Chronos: Sequential Identity-Preserving Video Generation by Temporal Anchoring

01Abstract

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 tied first place in Track 1 and first place in 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.

02Method

One temporal formulation for both tracks.

Given reference images ℐ, a raw prompt P and optional temporal captions 𝒞 = {c_t}_t=1^T, Chronos generates a video that keeps the target identity while following the scene, the actions and their order.

T = 1Track 1 · facial identity preservation, a single segment

T > 1Track 2 · sequential actions, keyframe anchors per action phase

Chronos pipeline. Inputs (reference images, raw prompt, temporal captions) flow into Semantic Prompt Anchoring (identity cue extraction and risk-aware prompt adaptation with levels light, medium, high), then Visual Frame Anchoring (cascaded keyframe anchors A1 to AT), then Metric-Guided Candidate Routing (several video generators, an automatic evaluator, weighted aggregation and exception-aware review). — **Overview of Chronos.** Semantic Prompt Anchoring refines the prompt with reliable identity cues, Visual Frame Anchoring turns conditions into keyframe anchors, and Metric-Guided Candidate Routing picks the best video from a pool of heterogeneous generators.

Stage 1 · SPA

Semantic Prompt Anchoring

A non-intrusive prompt refinement that strengthens identity constraints while keeping the original scene, style and motion.
- Identity cue extraction. A GPT model reads the reference images and keeps only stable, invariant identity cues H, dropping transient attributes such as clothing and background.
- Risk-aware prompt adaptation. It estimates the identity-drift risk of the prompt as a score s ∈ {1, 2, 3} and adds concise viewpoint and visibility cues to produce P̂.
1 Light · minimal correction 2 Medium · identity-aware adjustment 3 High · strong identity protection
Stage 2 · VFA

Visual Frame Anchoring

Text alone is ambiguous when the target pose, action or viewpoint departs from the reference. VFA uses an image-to-image operator Φ_I2I to turn ℐ, P̂ and an instruction template into visually grounded keyframes.
- Track 1. A single high-fidelity first-frame anchor A₁ initializes image-to-video generation.
- Track 2. Anchor A₁ comes from the first caption c₁; each later anchor A_t is generated from c_t and the previous keyframes, in a cascade.
A₁A₂⋯A_T
Stage 3 · MGCR

Metric-Guided Candidate Routing

Generators G₁…G_K each turn (P̂, 𝒜) into a candidate video. An automatic evaluator scores every candidate on text consistency, identity, subject preservation, motion, visual quality and an LLM-based holistic judgment.

R(V_k) = Σ_j λ_j s_j(V_k)

V^* = arg max_{V_k ∈ 𝒱} R(V_k)
- Exception-aware review. If the winner falls below a threshold τ_j on any critical metric, it goes to human review instead of being accepted on its total score alone.

Chronos is training-free: no model is fine-tuned. The generator pool uses Wan2.2 I2V-14B for both tracks, plus Seedance for Track 1 and Wan FLF2V for Track 2; SPA and VFA use GPT-based models.

03Visual comparison

Same reference and prompt for every method. Clips play at their native frame rate, synced to the Chronos clip on one shared timeline of input captions.

04Results

Track 1 · Challenge

Tied 1st

Final Score 2.5, tied with MIPL_Video (lower is better)

Track 2 · Challenge

1st

Final Score 1.25, ahead of WislabGDUT (1.8125; lower is better)

Track 1 · vs. baselines

0.6769

Total Score, +8.9% over the best baseline, RefAlign (0.6214)

Track 2 · vs. baselines

0.6102

Total Score, +6.6% over the best baseline, Phantom (0.5725)

Challenge leaderboard · top three teams · Final Score, lower is better

Track 1 · Facial Identity-Preserving
Team	Final Score ↓	Rank
USTC-CMI (Chronos)	2.5	1
MIPL_Video	2.5	1
xuanyuan_fuxi	3	3

Track 2 · Sequential Action Identity-Preserving
Team	Final Score ↓	Rank
USTC-CMI (Chronos)	1.25	1
WislabGDUT	1.8125	2
USTC-AC	2.9375	3

Comparison with existing methods

Bold marks the best and italics the second-best value in each column. RefAlign and Phantom keep a higher Subject Score (0.80 and 0.72 vs. 0.68) but move less (Motion Amp. 0.29 and 0.21 vs. 0.43); Stand-In moves the most (0.46) but loses the face (Subject 0.24). Chronos balances the two and leads on GMEScore, Aesthetic, Smoothness, Natural and Total.

Ablation study

SPA alone gives a small identity gain on Track 1 (Subject 0.751 → 0.756) and raises Natural Score (0.578 → 0.588).
VFA keyframe anchors with an image-to-video generator lift Track 1 Aesthetic (0.50 → 0.58–0.60) and Motion Amp. (0.18 → 0.37–0.50) at some cost in Subject. On Track 2, SPA and VFA together raise Subject from 0.458 to 0.534–0.591.
Routing adds +0.019 (Track 1) and +0.011 (Track 2) Total Score over the best single generator, so full Chronos has the best Total on both tracks.

Bold marks the best and italics the second-best value in each column. Ablations use VACE as the baseline and run on the first 100 cases, a representative subset whose trends match the full evaluation set, so absolute numbers differ from the comparison table.

05BibTeX

@misc{huang2026chronossequentialidentitypreservingvideo,
  title={Chronos: Sequential Identity-Preserving Video Generation by Temporal Anchoring},
  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},
  year={2026},
  eprint={2611.05893},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2611.05893},
}

BibTeX will be available here once the arXiv version is out.

Chronos Sequential Identity-Preserving Video Generation by Temporal Anchoring