 / N=4×1 / N=5×1——N=3 占两个（最常见档），2/4/5 各一，正好把四档铺开给你看。

25%N=2

40%N=3

20%N=4

15%N=5

配方参数（本批样品）

项	值	说明
数据源	ABO spins（8,116 干净 72 帧 spin）	真实电动转台，az 0..71 步长 5°，无需渲染
N 视角	2 / 3 / 4 / 5（按 25:40:20:15）	难度旋钮；每 spin 一个 N，metadata 逐条记录
INPUT A	N 视角 × 每视角 ~10° 真弧	相邻 3 真帧，各 hold 3 帧 → 每视角 ~0.9s
INPUT B	N 视角 × 每视角冻结	与 A 视角完全相同，唯一变量=动/不动
INPUT C	构成 B 的 N 张 raw JPG	N=2/3/4/5 时正好 2/3/4/5 张；保留原文件名、azimuth、raw URI 与 hash
TARGET	36 帧 @10fps（360°，10°/帧）	72 真帧抽半；A/B/C 共用同一文件
画幅/编码	640×640 letterbox 白底 · H.264 · yuv420p	原生尺寸不一 → 保长宽比 pad，不拉伸

5 个 group · 每组 A/B/C 三条 sample

浅木书桌（带抽屉） N=2 视角

group_id: abo_020fc400_N2 · samples: 020fc400__A/B/C

N=2 · 最难：只给正/背 2 个视角，侧面抽屉结构全靠推断 · 视角方位角 = [0, 180]°

SAMPLE A · input_video.mp4 · small_arc

SAMPLE B · input_video.mp4 · static

SHARED TARGET · A/B/C bytes 相同

INPUT (C) · 2 张 raw images · 正好构成 Input (B)

prompt.txt（A/B/C 三个文件内容逐字相同）

You are given two reference videos of a single product plus 2 source images from the same 2 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background.

metadata.json · 共同核心相同，sample/input 字段各自不同

metadata.json · sample A · 3107 B

{
  "schema_version": "v2v-metadata/2.0",
  "task_id": "Task_13_novel_view_generation",
  "sample_id": "020fc400__A",
  "pipeline_id": "winnie_task13_abo_pipeline",
  "created_at": "2026-07-15T00:00:00Z",
  "generation_mode": "video_to_video",
  "transformation": {
    "kind": "product_novel_view_synthesis",
    "input_variant": "A",
    "n_views": 2,
    "view_azimuths_deg": [
      0,
      180
    ],
    "per_view_arc_deg": 10.0,
    "target": "orbit_360",
    "target_frames": 36,
    "target_az_step_deg": 10,
    "canvas": {
      "size": [
        640,
        640
      ],
      "fit": "letterbox",
      "background": "white"
    },
    "source_media": "real_turntable_photography",
    "frame_source": "abo_electric_turntable_capture",
    "note": "A/B/C are separate samples linked by group_id and share prompt/target."
  },
  "grouping": {
    "group_id": "abo_020fc400_N2",
    "variants": [
      "A",
      "B",
      "C"
    ],
    "shared_prompt": true,
    "shared_target": true
  },
  "input": {
    "variant": "A",
    "type": "video",
    "path": "input_video.mp4",
    "mode": "small_arc",
    "num_frames": 18,
    "duration_s": 1.8
  },
  "video": {
    "num_frames_input": 18,
    "num_frames_target": 36,
    "fps": 10,
    "resolution": [
      640,
      640
    ],
    "native_resolution": [
      1173,
      884
    ],
    "duration_input_s": 1.8,
    "duration_target_s": 3.6,
    "source_media": "real_turntable_jpg_sequence"
  },
  "prompt": {
    "v2v_instruction": "You are given two reference videos of a single product plus 2 source images from the same 2 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background."
  },
  "domain": {
    "field": "object_centric_ecommerce",
    "kind": "real_product_turntable",
    "attributes": {
      "input_variant": "A",
      "n_views": 2
    },
    "source": {
      "dataset": "ABO",
      "spin_id": "020fc400"
    }
  },
  "source_dataset": {
    "dataset": "ABO (Amazon Berkeley Objects) spins",
    "license": "CC BY-NC 4.0",
    "origin": "Guillaumin et al. (Amazon.com) & Collins/Malik (UC Berkeley)",
    "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/02/020fc400/"
  },
  "integrity": {
    "input_sha256": "8f7bf55cefa74cc6e12a5db017b2b0ee97174000228a196c1ea6c241e8b3d8aa",
    "target_sha256": "94c9dd24f4d7283dc4cd6de4107aadd6358e559e0611b4b0590ffd9b6f5c8f0c"
  },
  "provenance": {
    "engine": "v2v_taxonomy",
    "engine_version": "1.0.0",
    "pipeline_version": "v1.0.0",
    "encode": {
      "codec": "libx264",
      "pix_fmt": "yuv420p",
      "crf": 18,
      "ffmpeg": "8.1.2"
    }
  }
}

metadata.json · sample B · 3104 B

{
  "schema_version": "v2v-metadata/2.0",
  "task_id": "Task_13_novel_view_generation",
  "sample_id": "020fc400__B",
  "pipeline_id": "winnie_task13_abo_pipeline",
  "created_at": "2026-07-15T00:00:00Z",
  "generation_mode": "video_to_video",
  "transformation": {
    "kind": "product_novel_view_synthesis",
    "input_variant": "B",
    "n_views": 2,
    "view_azimuths_deg": [
      0,
      180
    ],
    "per_view_arc_deg": 10.0,
    "target": "orbit_360",
    "target_frames": 36,
    "target_az_step_deg": 10,
    "canvas": {
      "size": [
        640,
        640
      ],
      "fit": "letterbox",
      "background": "white"
    },
    "source_media": "real_turntable_photography",
    "frame_source": "abo_electric_turntable_capture",
    "note": "A/B/C are separate samples linked by group_id and share prompt/target."
  },
  "grouping": {
    "group_id": "abo_020fc400_N2",
    "variants": [
      "A",
      "B",
      "C"
    ],
    "shared_prompt": true,
    "shared_target": true
  },
  "input": {
    "variant": "B",
    "type": "video",
    "path": "input_video.mp4",
    "mode": "static",
    "num_frames": 18,
    "duration_s": 1.8
  },
  "video": {
    "num_frames_input": 18,
    "num_frames_target": 36,
    "fps": 10,
    "resolution": [
      640,
      640
    ],
    "native_resolution": [
      1173,
      884
    ],
    "duration_input_s": 1.8,
    "duration_target_s": 3.6,
    "source_media": "real_turntable_jpg_sequence"
  },
  "prompt": {
    "v2v_instruction": "You are given two reference videos of a single product plus 2 source images from the same 2 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background."
  },
  "domain": {
    "field": "object_centric_ecommerce",
    "kind": "real_product_turntable",
    "attributes": {
      "input_variant": "B",
      "n_views": 2
    },
    "source": {
      "dataset": "ABO",
      "spin_id": "020fc400"
    }
  },
  "source_dataset": {
    "dataset": "ABO (Amazon Berkeley Objects) spins",
    "license": "CC BY-NC 4.0",
    "origin": "Guillaumin et al. (Amazon.com) & Collins/Malik (UC Berkeley)",
    "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/02/020fc400/"
  },
  "integrity": {
    "input_sha256": "f1a7dc25e271b19b2149c0f3ba937e7846b1bdda5b3db08ba4d5a217e54f9ea2",
    "target_sha256": "94c9dd24f4d7283dc4cd6de4107aadd6358e559e0611b4b0590ffd9b6f5c8f0c"
  },
  "provenance": {
    "engine": "v2v_taxonomy",
    "engine_version": "1.0.0",
    "pipeline_version": "v1.0.0",
    "encode": {
      "codec": "libx264",
      "pix_fmt": "yuv420p",
      "crf": 18,
      "ffmpeg": "8.1.2"
    }
  }
}

