README.md

TGI-Bench

TGI-Bench is a benchmark for text-conditioned generative inbetweening.

It provides image sequences together with text annotations for evaluating whether a model can generate intermediate frames that are both temporally plausible and aligned with natural-language descriptions.

The dataset includes four annotation files for different sequence lengths:

• 25_dataset_caption.json
• 33_dataset_caption.json
• 65_dataset_caption.json
• 81_dataset_caption.json

Each sample is stored in its own folder, and each JSON file contains annotations for the corresponding sequence length.

Dataset Structure

TGI-Dataset/
├── aerobatics/
│   ├── 00000.jpg
│   ├── 00001.jpg
│   └── ...
├── air_conditioner/
│   ├── 00000.jpg
│   ├── 00001.jpg
│   └── ...
├── ...
├── 25_dataset_caption.json
├── 33_dataset_caption.json
├── 65_dataset_caption.json
└── 81_dataset_caption.json

Annotation Format

Each JSON file is a list of entries like this:

{
  "folder": "aerobatics",
  "first_image_desc": "Man in cockpit of glider flying over patchwork fields.",
  "last_image_desc": "Man in cockpit of glider banking over fields and clouds.",
  "challenge": "Large motion",
  "caption": "Glider floats over fields with man holding control stick."
}

Fields

• folder: folder name containing the image sequence
• first_image_desc: description of the first frame
• last_image_desc: description of the last frame
• challenge: challenge category
• caption: text description of the sequence

Usage

This dataset can be used for research on:

• text-conditioned generative inbetweening
• video frame interpolation
• vision-language evaluation
• text-video alignment