Abstract
We propose Bi-AGMI (Bidirectional Attention-Gated Motion Injection), a lightweight and efficient framework for keyframe interpolation based on diffusion models. Bi-AGMI introduces a dual-path denoising process that sequentially connects forward and backward sampling trajectories via latent flipping, enabling temporally bounded generation from two keyframes. To enhance consistency between these two trajectories, we design a novel attention-gated fusion mechanism that dynamically injects and blends forward-path attention features into the backward UNet using a learnable gating module. This design improves temporal coherence, mitigates motion ambiguity, and eliminates the need for repeated re-noising. Experiments on DAVIS and Pexels datasets demonstrate that our method achieves competitive visual quality and inference efficiency compared to recent diffusion-based baselines, while requiring significantly fewer sampling steps. By enabling stable interpolation over large temporal gaps, Bi-AGMI expands the practical usability of diffusion models for long-range video completion.
ZIYUE ZENG
Ph.D. Student at Waseda University
Pursuing socially meaningful, grounded research with enduring passion and diligence.