metadata.json · sample C · 3806 B

{
  "schema_version": "v2v-metadata/2.0",
  "task_id": "Task_13_novel_view_generation",
  "sample_id": "020fc400__C",
  "pipeline_id": "winnie_task13_abo_pipeline",
  "created_at": "2026-07-15T00:00:00Z",
  "generation_mode": "image_set_to_video",
  "transformation": {
    "kind": "product_novel_view_synthesis",
    "input_variant": "C",
    "n_views": 2,
    "view_azimuths_deg": [
      0,
      180
    ],
    "per_view_arc_deg": 10.0,
    "target": "orbit_360",
    "target_frames": 36,
    "target_az_step_deg": 10,
    "canvas": {
      "size": [
        640,
        640
      ],
      "fit": "letterbox",
      "background": "white"
    },
    "source_media": "real_turntable_photography",
    "frame_source": "abo_electric_turntable_capture",
    "note": "A/B/C are separate samples linked by group_id and share prompt/target."
  },
  "grouping": {
    "group_id": "abo_020fc400_N2",
    "variants": [
      "A",
      "B",
      "C"
    ],
    "shared_prompt": true,
    "shared_target": true
  },
  "input": {
    "variant": "C",
    "type": "image_set",
    "path": "input_images/",
    "count": 2,
    "view_azimuths_deg": [
      0,
      180
    ],
    "items": [
      {
        "path": "input_images/01_020fc400_00.jpg",
        "source_filename": "020fc400_00.jpg",
        "azimuth_deg": 0,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/02/020fc400/020fc400_00.jpg"
      },
      {
        "path": "input_images/02_020fc400_36.jpg",
        "source_filename": "020fc400_36.jpg",
        "azimuth_deg": 180,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/02/020fc400/020fc400_36.jpg"
      }
    ]
  },
  "video": {
    "num_frames_input": 18,
    "num_frames_target": 36,
    "fps": 10,
    "resolution": [
      640,
      640
    ],
    "native_resolution": [
      1173,
      884
    ],
    "duration_input_s": 1.8,
    "duration_target_s": 3.6,
    "source_media": "real_turntable_jpg_sequence"
  },
  "prompt": {
    "v2v_instruction": "You are given two reference videos of a single product plus 2 source images from the same 2 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background."
  },
  "domain": {
    "field": "object_centric_ecommerce",
    "kind": "real_product_turntable",
    "attributes": {
      "input_variant": "C",
      "n_views": 2
    },
    "source": {
      "dataset": "ABO",
      "spin_id": "020fc400"
    }
  },
  "source_dataset": {
    "dataset": "ABO (Amazon Berkeley Objects) spins",
    "license": "CC BY-NC 4.0",
    "origin": "Guillaumin et al. (Amazon.com) & Collins/Malik (UC Berkeley)",
    "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/02/020fc400/"
  },
  "integrity": {
    "input_images_sha256": {
      "01_020fc400_00.jpg": "710188508db1a4af1a5190af0d0de70da02727b9bc3539bfc21875f6985066cd",
      "02_020fc400_36.jpg": "77a7372e0bccd09665e23ff14f5c1e2ba31bf1eca9e874d4ae3e713ff29048c4"
    },
    "target_sha256": "94c9dd24f4d7283dc4cd6de4107aadd6358e559e0611b4b0590ffd9b6f5c8f0c"
  },
  "provenance": {
    "engine": "v2v_taxonomy",
    "engine_version": "1.0.0",
    "pipeline_version": "v1.0.0",
    "encode": {
      "codec": "libx264",
      "pix_fmt": "yuv420p",
      "crf": 18,
      "ffmpeg": "8.1.2"
    }
  }
}

三条 metadata 共享 group_id、N、view azimuth、prompt 与 target hash；分别记录自己的 sample_id、input variant/type/path/hash。

棕色皮革扶手椅 N=3 视角

group_id: abo_050419d1_N3 · samples: 050419d1__A/B/C

N=3 · 常见档：3 视角，皮革扶手与靠背的体块要补全 · 视角方位角 = [0, 120, 240]°

SAMPLE A · input_video.mp4 · small_arc

SAMPLE B · input_video.mp4 · static

SHARED TARGET · A/B/C bytes 相同

INPUT (C) · 3 张 raw images · 正好构成 Input (B)

prompt.txt（A/B/C 三个文件内容逐字相同）

You are given two reference videos of a single product plus 3 source images from the same 3 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background.

metadata.json · 共同核心相同，sample/input 字段各自不同

metadata.json · sample A · 3118 B

{
  "schema_version": "v2v-metadata/2.0",
  "task_id": "Task_13_novel_view_generation",
  "sample_id": "050419d1__A",
  "pipeline_id": "winnie_task13_abo_pipeline",
  "created_at": "2026-07-15T00:00:00Z",
  "generation_mode": "video_to_video",
  "transformation": {
    "kind": "product_novel_view_synthesis",
    "input_variant": "A",
    "n_views": 3,
    "view_azimuths_deg": [
      0,
      120,
      240
    ],
    "per_view_arc_deg": 10.0,
    "target": "orbit_360",
    "target_frames": 36,
    "target_az_step_deg": 10,
    "canvas": {
      "size": [
        640,
        640
      ],
      "fit": "letterbox",
      "background": "white"
    },
    "source_media": "real_turntable_photography",
    "frame_source": "abo_electric_turntable_capture",
    "note": "A/B/C are separate samples linked by group_id and share prompt/target."
  },
  "grouping": {
    "group_id": "abo_050419d1_N3",
    "variants": [
      "A",
      "B",
      "C"
    ],
    "shared_prompt": true,
    "shared_target": true
  },
  "input": {
    "variant": "A",
    "type": "video",
    "path": "input_video.mp4",
    "mode": "small_arc",
    "num_frames": 27,
    "duration_s": 2.7
  },
  "video": {
    "num_frames_input": 27,
    "num_frames_target": 36,
    "fps": 10,
    "resolution": [
      640,
      640
    ],
    "native_resolution": [
      1117,
      761
    ],
    "duration_input_s": 2.7,
    "duration_target_s": 3.6,
    "source_media": "real_turntable_jpg_sequence"
  },
  "prompt": {
    "v2v_instruction": "You are given two reference videos of a single product plus 3 source images from the same 3 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background."
  },
  "domain": {
    "field": "object_centric_ecommerce",
    "kind": "real_product_turntable",
    "attributes": {
      "input_variant": "A",
      "n_views": 3
    },
    "source": {
      "dataset": "ABO",
      "spin_id": "050419d1"
    }
  },
  "source_dataset": {
    "dataset": "ABO (Amazon Berkeley Objects) spins",
    "license": "CC BY-NC 4.0",
    "origin": "Guillaumin et al. (Amazon.com) & Collins/Malik (UC Berkeley)",
    "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/05/050419d1/"
  },
  "integrity": {
    "input_sha256": "75e4752dd3c0ed9a4741cfc8737189ee6e42009c495eb9b540b75145ac500009",
    "target_sha256": "8d8d04a434ef37fea4ab397696827ee39f70dbadee27a15d5ab4131cb19d3b7d"
  },
  "provenance": {
    "engine": "v2v_taxonomy",
    "engine_version": "1.0.0",
    "pipeline_version": "v1.0.0",
    "encode": {
      "codec": "libx264",
      "pix_fmt": "yuv420p",
      "crf": 18,
      "ffmpeg": "8.1.2"
    }
  }
}

