Datasets:

Kytolly
/

FollowBench

dataset_info:
  - config_name: default
    data_files:
      - split: train
        path: train/*
      - split: test_seen
        path: test_seen/*
      - split: test_unseen
        path: test_unseen/*
tags:
  - video-generation
  - multimedia
  - benchmark
  - evaluation
  - camera-control
license: mit
language:
  - en
  - zh
task_categories:
  - video-to-video

FollowBench

Dataset Description

FollowBench is a benchmark for evaluating video generation models on viewpoint transformation and subject following tasks. It features diverse scenarios and appearances to test novel view synthesis and geometric spatial understanding capabilities.

📊 Data Splits

Split	Samples	Description
`train`	8644	Training set for model optimization and development
`test_seen`	1235	In-distribution test set for evaluating performance on familiar scenarios
`test_unseen`	1138	Out-of-distribution test set for assessing generalization to novel scenarios