Keyframe-Anchored Identity Preservation for Sequential-Action Video Generation
2026-07-20 • Computer Vision and Pattern Recognition
Computer Vision and Pattern RecognitionMultimedia
AI summaryⓘ
The authors worked on making videos from text descriptions that keep showing the same person doing a sequence of actions correctly over time. They created a method with three steps: first, changing the action descriptions into clear image prompts; second, generating key video frames that keep the person's look consistent; and third, filling in the frames between key moments while maintaining the person's identity. Their approach helps prevent the person’s appearance from changing or drifting as the video plays. Their method performed well in a competition focused on this kind of video generation.
text-to-video generationidentity preservationkeyframe generationprompt engineeringpose conditioningmulti-reference guidancenoise searchingvideo synthesis
Authors
Zhenjie Liu, Binyan Chen, Hao Chen, Tong Pan, Shangfei Wang
Abstract
Identity-preserving text-to-video generation aims to synthesize a video that accurately follows a textual description while maintaining the recognizability of a user-specified subject throughout. The IPVG26 challenge extends this framework from a single holistic prompt to a temporally structured specification. The model additionally receives a sequence of timestamped action captions and must render the subject performing these actions in the specified order. This temporal structure presents a challenge not encountered in previous identity-preserving generation tasks, as the subject must continuously perform a scripted sequence of distinct actions while maintaining a consistent identity. However, end-to-end video generators are prone to appearance drift as motion accumulates and the depicted actions change. We address this challenge with a training-free, three-stage pipeline framework. An action-aware prompt polishment stage first rewrites the inputs into image-generation prompts that specify the terminal state of each action. An identity-preserving generation stage then produces the keyframe sequence by conditioning each frame jointly on the reference identity and its predecessor, thereby decoupling time-invariant appearance from time-varying pose. Finally, an identity-aware inference enhancement stage synthesizes the intermediate segments using multi-reference guidance and identity-driven noise searching, both of which reinforce identity fidelity during sampling. Our method ranked third on the official Track 2 leaderboard, demonstrating competitive performance and strong generality.