metadata.json · sample B · 3115 B

{
  "schema_version": "v2v-metadata/2.0",
  "task_id": "Task_13_novel_view_generation",
  "sample_id": "050419d1__B",
  "pipeline_id": "winnie_task13_abo_pipeline",
  "created_at": "2026-07-15T00:00:00Z",
  "generation_mode": "video_to_video",
  "transformation": {
    "kind": "product_novel_view_synthesis",
    "input_variant": "B",
    "n_views": 3,
    "view_azimuths_deg": [
      0,
      120,
      240
    ],
    "per_view_arc_deg": 10.0,
    "target": "orbit_360",
    "target_frames": 36,
    "target_az_step_deg": 10,
    "canvas": {
      "size": [
        640,
        640
      ],
      "fit": "letterbox",
      "background": "white"
    },
    "source_media": "real_turntable_photography",
    "frame_source": "abo_electric_turntable_capture",
    "note": "A/B/C are separate samples linked by group_id and share prompt/target."
  },
  "grouping": {
    "group_id": "abo_050419d1_N3",
    "variants": [
      "A",
      "B",
      "C"
    ],
    "shared_prompt": true,
    "shared_target": true
  },
  "input": {
    "variant": "B",
    "type": "video",
    "path": "input_video.mp4",
    "mode": "static",
    "num_frames": 27,
    "duration_s": 2.7
  },
  "video": {
    "num_frames_input": 27,
    "num_frames_target": 36,
    "fps": 10,
    "resolution": [
      640,
      640
    ],
    "native_resolution": [
      1117,
      761
    ],
    "duration_input_s": 2.7,
    "duration_target_s": 3.6,
    "source_media": "real_turntable_jpg_sequence"
  },
  "prompt": {
    "v2v_instruction": "You are given two reference videos of a single product plus 3 source images from the same 3 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background."
  },
  "domain": {
    "field": "object_centric_ecommerce",
    "kind": "real_product_turntable",
    "attributes": {
      "input_variant": "B",
      "n_views": 3
    },
    "source": {
      "dataset": "ABO",
      "spin_id": "050419d1"
    }
  },
  "source_dataset": {
    "dataset": "ABO (Amazon Berkeley Objects) spins",
    "license": "CC BY-NC 4.0",
    "origin": "Guillaumin et al. (Amazon.com) & Collins/Malik (UC Berkeley)",
    "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/05/050419d1/"
  },
  "integrity": {
    "input_sha256": "54582f69cf771beb193196abbb36b20d7387790dff921e404cd7000cc3106933",
    "target_sha256": "8d8d04a434ef37fea4ab397696827ee39f70dbadee27a15d5ab4131cb19d3b7d"
  },
  "provenance": {
    "engine": "v2v_taxonomy",
    "engine_version": "1.0.0",
    "pipeline_version": "v1.0.0",
    "encode": {
      "codec": "libx264",
      "pix_fmt": "yuv420p",
      "crf": 18,
      "ffmpeg": "8.1.2"
    }
  }
}

metadata.json · sample C · 4192 B

{
  "schema_version": "v2v-metadata/2.0",
  "task_id": "Task_13_novel_view_generation",
  "sample_id": "050419d1__C",
  "pipeline_id": "winnie_task13_abo_pipeline",
  "created_at": "2026-07-15T00:00:00Z",
  "generation_mode": "image_set_to_video",
  "transformation": {
    "kind": "product_novel_view_synthesis",
    "input_variant": "C",
    "n_views": 3,
    "view_azimuths_deg": [
      0,
      120,
      240
    ],
    "per_view_arc_deg": 10.0,
    "target": "orbit_360",
    "target_frames": 36,
    "target_az_step_deg": 10,
    "canvas": {
      "size": [
        640,
        640
      ],
      "fit": "letterbox",
      "background": "white"
    },
    "source_media": "real_turntable_photography",
    "frame_source": "abo_electric_turntable_capture",
    "note": "A/B/C are separate samples linked by group_id and share prompt/target."
  },
  "grouping": {
    "group_id": "abo_050419d1_N3",
    "variants": [
      "A",
      "B",
      "C"
    ],
    "shared_prompt": true,
    "shared_target": true
  },
  "input": {
    "variant": "C",
    "type": "image_set",
    "path": "input_images/",
    "count": 3,
    "view_azimuths_deg": [
      0,
      120,
      240
    ],
    "items": [
      {
        "path": "input_images/01_050419d1_00.jpg",
        "source_filename": "050419d1_00.jpg",
        "azimuth_deg": 0,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/05/050419d1/050419d1_00.jpg"
      },
      {
        "path": "input_images/02_050419d1_24.jpg",
        "source_filename": "050419d1_24.jpg",
        "azimuth_deg": 120,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/05/050419d1/050419d1_24.jpg"
      },
      {
        "path": "input_images/03_050419d1_48.jpg",
        "source_filename": "050419d1_48.jpg",
        "azimuth_deg": 240,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/05/050419d1/050419d1_48.jpg"
      }
    ]
  },
  "video": {
    "num_frames_input": 27,
    "num_frames_target": 36,
    "fps": 10,
    "resolution": [
      640,
      640
    ],
    "native_resolution": [
      1117,
      761
    ],
    "duration_input_s": 2.7,
    "duration_target_s": 3.6,
    "source_media": "real_turntable_jpg_sequence"
  },
  "prompt": {
    "v2v_instruction": "You are given two reference videos of a single product plus 3 source images from the same 3 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background."
  },
  "domain": {
    "field": "object_centric_ecommerce",
    "kind": "real_product_turntable",
    "attributes": {
      "input_variant": "C",
      "n_views": 3
    },
    "source": {
      "dataset": "ABO",
      "spin_id": "050419d1"
    }
  },
  "source_dataset": {
    "dataset": "ABO (Amazon Berkeley Objects) spins",
    "license": "CC BY-NC 4.0",
    "origin": "Guillaumin et al. (Amazon.com) & Collins/Malik (UC Berkeley)",
    "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/05/050419d1/"
  },
  "integrity": {
    "input_images_sha256": {
      "01_050419d1_00.jpg": "44592735a3de8c1f95573364783bb21540344528055aa279d8c8536bbfac90b9",
      "02_050419d1_24.jpg": "2777f192bc7745bbab323b7366e5958332e9066bc7021e0ab3cb49fe0ed98075",
      "03_050419d1_48.jpg": "0f20d1de3f27f3084878cfca9dc4966479fbd0097610d142cd02ec5d99a73fcf"
    },
    "target_sha256": "8d8d04a434ef37fea4ab397696827ee39f70dbadee27a15d5ab4131cb19d3b7d"
  },
  "provenance": {
    "engine": "v2v_taxonomy",
    "engine_version": "1.0.0",
    "pipeline_version": "v1.0.0",
    "encode": {
      "codec": "libx264",
      "pix_fmt": "yuv420p",
      "crf": 18,
      "ffmpeg": "8.1.2"
    }
  }
}

三条 metadata 共享 group_id、N、view azimuth、prompt 与 target hash；分别记录自己的 sample_id、input variant/type/path/hash。

