本任务（物体新视角/360°）的 GSO 源样品：给 N 个离散视角 → 补出 360° 整圈；含 leader 要的 static/小弧双输入配对。另一源 ABO（真实转台）见 ABO 样品页；场景级 RE10K 已随重组移到 Task_16。
metadata.pairing.pair_id，事后可按 pair_id 拉齐做「运动视差有没有用」的对照实验。| 项 | 值 | 说明 |
|---|---|---|
| 数据源 | GSO（Google Scanned Objects, 1,030 件） | raw 无现成图，全部 Blender Cycles 渲 |
| N 视角 | 2 / 3 / 5（每物体一个） | 难度旋钮：本批故意跨样本变化，metadata 逐条记录 |
| INPUT | N×10 帧 @10fps（每视角 1s） | static=冻结帧重复；small_arc=±6° 扫 |
| TARGET | 36 帧 @10fps（360°，10°/帧） | 全批固定，两 sample 共享同一文件 |
| 取景 | cam_dist_k=1.5（较上版 1.9 拉近） | 全批统一，物体填满约 55%；非按 sample 变 |
| 画幅/编码 | 640×640 · H.264 · yuv420p | 浏览器可放，符合指南 |
gso_3D_Dollhouse_Lamp_N2N=2 · 最难：只给 2 个视角，背面基本靠推断
You are given a short reference video showing a single 3D object from 2 viewpoints spaced evenly around it (a fixed camera elevation, the azimuth changing between viewpoints). Generate a smooth 360-degree turntable video of the SAME object: rotate it one full turn about its vertical axis at constant camera elevation and distance, advancing the viewing azimuth steadily. Keep the object's identity, geometry, colour and texture consistent across every angle, and plausibly synthesize the sides not shown in the reference. Output 36 frames at 10 fps.
{
"schema_version": "v2v-metadata/2.0",
"task_id": "Task_13_novel_view_generation",
"sample_id": "3D_Dollhouse_Lamp__static",
"pipeline_id": "winnie_task13_pipeline",
"created_at": "2026-07-14T08:42:20.455196+00:00",
"generation_mode": "video_to_video",
"transformation": {
"kind": "novel_view_synthesis",
"input_mode": "static",
"n_views": 2,
"view_azimuths_deg": [
45.0,
225.0
],
"arc_span_deg": 0.0,
"hold_frames_per_view": 10,
"target": "orbit_360",
"target_frames": 36,
"target_az_start_deg": 45.0,
"target_az_step_deg": 10.0,
"camera": {
"axis": "world_z",
"elevation_k": 0.9,
"dist_k": 1.5
},
"note": "INPUT = N discrete viewpoints around the object (static held frames, or +/-arc sweep); TARGET = full 360-degree turntable, shared across the two paired input modes."
},
"pairing": {
"pair_id": "gso_3D_Dollhouse_Lamp_N2",
"input_mode": "static",
"counterpart_input_mode": "small_arc",
"shared_target": true,
"note": "static vs small_arc share the same target and pair_id; the only variable is the input \u2014 a research ablation on motion parallax."
},
"video": {
"num_frames_input": 20,
"num_frames_target": 36,
"fps": 10,
"resolution": [
640,
640
],
"duration_input_s": 2.0,
"duration_target_s": 3.6,
"source_media": "3d_scanned_mesh",
"frame_source": "blender_cycles_render"
},
"prompt": {
"v2v_instruction": "You are given a short reference video showing a single 3D object from 2 viewpoints spaced evenly around it (a fixed camera elevation, the azimuth changing between viewpoints). Generate a smooth 360-degree turntable video of the SAME object: rotate it one full turn about its vertical axis at constant camera elevation and distance, advancing the viewing azimuth steadily. Keep the object's identity, geometry, colour and texture consistent across every angle, and plausibly synthesize the sides not shown in the reference. Output 36 frames at 10 fps."
