Research map / Video / Video generation
VidProM: A Million-scale Real Prompt-Gallery Dataset for Text-to-Video Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces VidProM, a 1.67M-unique real-user text-to-video prompt dataset with 6.69M videos from four diffusion models, and compares it to DiffusionDB and outlines research directions.
Problem
There is no public dataset of real user prompts for text-to-video diffusion models, so the prompt distribution of T2V users and its implications for model development, efficiency, and safety are unstudied.
Approach
Scrapes real user prompts from Pika Discord channels, embeds them with text-embedding-3-large, generates videos with three open T2V models plus scraped Pika videos, and adds NSFW scores and a semantic-uniqueness subset (cosine similarity below 0.8). It also fine-tunes Mistral-7B as an automatic prompt completion model.
Key results
- VidProM contains 1,672,243 unique prompts and 6.69M generated videos (about 50,631 V100 GPU hours), with 1,038,805 semantically unique prompts versus 739,010 in DiffusionDB.
- Existing fake image detectors largely fail on fake video frames; average accuracy ranges from 45.32% (LNP) to 63.64% (Gram-Net), with near-chance results for most methods.
- Only about 2% of Text2Video-Zero videos exceed a normalized copy-detection score of 0, and Pika shows the lowest replication rate.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply existing fake-image detectors to video frames | Fake video detection · Fake video detection | accuracy (%) | Hurt |
| Use detectors relying on traditional cues, such as global texture statistics or frequency analysis | Fake image detection methods evaluated on fake videos · Fake video detection | accuracy (%) | Helped |
Surprising findings
- Most fake image detectors perform near chance on generated video frames, with frequency and texture-statistic methods performing best.
- Less than 0.5% of prompts have a NSFW probability above 0.2, despite no explicit filtering of NSFW prompts in the source channels.
Limitations
- Generated videos are short (1.6 to 3 seconds) and not of the highest quality.
- Prompts come only from Pika Discord channels, a single source, with Pika-specific bias.
- Fake image detection and copy detection experiments are preliminary and use only middle frames.
Experimental setup
Datasets: VidProM DiffusionDB Panda-70M DVSC2023 WebVid-10M HDVILA-100M
Benchmarks: VidProM (fake video detection) DVSC2023
Metrics: accuracy mAP normalized copy-detection score
Baselines: DiffusionDB CNNSpot FreDect Fusing Gram-Net LGrad LNP DIRE UnivFD
Builds on: DiffusionDB Panda-70M FCPL
Closest papers by meaning
- VideoUFO: A Million-Scale User-Focused Dataset for Text-to-Video Generation (NeurIPS 2025 · similarity 0.89)
- OpenVid-1M: A Large-Scale High-Quality Dataset for Text-to-video Generation (ICLR 2025 · similarity 0.89)
- Prompt-Free Diffusion: Taking "Text" out of Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.89)
- EIDT-V: Exploiting Intersections in Diffusion Trajectories for Model-Agnostic, Zero-Shot, Training-Free Text-to-Video Generation (CVPR 2025 · similarity 0.89)
- ShotAdapter: Text-to-Multi-Shot Video Generation with Diffusion Models (CVPR 2025 · similarity 0.89)
- VideoGuide: Improving Video Diffusion Models without Training Through a Teacher's Guide (CVPR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.