橙色坐垫木框椅 N=3 视角

group_id: abo_090c015f_N3 · samples: 090c015f__A/B/C

N=3 · 常见档：木扶手镂空 + 橙垫，3 视角推整圈 · 视角方位角 = [0, 120, 240]°

SAMPLE A · input_video.mp4 · small_arc

SAMPLE B · input_video.mp4 · static

SHARED TARGET · A/B/C bytes 相同

INPUT (C) · 3 张 raw images · 正好构成 Input (B)

prompt.txt（A/B/C 三个文件内容逐字相同）

You are given two reference videos of a single product plus 3 source images from the same 3 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background.

metadata.json · 共同核心相同，sample/input 字段各自不同

metadata.json · sample A · 3117 B

{
  "schema_version": "v2v-metadata/2.0",
  "task_id": "Task_13_novel_view_generation",
  "sample_id": "090c015f__A",
  "pipeline_id": "winnie_task13_abo_pipeline",
  "created_at": "2026-07-15T00:00:00Z",
  "generation_mode": "video_to_video",
  "transformation": {
    "kind": "product_novel_view_synthesis",
    "input_variant": "A",
    "n_views": 3,
    "view_azimuths_deg": [
      0,
      120,
      240
    ],
    "per_view_arc_deg": 10.0,
    "target": "orbit_360",
    "target_frames": 36,
    "target_az_step_deg": 10,
    "canvas": {
      "size": [
        640,
        640
      ],
      "fit": "letterbox",
      "background": "white"
    },
    "source_media": "real_turntable_photography",
    "frame_source": "abo_electric_turntable_capture",
    "note": "A/B/C are separate samples linked by group_id and share prompt/target."
  },
  "grouping": {
    "group_id": "abo_090c015f_N3",
    "variants": [
      "A",
      "B",
      "C"
    ],
    "shared_prompt": true,
    "shared_target": true
  },
  "input": {
    "variant": "A",
    "type": "video",
    "path": "input_video.mp4",
    "mode": "small_arc",
    "num_frames": 27,
    "duration_s": 2.7
  },
  "video": {
    "num_frames_input": 27,
    "num_frames_target": 36,
    "fps": 10,
    "resolution": [
      640,
      640
    ],
    "native_resolution": [
      918,
      912
    ],
    "duration_input_s": 2.7,
    "duration_target_s": 3.6,
    "source_media": "real_turntable_jpg_sequence"
  },
  "prompt": {
    "v2v_instruction": "You are given two reference videos of a single product plus 3 source images from the same 3 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background."
  },
  "domain": {
    "field": "object_centric_ecommerce",
    "kind": "real_product_turntable",
    "attributes": {
      "input_variant": "A",
      "n_views": 3
    },
    "source": {
      "dataset": "ABO",
      "spin_id": "090c015f"
    }
  },
  "source_dataset": {
    "dataset": "ABO (Amazon Berkeley Objects) spins",
    "license": "CC BY-NC 4.0",
    "origin": "Guillaumin et al. (Amazon.com) & Collins/Malik (UC Berkeley)",
    "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/09/090c015f/"
  },
  "integrity": {
    "input_sha256": "e6faec64bba48df43d5a18fae7ab3f202b4b85b6408481220bdeae4b0d00abd0",
    "target_sha256": "f20443c243c4e8065ba30eeb0ca32ae7aa7a2c1e910d601f6b28075c2bfeadbe"
  },
  "provenance": {
    "engine": "v2v_taxonomy",
    "engine_version": "1.0.0",
    "pipeline_version": "v1.0.0",
    "encode": {
      "codec": "libx264",
      "pix_fmt": "yuv420p",
      "crf": 18,
      "ffmpeg": "8.1.2"
    }
  }
}

metadata.json · sample B · 3114 B

{
  "schema_version": "v2v-metadata/2.0",
  "task_id": "Task_13_novel_view_generation",
  "sample_id": "090c015f__B",
  "pipeline_id": "winnie_task13_abo_pipeline",
  "created_at": "2026-07-15T00:00:00Z",
  "generation_mode": "video_to_video",
  "transformation": {
    "kind": "product_novel_view_synthesis",
    "input_variant": "B",
    "n_views": 3,
    "view_azimuths_deg": [
      0,
      120,
      240
    ],
    "per_view_arc_deg": 10.0,
    "target": "orbit_360",
    "target_frames": 36,
    "target_az_step_deg": 10,
    "canvas": {
      "size": [
        640,
        640
      ],
      "fit": "letterbox",
      "background": "white"
    },
    "source_media": "real_turntable_photography",
    "frame_source": "abo_electric_turntable_capture",
    "note": "A/B/C are separate samples linked by group_id and share prompt/target."
  },
  "grouping": {
    "group_id": "abo_090c015f_N3",
    "variants": [
      "A",
      "B",
      "C"
    ],
    "shared_prompt": true,
    "shared_target": true
  },
  "input": {
    "variant": "B",
    "type": "video",
    "path": "input_video.mp4",
    "mode": "static",
    "num_frames": 27,
    "duration_s": 2.7
  },
  "video": {
    "num_frames_input": 27,
    "num_frames_target": 36,
    "fps": 10,
    "resolution": [
      640,
      640
    ],
    "native_resolution": [
      918,
      912
    ],
    "duration_input_s": 2.7,
    "duration_target_s": 3.6,
    "source_media": "real_turntable_jpg_sequence"
  },
  "prompt": {
    "v2v_instruction": "You are given two reference videos of a single product plus 3 source images from the same 3 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background."
  },
  "domain": {
    "field": "object_centric_ecommerce",
    "kind": "real_product_turntable",
    "attributes": {
      "input_variant": "B",
      "n_views": 3
    },
    "source": {
      "dataset": "ABO",
      "spin_id": "090c015f"
    }
  },
  "source_dataset": {
    "dataset": "ABO (Amazon Berkeley Objects) spins",
    "license": "CC BY-NC 4.0",
    "origin": "Guillaumin et al. (Amazon.com) & Collins/Malik (UC Berkeley)",
    "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/09/090c015f/"
  },
  "integrity": {
    "input_sha256": "6bc6c764934b63fdbe5aeeee42dd59c6d0c78f092872bf8e3e3c14cd447ca9d8",
    "target_sha256": "f20443c243c4e8065ba30eeb0ca32ae7aa7a2c1e910d601f6b28075c2bfeadbe"
  },
  "provenance": {
    "engine": "v2v_taxonomy",
    "engine_version": "1.0.0",
    "pipeline_version": "v1.0.0",
    "encode": {
      "codec": "libx264",
      "pix_fmt": "yuv420p",
      "crf": 18,
      "ffmpeg": "8.1.2"
    }
  }
}

