Micro-World-Capsaicin

This dataset is part of the work Micro-World V2. Micro-World-Capsaicin is a synthetic video dataset rendered with the open-source AMD Capsaicin framework. It pairs rendered videos with English scene descriptions and frame-aligned control traces, making it suitable for research on action-conditioned video generation, world models, temporal prediction, and embodied-agent representation learning.

The dataset contains two sequence lengths, 200 frames and 400 frames. Every trajectory is provided in two video resolutions: a 360x640 resolution version under video/ and a 4K version under 4k_videos/. Files with the same basename represent the same trajectory and control sequence.

Dataset summary

Group	Unique trajectories	Frames per trajectory	Approximate size
200f	440	200	52.4 GB
400f	374	400	88.8 GB
Total	814	—	141 GB

The counts and sizes above describe the current main revision.

Repository structure

Micro-World-Capsaicin/
├── 200f/
│   ├── video/                  # 360x640 Standard-resolution MP4 files
│   ├── 4k_videos/             # 4K MP4 files
│   ├── metadata-detection/     # Frame-aligned control JSON files
│   └── video_captions_v0.csv   # Video descriptions and frame ranges
└── 400f/
    ├── video/
    ├── 4k_videos/
    ├── metadata-detection/
    └── video_captions_v0.csv

Within a frame group, the MP4 and JSON files are aligned by basename. For example:

200f/video/allow_gallery_pitch_allow_a000_v00.mp4
200f/4k_videos/allow_gallery_pitch_allow_a000_v00.mp4
200f/metadata-detection/allow_gallery_pitch_allow_a000_v00.json

Data fields

Caption CSV

Each video_captions_v0.csv file contains one row per unique trajectory:

Column	Description
original video name	MP4 filename used to locate both resolution variants and the corresponding JSON file.
prompt	Description of the rendered scene and its visible content.
start frame index	Inclusive first frame index.
end frame index	Inclusive final frame index.

Control JSON

Each file in metadata-detection/ contains normalization factors and a dictionary of actions indexed by frame number. A shortened example is shown below:

{
  "mouse_normalized_to_mc": {
    "yaw_factor": 0.06666666666666667,
    "pitch_factor": 0.06666666666666667
  },
  "actions": {
    "0": {
      "ws": 1,
      "ad": 0,
      "scs": 0,
      "pitch_delta": 0.0,
      "yaw_delta": 0.0,
      "collision": 0,
      "jump_invalid": 0
    }
  }
}

The action entries include encoded movement state (ws, ad, and scs), camera pitch and yaw changes, and collision or invalid-jump indicators. These values are pipeline-specific encodings; consumers should preserve the values as recorded and use the mapping supplied with the project code when it becomes available.