V-Warper: Appearance-Consistent Video Diffusion Personalization via Value Warping

2025-12-13

AI Summary

Video personalization faces two key challenges: high computational cost due to reliance on large-scale video fine-tuning, and difficulty in maintaining cross-frame fine-grained appearance consistency. This paper proposes a diffusion-based personalization method that requires no video fine-tuning—only a few reference images suffice for high-fidelity identity preservation. Our core contribution is a training-free coarse-to-fine appearance adaptation framework: (1) semantic alignment via LoRA-augmented image encoding and adaptive subject embedding; and (2) precise identity injection and cross-frame representation correction through RoPE-free mid-layer attention feature matching and mask-controllable value warping—both achieved without architectural modification. Experiments demonstrate that our method achieves personalized generation quality on par with full video fine-tuning, while significantly improving appearance consistency and text-motion alignment, all at substantially reduced computational overhead.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Generation

Application Category

• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems
• Search and Retrieval-Augmented AI: Personalized, context-aware and across-device search
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Video personalization aims to generate videos that faithfully reflect a user-provided subject while following a text prompt. However, existing approaches often rely on heavy video-based finetuning or large-scale video datasets, which impose substantial computational cost and are difficult to scale. Furthermore, they still struggle to maintain fine-grained appearance consistency across frames. To address these limitations, we introduce V-Warper, a training-free coarse-to-fine personalization framework for transformer-based video diffusion models. The framework enhances fine-grained identity fidelity without requiring any additional video training. (1) A lightweight coarse appearance adaptation stage leverages only a small set of reference images, which are already required for the task. This step encodes global subject identity through image-only LoRA and subject-embedding adaptation. (2) A inference-time fine appearance injection stage refines visual fidelity by computing semantic correspondences from RoPE-free mid-layer query--key features. These correspondences guide the warping of appearance-rich value representations into semantically aligned regions of the generation process, with masking ensuring spatial reliability. V-Warper significantly improves appearance fidelity while preserving prompt alignment and motion dynamics, and it achieves these gains efficiently without large-scale video finetuning.

Problem

• Enhances video personalization with fine-grained appearance consistency
• Reduces computational cost by avoiding heavy video-based finetuning
• Maintains prompt alignment and motion dynamics in generated videos

Innovation

• Training-free coarse-to-fine personalization framework
• Lightweight coarse adaptation using image-only LoRA
• Inference-time value warping via semantic correspondences