metadata.json · sample C · 4191 B

{
  "schema_version": "v2v-metadata/2.0",
  "task_id": "Task_13_novel_view_generation",
  "sample_id": "090c015f__C",
  "pipeline_id": "winnie_task13_abo_pipeline",
  "created_at": "2026-07-15T00:00:00Z",
  "generation_mode": "image_set_to_video",
  "transformation": {
    "kind": "product_novel_view_synthesis",
    "input_variant": "C",
    "n_views": 3,
    "view_azimuths_deg": [
      0,
      120,
      240
    ],
    "per_view_arc_deg": 10.0,
    "target": "orbit_360",
    "target_frames": 36,
    "target_az_step_deg": 10,
    "canvas": {
      "size": [
        640,
        640
      ],
      "fit": "letterbox",
      "background": "white"
    },
    "source_media": "real_turntable_photography",
    "frame_source": "abo_electric_turntable_capture",
    "note": "A/B/C are separate samples linked by group_id and share prompt/target."
  },
  "grouping": {
    "group_id": "abo_090c015f_N3",
    "variants": [
      "A",
      "B",
      "C"
    ],
    "shared_prompt": true,
    "shared_target": true
  },
  "input": {
    "variant": "C",
    "type": "image_set",
    "path": "input_images/",
    "count": 3,
    "view_azimuths_deg": [
      0,
      120,
      240
    ],
    "items": [
      {
        "path": "input_images/01_090c015f_00.jpg",
        "source_filename": "090c015f_00.jpg",
        "azimuth_deg": 0,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/09/090c015f/090c015f_00.jpg"
      },
      {
        "path": "input_images/02_090c015f_24.jpg",
        "source_filename": "090c015f_24.jpg",
        "azimuth_deg": 120,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/09/090c015f/090c015f_24.jpg"
      },
      {
        "path": "input_images/03_090c015f_48.jpg",
        "source_filename": "090c015f_48.jpg",
        "azimuth_deg": 240,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/09/090c015f/090c015f_48.jpg"
      }
    ]
  },
  "video": {
    "num_frames_input": 27,
    "num_frames_target": 36,
    "fps": 10,
    "resolution": [
      640,
      640
    ],
    "native_resolution": [
      918,
      912
    ],
    "duration_input_s": 2.7,
    "duration_target_s": 3.6,
    "source_media": "real_turntable_jpg_sequence"
  },
  "prompt": {
    "v2v_instruction": "You are given two reference videos of a single product plus 3 source images from the same 3 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background."
  },
  "domain": {
    "field": "object_centric_ecommerce",
    "kind": "real_product_turntable",
    "attributes": {
      "input_variant": "C",
      "n_views": 3
    },
    "source": {
      "dataset": "ABO",
      "spin_id": "090c015f"
    }
  },
  "source_dataset": {
    "dataset": "ABO (Amazon Berkeley Objects) spins",
    "license": "CC BY-NC 4.0",
    "origin": "Guillaumin et al. (Amazon.com) & Collins/Malik (UC Berkeley)",
    "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/09/090c015f/"
  },
  "integrity": {
    "input_images_sha256": {
      "01_090c015f_00.jpg": "f3021ea693b334f1409089fd6d7c7c04e0d3d6ea8ec26d0fbf8c3e22640c4513",
      "02_090c015f_24.jpg": "b47758209d37c944f7ffad99d67138efa744284e9440039d485d4e85337b2cea",
      "03_090c015f_48.jpg": "5bd6e438fb151cd022ac1d71893c394f8f2b96812a246021be2a219165bd6d64"
    },
    "target_sha256": "f20443c243c4e8065ba30eeb0ca32ae7aa7a2c1e910d601f6b28075c2bfeadbe"
  },
  "provenance": {
    "engine": "v2v_taxonomy",
    "engine_version": "1.0.0",
    "pipeline_version": "v1.0.0",
    "encode": {
      "codec": "libx264",
      "pix_fmt": "yuv420p",
      "crf": 18,
      "ffmpeg": "8.1.2"
    }
  }
}

三条 metadata 共享 group_id、N、view azimuth、prompt 与 target hash；分别记录自己的 sample_id、input variant/type/path/hash。

软包双人床 N=4 视角

group_id: abo_04008b81_N4 · samples: 04008b81__A/B/C

N=4 · 前后左右 4 视角，床头板/床尾落差要一致 · 视角方位角 = [0, 90, 180, 270]°

SAMPLE A · input_video.mp4 · small_arc

SAMPLE B · input_video.mp4 · static

SHARED TARGET · A/B/C bytes 相同

INPUT (C) · 4 张 raw images · 正好构成 Input (B)

prompt.txt（A/B/C 三个文件内容逐字相同）

You are given two reference videos of a single product plus 4 source images from the same 4 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background.

metadata.json · 共同核心相同，sample/input 字段各自不同

metadata.json · sample A · 3128 B

{
  "schema_version": "v2v-metadata/2.0",
  "task_id": "Task_13_novel_view_generation",
  "sample_id": "04008b81__A",
  "pipeline_id": "winnie_task13_abo_pipeline",
  "created_at": "2026-07-15T00:00:00Z",
  "generation_mode": "video_to_video",
  "transformation": {
    "kind": "product_novel_view_synthesis",
    "input_variant": "A",
    "n_views": 4,
    "view_azimuths_deg": [
      0,
      90,
      180,
      270
    ],
    "per_view_arc_deg": 10.0,
    "target": "orbit_360",
    "target_frames": 36,
    "target_az_step_deg": 10,
    "canvas": {
      "size": [
        640,
        640
      ],
      "fit": "letterbox",
      "background": "white"
    },
    "source_media": "real_turntable_photography",
    "frame_source": "abo_electric_turntable_capture",
    "note": "A/B/C are separate samples linked by group_id and share prompt/target."
  },
  "grouping": {
    "group_id": "abo_04008b81_N4",
    "variants": [
      "A",
      "B",
      "C"
    ],
    "shared_prompt": true,
    "shared_target": true
  },
  "input": {
    "variant": "A",
    "type": "video",
    "path": "input_video.mp4",
    "mode": "small_arc",
    "num_frames": 36,
    "duration_s": 3.6
  },
  "video": {
    "num_frames_input": 36,
    "num_frames_target": 36,
    "fps": 10,
    "resolution": [
      640,
      640
    ],
    "native_resolution": [
      1193,
      756
    ],
    "duration_input_s": 3.6,
    "duration_target_s": 3.6,
    "source_media": "real_turntable_jpg_sequence"
  },
  "prompt": {
    "v2v_instruction": "You are given two reference videos of a single product plus 4 source images from the same 4 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background."
  },
  "domain": {
    "field": "object_centric_ecommerce",
    "kind": "real_product_turntable",
    "attributes": {
      "input_variant": "A",
      "n_views": 4
    },
    "source": {
      "dataset": "ABO",
      "spin_id": "04008b81"
    }
  },
  "source_dataset": {
    "dataset": "ABO (Amazon Berkeley Objects) spins",
    "license": "CC BY-NC 4.0",
    "origin": "Guillaumin et al. (Amazon.com) & Collins/Malik (UC Berkeley)",
    "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/04/04008b81/"
  },
  "integrity": {
    "input_sha256": "93eba376404aad4fd1daf35500509845a5e566d3b3b0419657eb02a220b3d93e",
    "target_sha256": "e1bfaea2ed05de55bbbe5f0c7029f1598f0b7839cbda209784f20f92ab727a20"
  },
  "provenance": {
    "engine": "v2v_taxonomy",
    "engine_version": "1.0.0",
    "pipeline_version": "v1.0.0",
    "encode": {
      "codec": "libx264",
      "pix_fmt": "yuv420p",
      "crf": 18,
      "ffmpeg": "8.1.2"
    }
  }
}

