Identity-Preserving Text-to-Video Generation Guided by Simple yet Effective Spatial-Temporal Decoupled Representations

TL;DR

The paper tackles identity-preserving text-to-video generation by addressing the inherent spatial versus temporal trade-off in end-to-end IPT2V systems. It introduces a stage-wise decoupled framework that uses a T2I stage to anchor high-fidelity spatial identities and layouts, followed by an I2V stage to model temporal dynamics, guided by semantic prompt optimization. The Spatial Semantic Parser anchors static visual semantics for the first frame, while the Temporal Prompt Polisher enhances sequential motion cues for temporal realism, enabling cross-stage semantic consistency. Evaluations on VIP-200K demonstrate improved identity preservation, text relevance, and video quality, with the method achieving runner-up in the 2025 ACM Multimedia Challenge, underscoring its practical effectiveness and efficiency. The approach offers a simple, robust blueprint for IPT2V that leverages prompt-guided disentanglement and staged generation to balance spatial structure and temporal coherence.

Abstract

Identity-preserving text-to-video (IPT2V) generation, which aims to create high-fidelity videos with consistent human identity, has become crucial for downstream applications. However, current end-to-end frameworks suffer a critical spatial-temporal trade-off: optimizing for spatially coherent layouts of key elements (e.g., character identity preservation) often compromises instruction-compliant temporal smoothness, while prioritizing dynamic realism risks disrupting the spatial coherence of visual structures. To tackle this issue, we propose a simple yet effective spatial-temporal decoupled framework that decomposes representations into spatial features for layouts and temporal features for motion dynamics. Specifically, our paper proposes a semantic prompt optimization mechanism and stage-wise decoupled generation paradigm. The former module decouples the prompt into spatial and temporal components. Aligned with the subsequent stage-wise decoupled approach, the spatial prompts guide the text-to-image (T2I) stage to generate coherent spatial features, while the temporal prompts direct the sequential image-to-video (I2V) stage to ensure motion consistency. Experimental results validate that our approach achieves excellent spatiotemporal consistency, demonstrating outstanding performance in identity preservation, text relevance, and video quality. By leveraging this simple yet robust mechanism, our algorithm secures the runner-up position in 2025 ACM MultiMedia Challenge. Our code is available at https://github.com/rain152/IPVG.

Figures (5)

• Figure 1: The pipeline of our proposed framework, comprising a spatial-temporal decoupling architecture. In the T2I stage, a pre-trained HRNet wu2021optimized generates a human-centric Reference Image via background removal, while a Spatial Semantic Parser extracts character-relevant phrases from input prompts. These are fed into HyperLoRA to produce a high-fidelity first-frame image. The I2V stage employs a Temporal Prompt Polisher to refine instructions for logical refinement and dynamic enhancement, followed by VACE-based jiang2025vace first-frame generation to synthesize temporally coherent videos.
• Figure 2: Qualitative analysis between R2V and T2I+I2V pipeline. The T2I+I2V framework demonstrates better ID preservation capability, and the video content aligns more effectively with the text. All facial images are from the VIP-200K test set vip-200k.
• Figure 3: Qualitative analysis of the ablation study. Results show that videos driven by refined prompts have better facial details and higher quality. All facial images are from the VIP-200K test set vip-200k.
• Figure :
• Figure :