Chronos: Sequential Identity-Preserving Video Generation by Temporal Anchoring

Mengqi Huang, Wenchuan Wang, Zhendong Mao, Yijing Tu, Zhike Liu, Luoni Wang, Ning Tong, Xiaoxiao Sheng, Tianyu Huang, Xuecheng Qi, Yongdong ZhangarXiv 2026 · arXiv

TL;DR

Chronos is a training-free framework that solves the persistent problem of "identity drift" in AI-generated videos by using a novel temporal anchoring strategy. By combining LLM-based prompt refinement with the generation of visual keyframe anchors, it maintains high subject fidelity across complex, multi-stage actions, securing first place in both tracks of the ACM MM 2026 Identity-Preserving Video Generation Challenge.

What Problem Does This Paper Address?

While modern video generation models (like Sora or Gen-3) produce stunning visuals, they often struggle with identity preservation. If you ask an AI to generate a video of a specific person performing a sequence of actions, the person’s face often morphs or "drifts" as they move.

Existing methods typically face a "fundamental tension":

  1. Strong constraints keep the face consistent but result in stiff, robotic motion.
  2. Weak constraints allow for fluid motion but cause the subject's identity to dissolve, especially during large pose variations or multi-step sequences (e.g., "a woman sits down, then opens a box, then smiles").

Chronos tackles this by unifying single-segment and multi-segment generation under a common temporal anchoring formulation, ensuring the subject looks like themselves from the first frame to the last.

Methods

Chronos is a "training-free" framework, meaning it doesn't require expensive fine-tuning of base models. Instead, it orchestrates existing models through three key stages:

  1. Semantic Prompt Anchoring (SPA): Using a GPT-based model, Chronos extracts "invariant biometric cues" from reference images—ignoring transient details like clothing or background. It then adaptively refines the user's prompt based on a "risk score," adding specific visibility and viewpoint instructions to help the generator maintain the face.
  2. Visual Frame Anchoring (VFA): Text prompts are often ambiguous. VFA uses an Image-to-Image (I2I) operator to create high-fidelity "anchor frames." For complex sequences, it generates these anchors in a cascaded manner, providing a visual "north star" for each phase of the action.
  3. Metric-Guided Candidate Routing (MGCR): Recognizing that no single model is perfect, Chronos generates a pool of candidate videos using different backbones (like Wan2.1 or Seedance). An automated evaluator then scores these candidates across dimensions like identity consistency, motion amplitude, and aesthetic quality to "route" the best version to the user.
The Chronos pipeline: from reference images and prompts to Semantic Prompt Anchoring, Visual Frame Anchoring, and final Metric-Guided Routing.

Datasets

The authors evaluated Chronos using the benchmarks provided by the ACM MM 2026 Identity-Preserving Video Generation (IPVG) Challenge:

  • Track 1 (Facial Identity): Focused on maintaining a single human face across diverse motions.
  • Track 2 (Sequential Action): Focused on multi-subject interactions and multi-stage actions (e.g., two cats interacting over several seconds).

Evaluation relied on a suite of automated metrics including FaceSim-Cur (for human ID), YOLO-World crop similarity (for non-human subjects), and GMEScore (for text alignment).

Key Insights and Ideas

  • Identity vs. Context Decoupling: A major "aha" moment is the realization that identity preservation improves when you explicitly tell the LLM to ignore the subject's clothes and the room's decor, focusing strictly on permanent facial features during the prompt refinement stage.
  • The Power of the Visual Anchor: Moving from "text-to-video" to "anchor-frame-to-video" significantly reduces the search space for the generative model, preventing the "hallucinations" that typically cause identity drift.
  • Ensemble Intelligence: By using MGCR, the framework acknowledges that different generative architectures have complementary strengths. One model might be better at fluid motion, while another excels at skin textures; Chronos picks the winner for each specific prompt.

Key Results

Chronos achieved first place in both tracks of the ACM MM 2026 Challenge.

  • Track 1: It outperformed the runner-up (RefAlign) by 9.0% in Total Score, striking a better balance between facial fidelity and natural motion smoothness.
  • Track 2: It improved the Total Score by 6.6% over the runner-up (Phantom), specifically excelling in multi-reference consistency for complex scenes.
  • Qualitative Superiority: As shown in Figure 3, while baseline models often failed to follow specific instructions (like "placing a golf ball") or lost the subject's likeness, Chronos maintained high identity fidelity while accurately executing sequential motions.
Qualitative comparison showing Chronos maintaining identity and action accuracy better than baselines like Stand-In, RefAlign, and Phantom.

Limitations and Future Work

  • Inference Latency: Because the framework generates multiple candidates and runs an automated evaluation suite (MGCR), the time required to produce a final video is likely higher than a single-pass model.
  • LLM Dependency: The effectiveness of the Semantic Prompt Anchoring is tied to the reasoning capabilities of the underlying LLM.
  • Human-in-the-loop: The authors include an "exception-aware review mechanism," suggesting that automated metrics still occasionally miss subtle artifacts that a human eye would catch. Future work may involve tighter integration of these metrics into the diffusion process itself.

Summary updated

Abstract

Identity-preserving video generation aims to synthesize high-quality and visually realistic videos from textual and visual conditions while faithfully maintaining the subject's identity across diverse motions. Prior methods have extensively introduced reference-guided control signals and identity-aware conditioning to improve facial alignment from reference images. However, improving static identity alignment alone creates a fundamental tension between identity fidelity and temporal controllability, which frequently causes identity drift or motion suppression under large pose variations and multi-segment sequential actions. To address this, we propose Chronos, a framework that unifies single-segment facial identity preservation and multi-segment sequential action generation under a common temporal formulation. Specifically, Chronos integrates Semantic Prompt Anchoring to adaptively refine the user prompt, Visual Frame Anchoring to transform conditions into visually grounded keyframe anchors for visual guidance, and Metric-Guided Candidate Routing to select the highest-quality videos from heterogeneous generators via multi-dimensional evaluation. Extensive evaluations demonstrate the superiority of our approach, which achieved first place in both Track 1 and Track 2 of the ACM MM 2026 Identity-Preserving Video Generation Challenge, demonstrating its effectiveness across both facial identity preservation and sequential action generation settings.
BibTeX
@misc{huang2026chronos,
  author = {Mengqi Huang and Wenchuan Wang and Zhendong Mao and Yijing Tu and Zhike Liu and Luoni Wang and Ning Tong and Xiaoxiao Sheng and Tianyu Huang and Xuecheng Qi and Yongdong Zhang},
  title = {{Chronos: Sequential Identity-Preserving Video Generation by Temporal Anchoring}},
  year = {2026},
  eprint = {2611.05893},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.05893},
}
Figure