metadata.json · sample B · 3125 B

{
  "schema_version": "v2v-metadata/2.0",
  "task_id": "Task_13_novel_view_generation",
  "sample_id": "04008b81__B",
  "pipeline_id": "winnie_task13_abo_pipeline",
  "created_at": "2026-07-15T00:00:00Z",
  "generation_mode": "video_to_video",
  "transformation": {
    "kind": "product_novel_view_synthesis",
    "input_variant": "B",
    "n_views": 4,
    "view_azimuths_deg": [
      0,
      90,
      180,
      270
    ],
    "per_view_arc_deg": 10.0,
    "target": "orbit_360",
    "target_frames": 36,
    "target_az_step_deg": 10,
    "canvas": {
      "size": [
        640,
        640
      ],
      "fit": "letterbox",
      "background": "white"
    },
    "source_media": "real_turntable_photography",
    "frame_source": "abo_electric_turntable_capture",
    "note": "A/B/C are separate samples linked by group_id and share prompt/target."
  },
  "grouping": {
    "group_id": "abo_04008b81_N4",
    "variants": [
      "A",
      "B",
      "C"
    ],
    "shared_prompt": true,
    "shared_target": true
  },
  "input": {
    "variant": "B",
    "type": "video",
    "path": "input_video.mp4",
    "mode": "static",
    "num_frames": 36,
    "duration_s": 3.6
  },
  "video": {
    "num_frames_input": 36,
    "num_frames_target": 36,
    "fps": 10,
    "resolution": [
      640,
      640
    ],
    "native_resolution": [
      1193,
      756
    ],
    "duration_input_s": 3.6,
    "duration_target_s": 3.6,
    "source_media": "real_turntable_jpg_sequence"
  },
  "prompt": {
    "v2v_instruction": "You are given two reference videos of a single product plus 4 source images from the same 4 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background."
  },
  "domain": {
    "field": "object_centric_ecommerce",
    "kind": "real_product_turntable",
    "attributes": {
      "input_variant": "B",
      "n_views": 4
    },
    "source": {
      "dataset": "ABO",
      "spin_id": "04008b81"
    }
  },
  "source_dataset": {
    "dataset": "ABO (Amazon Berkeley Objects) spins",
    "license": "CC BY-NC 4.0",
    "origin": "Guillaumin et al. (Amazon.com) & Collins/Malik (UC Berkeley)",
    "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/04/04008b81/"
  },
  "integrity": {
    "input_sha256": "0fa903dae915b5fca236fb046830589e856b5560b134afd1541e47928c5717c6",
    "target_sha256": "e1bfaea2ed05de55bbbe5f0c7029f1598f0b7839cbda209784f20f92ab727a20"
  },
  "provenance": {
    "engine": "v2v_taxonomy",
    "engine_version": "1.0.0",
    "pipeline_version": "v1.0.0",
    "encode": {
      "codec": "libx264",
      "pix_fmt": "yuv420p",
      "crf": 18,
      "ffmpeg": "8.1.2"
    }
  }
}

metadata.json · sample C · 4575 B

{
  "schema_version": "v2v-metadata/2.0",
  "task_id": "Task_13_novel_view_generation",
  "sample_id": "04008b81__C",
  "pipeline_id": "winnie_task13_abo_pipeline",
  "created_at": "2026-07-15T00:00:00Z",
  "generation_mode": "image_set_to_video",
  "transformation": {
    "kind": "product_novel_view_synthesis",
    "input_variant": "C",
    "n_views": 4,
    "view_azimuths_deg": [
      0,
      90,
      180,
      270
    ],
    "per_view_arc_deg": 10.0,
    "target": "orbit_360",
    "target_frames": 36,
    "target_az_step_deg": 10,
    "canvas": {
      "size": [
        640,
        640
      ],
      "fit": "letterbox",
      "background": "white"
    },
    "source_media": "real_turntable_photography",
    "frame_source": "abo_electric_turntable_capture",
    "note": "A/B/C are separate samples linked by group_id and share prompt/target."
  },
  "grouping": {
    "group_id": "abo_04008b81_N4",
    "variants": [
      "A",
      "B",
      "C"
    ],
    "shared_prompt": true,
    "shared_target": true
  },
  "input": {
    "variant": "C",
    "type": "image_set",
    "path": "input_images/",
    "count": 4,
    "view_azimuths_deg": [
      0,
      90,
      180,
      270
    ],
    "items": [
      {
        "path": "input_images/01_04008b81_00.jpg",
        "source_filename": "04008b81_00.jpg",
        "azimuth_deg": 0,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/04/04008b81/04008b81_00.jpg"
      },
      {
        "path": "input_images/02_04008b81_18.jpg",
        "source_filename": "04008b81_18.jpg",
        "azimuth_deg": 90,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/04/04008b81/04008b81_18.jpg"
      },
      {
        "path": "input_images/03_04008b81_36.jpg",
        "source_filename": "04008b81_36.jpg",
        "azimuth_deg": 180,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/04/04008b81/04008b81_36.jpg"
      },
      {
        "path": "input_images/04_04008b81_54.jpg",
        "source_filename": "04008b81_54.jpg",
        "azimuth_deg": 270,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/04/04008b81/04008b81_54.jpg"
      }
    ]
  },
  "video": {
    "num_frames_input": 36,
    "num_frames_target": 36,
    "fps": 10,
    "resolution": [
      640,
      640
    ],
    "native_resolution": [
      1193,
      756
    ],
    "duration_input_s": 3.6,
    "duration_target_s": 3.6,
    "source_media": "real_turntable_jpg_sequence"
  },
  "prompt": {
    "v2v_instruction": "You are given two reference videos of a single product plus 4 source images from the same 4 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background."
  },
  "domain": {
    "field": "object_centric_ecommerce",
    "kind": "real_product_turntable",
    "attributes": {
      "input_variant": "C",
      "n_views": 4
    },
    "source": {
      "dataset": "ABO",
      "spin_id": "04008b81"
    }
  },
  "source_dataset": {
    "dataset": "ABO (Amazon Berkeley Objects) spins",
    "license": "CC BY-NC 4.0",
    "origin": "Guillaumin et al. (Amazon.com) & Collins/Malik (UC Berkeley)",
    "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/04/04008b81/"
  },
  "integrity": {
    "input_images_sha256": {
      "01_04008b81_00.jpg": "31e9c7c843a33a7ce7f97a7e240642ead71979bceeee7a4c5edbdc0053d62345",
      "02_04008b81_18.jpg": "be6e4ea2101ad6df2a543a992ee794f3045966d6574e0134f57cde45f1f8ffeb",
      "03_04008b81_36.jpg": "d90bfd15c57b8d632adf480def2cd582cbb728c838c0ad9c64b2e26dae2bd175",
      "04_04008b81_54.jpg": "9814ffc99ce7310e07db8b0136402e7b552eba8255bad89f6417eea7786e0737"
    },
    "target_sha256": "e1bfaea2ed05de55bbbe5f0c7029f1598f0b7839cbda209784f20f92ab727a20"
  },
  "provenance": {
    "engine": "v2v_taxonomy",
    "engine_version": "1.0.0",
    "pipeline_version": "v1.0.0",
    "encode": {
      "codec": "libx264",
      "pix_fmt": "yuv420p",
      "crf": 18,
      "ffmpeg": "8.1.2"
    }
  }
}

三条 metadata 共享 group_id、N、view azimuth、prompt 与 target hash；分别记录自己的 sample_id、input variant/type/path/hash。

米色翼背单椅 N=5 视角

group_id: abo_0b10d9ac_N5 · samples: 0b10d9ac__A/B/C

N=5 · 最易：5 视角密集，翼状椅耳在多数角度可见 · 视角方位角 = [0, 70, 145, 215, 290]°

SAMPLE A · input_video.mp4 · small_arc

SAMPLE B · input_video.mp4 · static

SHARED TARGET · A/B/C bytes 相同

INPUT (C) · 5 张 raw images · 正好构成 Input (B)

prompt.txt（A/B/C 三个文件内容逐字相同）

You are given two reference videos of a single product plus 5 source images from the same 5 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background.

