Identity-preserving video generation (IPVG) requires synthesizing videos that are faithful to both reference subjects and text prompts. Existing methods are often hindered by high tuning costs or limited input-level enhancements, struggling to maintain rigid identity consistency during complex, long-sequence actions. To address these limitations, we propose KeyID, a training-free IPVG framework that decouples the synthesis of video dynamics from the injection of identity. Specifically, KeyID comprises two components: (1) Reference-Aware Video Generation, which produces an identity-agnostic video draft aligned with multiple references, and (2) Identity-Preserved Keyframe Editing, which integrates the target identity via sparse keyframe correction and subsequent motion interpolation. By shifting from dense frame-level supervision to sparse keyframe-level refinement, KeyID effectively resolves the capacity conflict between prompt adherence and identity fidelity. Crucially, our modular design allows seamless extension to multi-subject references and complex sequential action generation without additional training. KeyID outperforms prior works and is validated by automatic and human evaluations on the official challenge benchmark, ultimately securing the runner-up position in the Track 2 (Sequential Action) of the ACM Multimedia 2026 IPVG Grand Challenge. Source code is available at https://github.com/WISLab-GDUT/KeyID.
Figures & tables
Figure 1. KeyID generates higher-quality identity-preserving videos from (a) Single Identity, (b) Multiple References, and (c) Sequential Actions. Red text highlights key attributes and actions in long instructions.
Figure 2. Overview of KeyID. (a) Reference-Aware Video Generation. A global prompt, optional temporal prompts, and a face reference are fused by GPT-5 into an enhanced prompt, from which a first frame is generated and optionally edited with non-human objects. The edited frame and the global prompt, along with optional temporal prompts, drive I2V to produce a video draft. (b) Identity Preserved Keyframe Editing. Keyframe windows are uniformly sampled in groups along the video draft, optimal keyframes are selected in keyframe identity refinement, and motion interpolation produces the final temporally coherent video with preserved identity and prompt adherence.
Method
Face-Cur ↑
Face-Arc ↑
CLIPScore ↑
Concat-ID ( Zhong et al., 2025 )
0.242
0.228
29.0
Xu et al. ( Xu et al., 2025 )
0.285
0.269
28.6
Wang et al. ( Wang et al., 2025 )
0.467
0.441
28.0
TPIGE ( Gao et al., 2025 )
0.492
0.473
27.8
KeyID (Ours)
0.633
0.630
30.9
Table 1. Comparison results of KeyID with other state-of-the-art Facial IPVG approaches on the VIP-200K test dataset.
Team
Final Score ↓
Rank
USTC-CMI
1.25
1
WislabGDUT (Ours)
1.81
2
USTC-AC
2.94
3
Table 2. Comparison results of KeyID with other state-of-the-art Sequential Action IPVG approaches on the leaderboard of IPVG 2026 Challenge (Track 2).
Figure 3. Qualitative comparison. Four evenly sampled frames are shown for each video. KeyID better preserves identity consistency and prompt adherence across complex actions. Four frames each from videos generated by KeyID and baselines, showing identity consistency across action segments.
Base
IPKE
RAVG
Face-Cur ↑
Face-Arc ↑
CLIPScore ↑
✓
0.279
0.260
28.6
✓
✓
0.596
0.594
28.3
✓
✓
✓
0.636
0.643
30.9
Table 3. Ablation study on each component in KeyID on the subset of the VIP-200K test dataset.
Configuration
Face-Cur ↑
Face-Arc ↑
CLIPScore ↑
T2V + IPKE
0.410
0.395
29.2
RAVG + DreamID-V ( Guo et al., 2026 )
0.503
0.484
30.8
RAVG + IPKE (KeyID)
0.636
0.643
30.9
Table 4. Performance comparison of KeyID and different variants on the subset of the VIP-200K test dataset.
Shanghai Jiao Tong University Shanghai, Shanghai, China · University of Electronic Science and Technology of China Chengdu, Sichuan, China · Zhejiang University Hangzhou, Zhejiang, China