# switch_inference 数据格式说明 本目录包含 8 条第一视角 prompt switch 推理数据:原始数据中每个输入场景保留 1 条代表样本,共 6 条;再追加 2 条新设计样本。每条样本由一张输入图像、一个相机动作轨迹和一份包含 switch 前后 prompt 的 JSON 文件组成。 ## 数据点 ### 原始 6 条 原始 6 条按源数据每个 `scene_index` 的第一条 prompt-switch case 保留,来源为: `https://huggingface.co/datasets/Andyson/lingbot/tree/main/prompt-switch/2026-05-25_magic_prompt_switch_training_free_inputs` | case_id | 场景 | action_label | | --- | --- | --- | | `ps_00_act_demo0_npy_00` | 原始 scene 0 | `fireball` | | `ps_05_act_demo1_npy_05` | 原始 scene 1 | `fireball` | | `ps_10_act_demo2_npy_10` | 原始 scene 2 | `fireball` | | `ps_15_act_demo0_npy_15` | 原始 scene 3 | `fireball` | | `ps_20_act_demo1_npy_20` | 原始 scene 4 | `fireball` | | `ps_25_act_demo2_npy_25` | 原始 scene 5 | `fireball` | ### 新增 2 条 | case_id | 场景 | switch 前 | switch 后 | | --- | --- | --- | --- | | `custom_00_court_present_painting` | 欧洲风格宫廷大厅 | 第一视角在宫廷中行走 | 掏出画作并双手献上 | | `custom_01_seaside_jump_swim` | 海边面向大海 | 第一视角在海边散步 | 跳入水中并向前游泳 | ## 目录结构 ```text switch_inference/ images/ # 输入图像,共 8 张 actions/ # 每个 case 一条 camera trajectory,共 8 个 .npy prompts/ # 每个 case 一份 prompt JSON,共 8 个 source_metadata/ # 原始选择、图像生成 prompt、轨迹生成说明和构建摘要 manifest.jsonl # 逐行 JSON manifest,推荐程序读取 manifest.csv # CSV 版本 manifest,方便人工查看 dataset_info.json # 全局信息 README.md # 本说明文档 ``` ## Manifest 字段 `manifest.jsonl` 每一行对应一个 case: | 字段 | 含义 | | --- | --- | | `case_id` | 样本唯一 ID | | `image_path` | 输入图像路径,相对 `switch_inference/` | | `action_path` | 相机动作轨迹路径,相对 `switch_inference/` | | `prompt_path` | prompt JSON 路径,相对 `switch_inference/` | | `action_label` | switch 后动作标签 | | `switch_frame` | prompt 切换帧,当前为 160 | | `prompt_before_en` / `prompt_before_zh` | switch 前 prompt | | `prompt_after_en` / `prompt_after_zh` | switch 后 prompt | ## Prompt JSON `prompts/.json` 包含完整 prompt 和调度信息: ```json { "case_id": "custom_00_court_present_painting", "image_path": "images/scene_06_court.png", "action_path": "actions/custom_00_court_present_painting_action.npy", "action_type": "camera_trajectory", "switch_frame": 160, "prompt_before": {"en": "...", "zh": "..."}, "prompt_after": {"en": "...", "zh": "..."}, "prompt_schedule": [ {"frame_start": 0, "frame_end": 159, "prompt_key": "prompt_before"}, {"frame_start": 160, "frame_end": 319, "prompt_key": "prompt_after"} ] } ``` ## Action 文件 `actions/*.npy` 是程序生成的第一视角相机轨迹: - `dtype`: `float32` - `shape`: `[321, 4, 4]` - 含义:每帧相机位姿矩阵;321 个 pose 对应 320 帧生成视频加初始 pose。 - 原始 6 条:保留源数据对应的 camera trajectory。 - 宫廷样本:前半段平稳向前行走,后半段放慢并轻微俯视/前倾,模拟取出并献上画作。 - 海边样本:前半段向海边行走,后半段向前入水、下沉到近水面高度并带轻微游泳摆动。 数据本身不绑定 fps;如果推理配置使用 16 fps,则 320 帧约为 20 秒,`switch_frame=160` 对应约第 10 秒。 ## 数据规模 - case 数量:8(原始 6 条 + 新增 2 条) - 输入图像:8 - 图像尺寸:`1280 x 704` - 每条目标帧数:320 - switch frame:160 - 视角:first-person POV