Hugging Face's logo

Datasets:
WenhaoWang
/
TIP-I2V

Tasks:
Image-to-Video
Text-to-Video
Text-to-Image
Modalities:
Image
Text
Formats:
parquet
Languages:
English
Size:
1M - 10M
ArXiv:
Tags:
prompt
image-to-video
text-to-video
visual-generation
video-generation
Libraries:
Datasets
Dask
License:
Dataset card Data Studio Files
xet
Community
2
TIP-I2V / README.md
WenhaoWang's picture
WenhaoWang
Enhance dataset card: Add project page, abstract, and `library_name` metadata (#2)
958f51d verified
preview code
|
raw
history blame
10.5 kB
metadata
language:
  - en
license: cc-by-nc-4.0
size_categories:
  - 1M<n<10M
task_categories:
  - image-to-video
  - text-to-video
  - text-to-image
  - image-to-image
pretty_name: TIP-I2V
library_name: datasets
dataset_info:
  features:
    - name: UUID
      dtype: string
    - name: Text_Prompt
      dtype: string
    - name: Image_Prompt
      dtype: image
    - name: Subject
      dtype: string
    - name: Timestamp
      dtype: string
    - name: Text_NSFW
      dtype: float32
    - name: Image_NSFW
      dtype: string
  splits:
    - name: Full
      num_bytes: 13440652664.125
      num_examples: 1701935
    - name: Subset
      num_bytes: 790710630
      num_examples: 100000
    - name: Eval
      num_bytes: 78258893
      num_examples: 10000
  download_size: 27500759907
  dataset_size: 27750274851.25
configs:
  - config_name: default
    data_files:
      - split: Full
        path: data/Full-*
      - split: Subset
        path: data/Subset-*
      - split: Eval
        path: data/Eval-*
tags:
  - prompt
  - image-to-video
  - text-to-video
  - visual-generation
  - video-generation

Summary

This is the dataset proposed in our paper TIP-I2V: A Million-Scale Real Text and Image Prompt Dataset for Image-to-Video Generation.

Project page | Paper

TIP-I2V is the first dataset comprising over 1.70 million unique user-provided text and image prompts. Besides the prompts, TIP-I2V also includes videos generated by five state-of-the-art image-to-video models (Pika, Stable Video Diffusion, Open-Sora, I2VGen-XL, and CogVideoX-5B). The TIP-I2V contributes to the development of better and safer image-to-video models.

Abstract

Video generation models are revolutionizing content creation, with image-to-video models drawing increasing attention due to their enhanced controllability, visual consistency, and practical applications. However, despite their popularity, these models rely on user-provided text and image prompts, and there is currently no dedicated dataset for studying these prompts. In this paper, we introduce TIP-I2V, the first large-scale dataset of over 1.70 million unique user-provided Text and Image Prompts specifically for Image-to-Video generation. Additionally, we provide the corresponding generated videos from five state-of-the-art image-to-video models. We begin by outlining the time-consuming and costly process of curating this large-scale dataset. Next, we compare TIP-I2V to two popular prompt datasets, VidProM (text-to-video) and DiffusionDB (text-to-image), highlighting differences in both basic and semantic information. This dataset enables advancements in image-to-video research. For instance, to develop better models, researchers can use the prompts in TIP-I2V to analyze user preferences and evaluate the multi-dimensional performance of their trained models; and to enhance model safety, they may focus on addressing the misinformation issue caused by image-to-video models. The new research inspired by TIP-I2V and the differences with existing datasets emphasize the importance of a specialized image-to-video prompt dataset. The project is available at this https URL .

Datapoint

Statistics

Download

For users in mainland China, try setting export HF_ENDPOINT=https://hf-mirror.com to successfully download the datasets.

Download the text and (compressed) image prompts with related information 

# Full (text and compressed image) prompts: ~13.4G
from datasets import load_dataset
ds = load_dataset("WenhaoWang/TIP-I2V", split='Full', streaming=True)

# Convert to Pandas format (it may be slow)
import pandas as pd
df = pd.DataFrame(ds)

# 100k subset (text and compressed image) prompts: ~0.8G
from datasets import load_dataset
ds = load_dataset("WenhaoWang/TIP-I2V", split='Subset', streaming=True)

# Convert to Pandas format (it may be slow)
import pandas as pd
df = pd.DataFrame(ds)

# 10k TIP-Eval (text and compressed image) prompts: ~0.08G
from datasets import load_dataset
ds = load_dataset("WenhaoWang/TIP-I2V", split='Eval', streaming=True)