},
"domain": {
"field": "object_centric_3d",
"kind": "synthetic_turntable_render",
"attributes": {
"subtype": [
"novel_view_synthesis",
"multi_view_input"
],
"input_mode": "static",
"n_views": 2
},
"source": {
"dataset": "GSO",
"object": "3D_Dollhouse_Lamp"
}
},
"source_dataset": {
"dataset": "GSO (Google Scanned Objects)",
"license": "CC BY 4.0",
"origin": "Downs et al., Google Research",
"paper": "Downs et al., 'Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items', ICRA 2022 (arXiv:2204.11918)",
"raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/GSO/3D_Dollhouse_Lamp.zip"
},
"integrity": {
"input_sha256": "99bdb4f762736f3387312a75d6cd02d380e46ba7dbe6bf3ffe011ea4657d679e",
"target_sha256": "e0ed25ddb7e1fc52190e9aeada788e65fdfffc319f7cbba60831aeb44c432f85"
},
"provenance": {
"engine": "v2v_taxonomy",
"engine_version": "1.0.0",
"pipeline_version": "v1.0.0",
"render": {
"engine": "cycles",
"samples": 16,
"blender": "5.1.2"
}
}
}
input_mode / pairing / integrity 三处不同。采用全组通用 v2v-metadata/2.0（与 Task_6 / 12 / 31 顶层结构一致；因无逐帧动作流故比 Task_12 短——这是完整文件，不是节选）。gso_11pro_SL_TRX_FG_N3N=3 · 中等：高度不对称（鞋底钉 vs 鞋口鞋带），3 视角
You are given a short reference video showing a single 3D object from 3 viewpoints spaced evenly around it (a fixed camera elevation, the azimuth changing between viewpoints). Generate a smooth 360-degree turntable video of the SAME object: rotate it one full turn about its vertical axis at constant camera elevation and distance, advancing the viewing azimuth steadily. Keep the object's identity, geometry, colour and texture consistent across every angle, and plausibly synthesize the sides not shown in the reference. Output 36 frames at 10 fps.
{
"schema_version": "v2v-metadata/2.0",
"task_id": "Task_13_novel_view_generation",
"sample_id": "11pro_SL_TRX_FG__static",
"pipeline_id": "winnie_task13_pipeline",
"created_at": "2026-07-14T08:44:47.626989+00:00",
"generation_mode": "video_to_video",
"transformation": {
"kind": "novel_view_synthesis",
"input_mode": "static",
"n_views": 3,
"view_azimuths_deg": [
45.0,
165.0,
285.0
],
"arc_span_deg": 0.0,
"hold_frames_per_view": 10,
"target": "orbit_360",
"target_frames": 36,
"target_az_start_deg": 45.0,
"target_az_step_deg": 10.0,
"camera": {
"axis": "world_z",
"elevation_k": 0.9,
"dist_k": 1.5
},
"note": "INPUT = N discrete viewpoints around the object (static held frames, or +/-arc sweep); TARGET = full 360-degree turntable, shared across the two paired input modes."
},
"pairing": {
"pair_id": "gso_11pro_SL_TRX_FG_N3",
"input_mode": "static",
"counterpart_input_mode": "small_arc",
"shared_target": true,
"note": "static vs small_arc share the same target and pair_id; the only variable is the input \u2014 a research ablation on motion parallax."
},
"video": {
"num_frames_input": 30,
"num_frames_target": 36,
"fps": 10,
"resolution": [
640,
640
],
"duration_input_s": 3.0,
"duration_target_s": 3.6,
"source_media": "3d_scanned_mesh",
"frame_source": "blender_cycles_render"
},
"prompt": {
"v2v_instruction": "You are given a short reference video showing a single 3D object from 3 viewpoints spaced evenly around it (a fixed camera elevation, the azimuth changing between viewpoints). Generate a smooth 360-degree turntable video of the SAME object: rotate it one full turn about its vertical axis at constant camera elevation and distance, advancing the viewing azimuth steadily. Keep the object's identity, geometry, colour and texture consistent across every angle, and plausibly synthesize the sides not shown in the reference. Output 36 frames at 10 fps."
},
"domain": {
"field": "object_centric_3d",
"kind": "synthetic_turntable_render",
"attributes": {
"subtype": [
"novel_view_synthesis",
"multi_view_input"
],
"input_mode": "static",
"n_views": 3
},
"source": {
"dataset": "GSO",
"object": "11pro_SL_TRX_FG"
}
},
"source_dataset": {
"dataset": "GSO (Google Scanned Objects)",
"license": "CC BY 4.0",
"origin": "Downs et al., Google Research",
"paper": "Downs et al., 'Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items', ICRA 2022 (arXiv:2204.11918)",
"raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/GSO/11pro_SL_TRX_FG.zip"
},
"integrity": {
"input_sha256": "f68ca275ad1d264c9aa23676fd914fbabdb34d70415b330ec5b5c3ae443aba82",
"target_sha256": "b5445b0edb0b567ba0816a9d25ed7482aafc7d91d918ac012da84a59335324a9"
},
"provenance": {
"engine": "v2v_taxonomy",
"engine_version": "1.0.0",
"pipeline_version": "v1.0.0",
"render": {
"engine": "cycles",
"samples": 16,
"blender": "5.1.2"
}
}
}
input_mode / pairing / integrity 三处不同。采用全组通用 v2v-metadata/2.0（与 Task_6 / 12 / 31 顶层结构一致；因无逐帧动作流故比 Task_12 短——这是完整文件，不是节选）。gso_3D_Dollhouse_Sofa_N5N=5 · 最易：5 视角，紫色垫只在部分角度露出
You are given a short reference video showing a single 3D object from 5 viewpoints spaced evenly around it (a fixed camera elevation, the azimuth changing between viewpoints). Generate a smooth 360-degree turntable video of the SAME object: rotate it one full turn about its vertical axis at constant camera elevation and distance, advancing the viewing azimuth steadily. Keep the object's identity, geometry, colour and texture consistent across every angle, and plausibly synthesize the sides not shown in the reference. Output 36 frames at 10 fps.