metadata.json · 共同核心相同，sample/input 字段各自不同

metadata.json · sample A · 3138 B

{
  "schema_version": "v2v-metadata/2.0",
  "task_id": "Task_13_novel_view_generation",
  "sample_id": "0b10d9ac__A",
  "pipeline_id": "winnie_task13_abo_pipeline",
  "created_at": "2026-07-15T00:00:00Z",
  "generation_mode": "video_to_video",
  "transformation": {
    "kind": "product_novel_view_synthesis",
    "input_variant": "A",
    "n_views": 5,
    "view_azimuths_deg": [
      0,
      70,
      145,
      215,
      290
    ],
    "per_view_arc_deg": 10.0,
    "target": "orbit_360",
    "target_frames": 36,
    "target_az_step_deg": 10,
    "canvas": {
      "size": [
        640,
        640
      ],
      "fit": "letterbox",
      "background": "white"
    },
    "source_media": "real_turntable_photography",
    "frame_source": "abo_electric_turntable_capture",
    "note": "A/B/C are separate samples linked by group_id and share prompt/target."
  },
  "grouping": {
    "group_id": "abo_0b10d9ac_N5",
    "variants": [
      "A",
      "B",
      "C"
    ],
    "shared_prompt": true,
    "shared_target": true
  },
  "input": {
    "variant": "A",
    "type": "video",
    "path": "input_video.mp4",
    "mode": "small_arc",
    "num_frames": 45,
    "duration_s": 4.5
  },
  "video": {
    "num_frames_input": 45,
    "num_frames_target": 36,
    "fps": 10,
    "resolution": [
      640,
      640
    ],
    "native_resolution": [
      891,
      942
    ],
    "duration_input_s": 4.5,
    "duration_target_s": 3.6,
    "source_media": "real_turntable_jpg_sequence"
  },
  "prompt": {
    "v2v_instruction": "You are given two reference videos of a single product plus 5 source images from the same 5 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background."
  },
  "domain": {
    "field": "object_centric_ecommerce",
    "kind": "real_product_turntable",
    "attributes": {
      "input_variant": "A",
      "n_views": 5
    },
    "source": {
      "dataset": "ABO",
      "spin_id": "0b10d9ac"
    }
  },
  "source_dataset": {
    "dataset": "ABO (Amazon Berkeley Objects) spins",
    "license": "CC BY-NC 4.0",
    "origin": "Guillaumin et al. (Amazon.com) & Collins/Malik (UC Berkeley)",
    "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/0b/0b10d9ac/"
  },
  "integrity": {
    "input_sha256": "706b18ee70aeec6f4e52f9658df1ebbe9e656e474a816a96fb4c6dac7bdb4ffd",
    "target_sha256": "fd69ff298e75e1895da486a5a01dfeb66320df5e314ce6def129305b9dd40ed2"
  },
  "provenance": {
    "engine": "v2v_taxonomy",
    "engine_version": "1.0.0",
    "pipeline_version": "v1.0.0",
    "encode": {
      "codec": "libx264",
      "pix_fmt": "yuv420p",
      "crf": 18,
      "ffmpeg": "8.1.2"
    }
  }
}

metadata.json · sample B · 3135 B

{
  "schema_version": "v2v-metadata/2.0",
  "task_id": "Task_13_novel_view_generation",
  "sample_id": "0b10d9ac__B",
  "pipeline_id": "winnie_task13_abo_pipeline",
  "created_at": "2026-07-15T00:00:00Z",
  "generation_mode": "video_to_video",
  "transformation": {
    "kind": "product_novel_view_synthesis",
    "input_variant": "B",
    "n_views": 5,
    "view_azimuths_deg": [
      0,
      70,
      145,
      215,
      290
    ],
    "per_view_arc_deg": 10.0,
    "target": "orbit_360",
    "target_frames": 36,
    "target_az_step_deg": 10,
    "canvas": {
      "size": [
        640,
        640
      ],
      "fit": "letterbox",
      "background": "white"
    },
    "source_media": "real_turntable_photography",
    "frame_source": "abo_electric_turntable_capture",
    "note": "A/B/C are separate samples linked by group_id and share prompt/target."
  },
  "grouping": {
    "group_id": "abo_0b10d9ac_N5",
    "variants": [
      "A",
      "B",
      "C"
    ],
    "shared_prompt": true,
    "shared_target": true
  },
  "input": {
    "variant": "B",
    "type": "video",
    "path": "input_video.mp4",
    "mode": "static",
    "num_frames": 45,
    "duration_s": 4.5
  },
  "video": {
    "num_frames_input": 45,
    "num_frames_target": 36,
    "fps": 10,
    "resolution": [
      640,
      640
    ],
    "native_resolution": [
      891,
      942
    ],
    "duration_input_s": 4.5,
    "duration_target_s": 3.6,
    "source_media": "real_turntable_jpg_sequence"
  },
  "prompt": {
    "v2v_instruction": "You are given two reference videos of a single product plus 5 source images from the same 5 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background."
  },
  "domain": {
    "field": "object_centric_ecommerce",
    "kind": "real_product_turntable",
    "attributes": {
      "input_variant": "B",
      "n_views": 5
    },
    "source": {
      "dataset": "ABO",
      "spin_id": "0b10d9ac"
    }
  },
  "source_dataset": {
    "dataset": "ABO (Amazon Berkeley Objects) spins",
    "license": "CC BY-NC 4.0",
    "origin": "Guillaumin et al. (Amazon.com) & Collins/Malik (UC Berkeley)",
    "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/0b/0b10d9ac/"
  },
  "integrity": {
    "input_sha256": "5fba66ff68087428cdd91e117c19426fedf21282c95a2a5a612281fa282df20f",
    "target_sha256": "fd69ff298e75e1895da486a5a01dfeb66320df5e314ce6def129305b9dd40ed2"
  },
  "provenance": {
    "engine": "v2v_taxonomy",
    "engine_version": "1.0.0",
    "pipeline_version": "v1.0.0",
    "encode": {
      "codec": "libx264",
      "pix_fmt": "yuv420p",
      "crf": 18,
      "ffmpeg": "8.1.2"
    }
  }
}

