Sign in to view source links and access this dataset
Description
PexelsCustom-1M is a large-scale dataset for customized video generation, containing 1,036,431 curated (identity, text, video) triplets. It spans 8,373 identity categories derived from approximately 320,000 Pexels HD videos. The dataset was created by carpedkm and is associated with research presented at ICASSP 2026.
Use Cases
Training text-to-video models based on the provided (text, video) pairs.
Developing identity-consistent video generation systems based on the curated identity categories.
Benchmarking customized video generation models using the large-scale, publicly available triplets.
Researching multimodal alignment between text descriptions and video content.
Strengths
Large scale with 1,036,431 training triplets.
Curated across 8,373 distinct identity categories.
Sourced from a substantial base of approximately 320,000 high-definition videos.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
Source
Pexels HD videos.
Collection Method
Curated (identity, text, video) triplets extracted from source videos.
Freshness
Last updated 2026-06-09 08:45:20; freshness should be verified.
License is unknown and should be verified before use.