Challenge Overview
Recent advances in multimodal Artificial Intelligence-Generated Content have transformed digital content creation. Modern text-to-video systems can synthesize visually compelling videos directly from textual prompts, but the appearance of key subjects may still change across frames and scenes. Facial features can drift over time, object attributes can degrade during long sequences, and identity can be lost during complex motion.
The Identity-Preserving Video Generation (IPVG) task addresses this limitation by maintaining the consistency of a given reference identity throughout the text-to-video generation process. This capability is important for applications such as personalized avatar animation, historical figure reenactment, and brand-specific content creation, where identity deviations can compromise usability or authenticity.
The challenge brings together the research community around identity-preserving video generation datasets and controllable generative models that support precise identity binding. It contains two tracks:
The VIP-200K and ReactID-Data datasets can only be used for research purposes.
Task Description
The challenge focuses on two identity-preserving video generation tasks:
Facial Identity-Preserving Text-to-Video Generation
Given videos, corresponding prompts, and reference facial identity images, the goal is to learn a video generation model that synthesizes temporally consistent videos. The generated videos must align semantically with the prompts and remain faithful to the identity constraints derived from the reference facial images.
Sequential Action Identity-Preserving Video Generation
Given reference images and structured timeline prompts containing multiple timestamped sub-actions, the goal is to synthesize videos in which the subject performs the specified sequence while preserving strict visual consistency. The subject may be a full-body human, an animal, or a specific object.
Contestants must develop identity-preserving video generation systems using the provided datasets. For evaluation, a system must generate at least one video for every <reference identity, prompt> pair in the corresponding testing set.
Datasets
The following datasets are provided for the two challenge tracks:
Track 1 — Facial Identity-Preserving Text-to-Video Generation
Training Dataset
VIP-200K contains 500,000 videos. Each video is coupled with a textual prompt and one or more identity images. Each identity is defined by one or more video frames with bounding boxes around the face.
Access VIP-200K on Hugging FaceTesting Dataset
The testing set contains 200 person identities that are unseen in the training data. Each identity is represented by one or more portrait images and five textual prompts, resulting in 1,000 identity-image and prompt pairs for testing.
Track 2 — Sequential Action Identity-Preserving Video Generation
Training Dataset
ReactID-Data is a large-scale dataset constructed with a high-precision pipeline for reliable subject-to-video correspondence. It contains subject-to-video pairs with timeline annotations describing sub-actions and their timestamps. The data is filtered for aesthetic quality and precise subject alignment.
Access ReactID-Data on Hugging FaceTesting Dataset
The testing set contains unseen identities paired with complex multi-action timeline prompts, such as 0-2s: [Action A]; 2-5s: [Action B], to evaluate sequential dynamics and identity preservation simultaneously.
Submission Format
For each track, each team may submit the results of at most three runs. The team must select one run as its primary run for performance comparison across teams.
Submit Up to Three Runs
A team may submit no more than three runs for each challenge track.
Select the Primary Run
One run must be designated as the primary submission. Evaluation of the second and third runs is not guaranteed.
Evaluation Metric
For both tracks, generated videos will be evaluated using a combination of objective metrics and human studies. The evaluation focuses on the following dimensions:
Identity Preservation
Feature similarity between the generated video and the reference identity image, together with manual identity-preservation scores from multiple annotators.
Video Quality
Visual quality, motion dynamics, and text alignment, assessed by objective scoring models and subjective evaluation from multiple annotators.
Action-Timeline Alignment (Track 2)
Text-to-video alignment and temporal consistency, measuring how accurately generated actions match the timestamps specified by the timeline prompts.
The final score for each submission will be derived from a weighted combination of objective evaluation results and a subjective user study conducted on the testing set.