metadata.json · sample C · 4960 B

{
  "schema_version": "v2v-metadata/2.0",
  "task_id": "Task_13_novel_view_generation",
  "sample_id": "0b10d9ac__C",
  "pipeline_id": "winnie_task13_abo_pipeline",
  "created_at": "2026-07-15T00:00:00Z",
  "generation_mode": "image_set_to_video",
  "transformation": {
    "kind": "product_novel_view_synthesis",
    "input_variant": "C",
    "n_views": 5,
    "view_azimuths_deg": [
      0,
      70,
      145,
      215,
      290
    ],
    "per_view_arc_deg": 10.0,
    "target": "orbit_360",
    "target_frames": 36,
    "target_az_step_deg": 10,
    "canvas": {
      "size": [
        640,
        640
      ],
      "fit": "letterbox",
      "background": "white"
    },
    "source_media": "real_turntable_photography",
    "frame_source": "abo_electric_turntable_capture",
    "note": "A/B/C are separate samples linked by group_id and share prompt/target."
  },
  "grouping": {
    "group_id": "abo_0b10d9ac_N5",
    "variants": [
      "A",
      "B",
      "C"
    ],
    "shared_prompt": true,
    "shared_target": true
  },
  "input": {
    "variant": "C",
    "type": "image_set",
    "path": "input_images/",
    "count": 5,
    "view_azimuths_deg": [
      0,
      70,
      145,
      215,
      290
    ],
    "items": [
      {
        "path": "input_images/01_0b10d9ac_00.jpg",
        "source_filename": "0b10d9ac_00.jpg",
        "azimuth_deg": 0,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/0b/0b10d9ac/0b10d9ac_00.jpg"
      },
      {
        "path": "input_images/02_0b10d9ac_14.jpg",
        "source_filename": "0b10d9ac_14.jpg",
        "azimuth_deg": 70,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/0b/0b10d9ac/0b10d9ac_14.jpg"
      },
      {
        "path": "input_images/03_0b10d9ac_29.jpg",
        "source_filename": "0b10d9ac_29.jpg",
        "azimuth_deg": 145,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/0b/0b10d9ac/0b10d9ac_29.jpg"
      },
      {
        "path": "input_images/04_0b10d9ac_43.jpg",
        "source_filename": "0b10d9ac_43.jpg",
        "azimuth_deg": 215,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/0b/0b10d9ac/0b10d9ac_43.jpg"
      },
      {
        "path": "input_images/05_0b10d9ac_58.jpg",
        "source_filename": "0b10d9ac_58.jpg",
        "azimuth_deg": 290,
        "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/0b/0b10d9ac/0b10d9ac_58.jpg"
      }
    ]
  },
  "video": {
    "num_frames_input": 45,
    "num_frames_target": 36,
    "fps": 10,
    "resolution": [
      640,
      640
    ],
    "native_resolution": [
      891,
      942
    ],
    "duration_input_s": 4.5,
    "duration_target_s": 3.6,
    "source_media": "real_turntable_jpg_sequence"
  },
  "prompt": {
    "v2v_instruction": "You are given two reference videos of a single product plus 5 source images from the same 5 viewpoints. Input A contains a small real arc around each viewpoint, Input B freezes each viewpoint, and Input C contains the exact source images that compose Input B. Generate a smooth 360-degree turntable video of the SAME product: rotate it one full turn about its vertical axis at constant elevation, advancing the azimuth steadily. Keep the product's identity, geometry, colour and material consistent from every angle, and plausibly synthesize the sides not shown. Output 36 frames at 10 fps on a clean white background."
  },
  "domain": {
    "field": "object_centric_ecommerce",
    "kind": "real_product_turntable",
    "attributes": {
      "input_variant": "C",
      "n_views": 5
    },
    "source": {
      "dataset": "ABO",
      "spin_id": "0b10d9ac"
    }
  },
  "source_dataset": {
    "dataset": "ABO (Amazon Berkeley Objects) spins",
    "license": "CC BY-NC 4.0",
    "origin": "Guillaumin et al. (Amazon.com) & Collins/Malik (UC Berkeley)",
    "raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/spins/original/0b/0b10d9ac/"
  },
  "integrity": {
    "input_images_sha256": {
      "01_0b10d9ac_00.jpg": "8cd687182d1fc7f7e68eb5dbc77895691ad4373e60edb1a1a6cf9cc6fb9833e1",
      "02_0b10d9ac_14.jpg": "e45eb19380f104b5166f4c82c1671fa3d9d27733c2d3c797d28c62f5aa54792f",
      "03_0b10d9ac_29.jpg": "d7ffa27f5c471e8cc4e238952dc8f67e2d66bff81ce72bd533780e7433c391d6",
      "04_0b10d9ac_43.jpg": "89f59f25a6b2b2dc7ffce48be530f54027bf9b36e42eb21f9ee03c787b71abaa",
      "05_0b10d9ac_58.jpg": "b432d71358bc4bcd835294750f2783543d1fea638eea19830112098d2aa529f8"
    },
    "target_sha256": "fd69ff298e75e1895da486a5a01dfeb66320df5e314ce6def129305b9dd40ed2"
  },
  "provenance": {
    "engine": "v2v_taxonomy",
    "engine_version": "1.0.0",
    "pipeline_version": "v1.0.0",
    "encode": {
      "codec": "libx264",
      "pix_fmt": "yuv420p",
      "crf": 18,
      "ffmpeg": "8.1.2"
    }
  }
}

三条 metadata 共享 group_id、N、view azimuth、prompt 与 target hash；分别记录自己的 sample_id、input variant/type/path/hash。

可调旋钮（你说了算）

旋钮	现值 → 可选	影响
N 分布	25:40:20:15（N=2/3/4/5）	难度混合比例；可调整各档占比
per_view_arc	~10°（相邻 3 真帧）→ ±5°/±15°	A 的每视角"小转动量"
target_frames	36 → 48/72	整圈更顺（72=用满真帧，但比 Task_13 长一倍）
canvas	640² letterbox 白底 → 768²/保留原生	归一策略（方案1 已选：pad 到方形）

还没做（等你点头开批量）

ABO 批量：8,116 干净 spin × A/B/C 三 variant = 约 24,348 samples；N 按 25:40:20:15 分配。
现存 GSO final 迁移：旧 2,060 条 A/B 配对保留并新增约 1,030 条 C image-set sample，总计约 3,090。
S3 写入尚未执行：需在有实例角色的 us-east-2 EC2 上生成新 v2 前缀，完成计数/hash/媒体 QA 后再切换，旧数据保留。

▶ 以后要做 raw→final 时：怎么跟 agent 说 / 给什么参考

规格已更新为 A/B/C 三条 sample 组成一个 group。后续 batch 必须遵守下面的结构：

给 agent 的话（可直接复制）

帮我把 Task_13（ABO 商品转盘）从 raw 批量生成 A/B/C 三样本 group：

- 每个 spin 生成连续三个 sample_id，用同一个 group_id 关联
- sample A/B 各含 input_video.mp4；sample C 含 input_images/ 的 N 张 raw JPG
- 三条 sample 各自保存 target_video.mp4、prompt.txt、metadata.json
- 三份 target bytes 与 prompt 内容完全相同
- metadata 共同核心相同，但 sample_id/input_variant/input_type/input hash 各自不同
- 规格权威 = build16_preview.py + abo_demo.py
- 数据实测 = task16-phase0-data-verification.md（8,116 干净 72 帧 spin；93 个 24 帧的剔除；
  分辨率不一按每条原生 letterbox；license CC BY-NC 4.0）
- 从 s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/ABO/ 读
- 写新 v2 前缀，先跑 5 个真样本 QA，批准后再全量；不要覆盖旧 final_v1。

冻结的规格（agent 应遵守，也是你日后核对的清单）：

项	值
N 视角分布	N=2/3/4/5 按 25:40:20:15 确定性分配（每 spin 一个 N）
INPUT A（small_arc）	N 个绕圈均匀视角，每视角相邻 3 真帧、各 hold 3 → ~10° 真转 / ~0.9s
INPUT B（static）	同样 N 个视角，每视角冻结；A/B 唯一差别 = 动/不动
INPUT C（raw images）	构成 B 的 N 张中心视角 JPG；数量必须等于 N
TARGET	72 真帧抽半 → 36 帧 @10fps（10°/帧，3.6s），A/B/C 共用
画布	letterbox 到 640² 白底，保长宽比不拉伸（原生尺寸不一）
取舍	只用 8,116 个满 72 帧 spin；93 个仅 24 帧的剔除
样本结构	一个 group 含 A/B/C 三个连续 sample_id；A/B 为 input_video.mp4，C 为 input_images/
输出	Task_13 新 v2 前缀 samples/<9位id>/ + index.jsonl；旧 final_v1 保留至 QA 完成

参考文件全部在 research/2026-07-13_task13-16-ingest/： build16_preview.py · abo_demo.py · task16-phase0-data-verification.md · render_task13.py（批量骨架范例）· 本 preview 页（已批准的切法快照）。唯一需要重做 preview 的情况：你改了上表任一规格。

Yilan Zhang (Winnie) × Codex · 2026-07-17 · 编码器 abo_demo.py + 样品/页构建 build16_preview.py · 数据核验 task16-phase0-data-verification.md · 报告 Task_13 dataset report