VRM-10K

VRM-10K is the video referring matting dataset introduced in VRMDiff: Text-Guided Video Referring Matting Generation of Diffusion. Each sample is a video with one or more people. Every person has a text caption that refers to them and a per-frame alpha matte.

Split	Videos
train	9,000
test	1,000

Data fields

Each row is one video (120 frames, 30 fps).

• video_id: video index within the split.
• video: the video (MP4).
• width, height: video resolution.
• background: source_file and start_frame of the background clip (from DVM).
• objects: the people in the video, listed in compositing order. Each entry has:
  • object_id: ID of the foreground clip (from VideoMatte240K).
  • caption: text describing this person.
  • layer: compositing order, where 0 is the bottom layer.
  • source_start_frame: first frame used from the foreground clip.
  • scale, x, y, box_width, box_height: resize factor and placement of the foreground clip in the video.
  • mattes: 120 alpha mattes, one per frame. Each is an 8-bit grayscale PNG at the video's resolution.

Usage

from datasets import load_dataset, Video

ds = load_dataset("xsourse/VRM-10K", split="test", streaming=True)
ds = ds.cast_column("video", Video(decode=False))  # raw MP4 bytes; drop this line to decode with torchcodec

sample = next(iter(ds))
mp4_bytes = sample["video"]["bytes"]
for obj in sample["objects"]:
    caption = obj["caption"]
    mattes = obj["mattes"]  # list of 120 PIL images

Source data

Foreground videos and alpha mattes come from VideoMatte240K (footage provided by Storyblocks). Background videos come from DVM. Captions were generated with Tarsier-34B. Please also cite these works when you use this dataset.

Citation

@article{yang2025vrmdiff,
  title   = {VRMDiff: Text-Guided Video Referring Matting Generation of Diffusion},
  author  = {Yang, Lehan and Song, Jincen and Wang, Tianlong and Qi, Daiqing and Shi, Weili and Liu, Yuheng and Li, Sheng},
  journal = {arXiv preprint arXiv:2503.10678},
  year    = {2025}
}