# Convert to Pandas format (it may be slow)
import pandas as pd
df = pd.DataFrame(ds)

Download the embeddings for text and image prompts 

# Embeddings for full text prompts (~21G) and image prompts (~3.5G)
from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="Embedding/Full_Text_Embedding.parquet", repo_type="dataset")
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="Embedding/Full_Image_Embedding.parquet", repo_type="dataset")

# Embeddings for 100k subset text prompts (~1.2G) and image prompts (~0.2G)
from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="Embedding/Subset_Text_Embedding.parquet", repo_type="dataset")
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="Embedding/Subset_Image_Embedding.parquet", repo_type="dataset")

# Embeddings for 10k TIP-Eval text prompts (~0.1G) and image prompts (~0.02G)
from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="Embedding/Eval_Text_Embedding.parquet", repo_type="dataset")
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="Embedding/Eval_Image_Embedding.parquet", repo_type="dataset")

Download uncompressed image prompts 

# Full uncompressed image prompts: ~1T
from huggingface_hub import hf_hub_download
for i in range(1,52):
    hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="image_prompt_tar/image_prompt_%d.tar"%i, repo_type="dataset")

# 100k subset uncompressed image prompts: ~69.6G
from huggingface_hub import hf_hub_download
for i in range(1,3):
    hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="sub_image_prompt_tar/sub_image_prompt_%d.tar"%i, repo_type="dataset")

# 10k TIP-Eval uncompressed image prompts: ~6.5G
from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="eval_image_prompt_tar/eval_image_prompt.tar", repo_type="dataset")

Download generated videos 

# Full videos generated by Pika: ~1T
from huggingface_hub import hf_hub_download
for i in range(1,52):
    hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="pika_videos_tar/pika_videos_%d.tar"%i, repo_type="dataset")

# 100k subset videos generated by Pika (~57.6G), Stable Video Diffusion (~38.9G), Open-Sora (~47.2G), I2VGen-XL (~54.4G), and CogVideoX-5B (~36.7G)
from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="subset_videos_tar/pika_videos_subset_1.tar", repo_type="dataset")
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="subset_videos_tar/pika_videos_subset_2.tar", repo_type="dataset")
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="subset_videos_tar/svd_videos_subset.tar", repo_type="dataset")
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="subset_videos_tar/opensora_videos_subset.tar", repo_type="dataset")
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="subset_videos_tar/i2vgenxl_videos_subset_1.tar", repo_type="dataset")
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="subset_videos_tar/i2vgenxl_videos_subset_2.tar", repo_type="dataset")
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="subset_videos_tar/cog_videos_subset.tar", repo_type="dataset")

# 10k TIP-Eval videos generated by Pika (~5.8G), Stable Video Diffusion (~3.9G), Open-Sora (~4.7G), I2VGen-XL (~5.4G), and CogVideoX-5B (~3.6G)
from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="eval_videos_tar/pika_videos_eval.tar", repo_type="dataset")
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="eval_videos_tar/svd_videos_eval.tar", repo_type="dataset")
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="eval_videos_tar/opensora_videos_eval.tar", repo_type="dataset")
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="eval_videos_tar/i2vgenxl_videos_eval.tar", repo_type="dataset")
hf_hub_download(repo_id="WenhaoWang/TIP-I2V", filename="eval_videos_tar/cog_videos_eval.tar", repo_type="dataset")

Comparison with VidProM and DiffusionDB

Click the WizMap (TIP-I2V VS VidProM) and WizMap (TIP-I2V VS DiffusionDB) (wait for 5 seconds) for an interactive visualization of our 1.70 million prompts.

License

The prompts and videos in our TIP-I2V are licensed under the CC BY-NC 4.0 license.

Curators

TIP-I2V is created by Wenhao Wang and Professor Yi Yang. We retain the Pika logos in our release to (1) acknowledge its contribution and (2) comply with its attribution requirements.

Citation 

@article{wang2024tipi2v,
  title={TIP-I2V: A Million-Scale Real Text and Image Prompt Dataset for Image-to-Video Generation},
  author={Wang, Wenhao and Yang, Yi},
  booktitle={arXiv preprint arXiv:2411.04709},
  year={2024}
}

Contact

If you have any questions, feel free to contact Wenhao Wang ([email protected]).