{
"schema_version": "v2v-metadata/2.0",
"task_id": "Task_13_novel_view_generation",
"sample_id": "3D_Dollhouse_Sofa__static",
"pipeline_id": "winnie_task13_pipeline",
"created_at": "2026-07-14T08:47:20.257083+00:00",
"generation_mode": "video_to_video",
"transformation": {
"kind": "novel_view_synthesis",
"input_mode": "static",
"n_views": 5,
"view_azimuths_deg": [
45.0,
117.0,
189.0,
261.0,
333.0
],
"arc_span_deg": 0.0,
"hold_frames_per_view": 10,
"target": "orbit_360",
"target_frames": 36,
"target_az_start_deg": 45.0,
"target_az_step_deg": 10.0,
"camera": {
"axis": "world_z",
"elevation_k": 0.9,
"dist_k": 1.5
},
"note": "INPUT = N discrete viewpoints around the object (static held frames, or +/-arc sweep); TARGET = full 360-degree turntable, shared across the two paired input modes."
},
"pairing": {
"pair_id": "gso_3D_Dollhouse_Sofa_N5",
"input_mode": "static",
"counterpart_input_mode": "small_arc",
"shared_target": true,
"note": "static vs small_arc share the same target and pair_id; the only variable is the input \u2014 a research ablation on motion parallax."
},
"video": {
"num_frames_input": 50,
"num_frames_target": 36,
"fps": 10,
"resolution": [
640,
640
],
"duration_input_s": 5.0,
"duration_target_s": 3.6,
"source_media": "3d_scanned_mesh",
"frame_source": "blender_cycles_render"
},
"prompt": {
"v2v_instruction": "You are given a short reference video showing a single 3D object from 5 viewpoints spaced evenly around it (a fixed camera elevation, the azimuth changing between viewpoints). Generate a smooth 360-degree turntable video of the SAME object: rotate it one full turn about its vertical axis at constant camera elevation and distance, advancing the viewing azimuth steadily. Keep the object's identity, geometry, colour and texture consistent across every angle, and plausibly synthesize the sides not shown in the reference. Output 36 frames at 10 fps."
},
"domain": {
"field": "object_centric_3d",
"kind": "synthetic_turntable_render",
"attributes": {
"subtype": [
"novel_view_synthesis",
"multi_view_input"
],
"input_mode": "static",
"n_views": 5
},
"source": {
"dataset": "GSO",
"object": "3D_Dollhouse_Sofa"
}
},
"source_dataset": {
"dataset": "GSO (Google Scanned Objects)",
"license": "CC BY 4.0",
"origin": "Downs et al., Google Research",
"paper": "Downs et al., 'Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items', ICRA 2022 (arXiv:2204.11918)",
"raw_uri": "s3://tv2v-data/raw/taxonomy/Task_13_novel_view_generation/GSO/3D_Dollhouse_Sofa.zip"
},
"integrity": {
"input_sha256": "534d621fd4366bab1e521f6ebbcdb4f36fec0c5ed585d17fd9344229df411fd9",
"target_sha256": "b36248bb54f09de3f5ed00db157a28a7c716c0c6c395a2a89610d88a8485efea"
},
"provenance": {
"engine": "v2v_taxonomy",
"engine_version": "1.0.0",
"pipeline_version": "v1.0.0",
"render": {
"engine": "cycles",
"samples": 16,
"blender": "5.1.2"
}
}
}
input_mode / pairing / integrity 三处不同。采用全组通用 v2v-metadata/2.0（与 Task_6 / 12 / 31 顶层结构一致；因无逐帧动作流故比 Task_12 短——这是完整文件，不是节选）。render_task13.py + t13_blender.py（--preview 模式）· 遵循
raw→final 指南 · 改进档案见 Task_13 定稿记 #02