Research map / Video / Video generation
Local Consistency Guidance: Personalized Stylization Method of Face Video (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Local Consistency Guidance (LCG), combining local cross-frame attention and inter-frame example-based style transfer on a personalized DreamBooth+ControlNet diffusion model, to stylize face videos with temporal consistency.
Problem
Face video stylization into reference styles while keeping continuity between frames and preserving original facial expressions; one-shot and GAN-based methods struggle with video consistency.
Approach
Replaces U-Net self-attention with a local cross-frame attention window whose weights decay from a moving center node, and adds inter-framed example-based style transfer (Ebsynth-style) inside iterative DDIM denoising to enforce detail-level consistency.
Key results
- Frame Consistency of 0.97 for the proposed method vs 0.83 for Dreambooth+ControlNet and 0.91 for the source video (Table 1).
- Qualitative comparison against Ebsynth and JOJOGan on VFHQ videos shows better facial feature and makeup capture and expression control (visual only).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Local Consistency Guidance (local cross-attention and inter-framed style transfer) | Personalized diffusion model with DreamBooth+ControlNet · VFHQ | Frame Consistency (%): 0.83 -> 0.97 | Helped |
Limitations
- Evaluation is limited to qualitative visualizations plus a single frame-consistency metric; no expression-preservation or style-fidelity measures reported.
- Method description is high-level, with no hyperparameters, window size, or training details given.
- Frame consistency score for the source video (0.91) is higher than Dreambooth+ControlNet, but the Ebsynth comparison is only visual.
Experimental setup
Datasets: VFHQ
Benchmarks: VFHQ
Metrics: Frame Consistency (%)
Baselines: Ebsynth JOJOGan Dreambooth+Controlnet
Builds on: DreamBooth ControlNet Ebsynth zero-shot video editing
Closest papers by meaning
- StyleMaster: Stylize Your Video with Artistic Generation and Translation (CVPR 2025 · similarity 0.85)
- StyO: Stylize Your Face in Only One-Shot (AAAI 2025 · similarity 0.85)
- ACFun: Abstract-Concrete Fusion Facial Stylization (NeurIPS 2024 · similarity 0.84)
- RealPortrait: Realistic Portrait Animation with Diffusion Transformers (AAAI 2025 · similarity 0.84)
- MimicTalk: Mimicking a personalized and expressive 3D talking face in minutes (NeurIPS 2024 · similarity 0.84)
- OmniConsistency: Learning Style-Agnostic Consistency from Paired Stylization Data (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.