Adaptive Interpolation-Synthesis for Motion In-Betweening on Keyframe-Based Animation

Problem
This preprint addresses the gap in capability for motion in-betweening in keyframe-based animation, a process that is both artistically demanding and time-consuming. Existing deep learning methods for motion synthesis often misalign with the practical constraints and workflows of professional animators, leading to inefficiencies and a lack of creative control. The authors aim to provide a solution that integrates seamlessly into production environments, thereby enhancing the animator’s workflow.

Method
The core technical contribution is the Adaptive Interpolation-Synthesis (AIS) layer, which dynamically balances learned interpolation with direct pose synthesis. This architecture is designed to reflect the animator’s creative process, allowing for more intuitive control over the in-betweening task. The method employs a domain-based input keypose schedule that captures the distribution of production data, ensuring stylistic consistency and better alignment with real-world animation practices. The training process leverages a dataset reflective of professional animation workflows, although specific details regarding the dataset size, training compute, and loss functions are not disclosed.

Results
The proposed method achieves state-of-the-art performance on production data, demonstrating a 3.5x speedup in completing in-betweening tasks when integrated into Autodesk Maya, compared to traditional methods. While specific quantitative metrics against named baselines are not provided in the abstract, the qualitative improvements in stylistic consistency and animator satisfaction are emphasized, suggesting significant effect sizes in practical applications.

Limitations
The authors acknowledge that their method may still be limited by the diversity of motion styles present in the training data, which could affect generalization to less common animation styles. Additionally, the reliance on Autodesk Maya for integration may limit accessibility for animators using other software. The paper does not discuss potential computational overhead introduced by the AIS layer or the scalability of the method across different animation projects.

Why it matters
This work has significant implications for the field of computer graphics and animation, particularly in enhancing the efficiency of the animation pipeline. By aligning deep learning techniques with the creative processes of animators, the AIS method not only reduces production time but also preserves artistic intent, potentially leading to broader adoption of AI tools in animation. This could pave the way for further research into adaptive systems that cater to the nuanced requirements of creative professionals across various domains.

Authors: Anton Raël, Julien Boucher, Antoine Lhermitte
Source: arXiv:2605.02742
URL: https://arxiv.org/abs/2605.02742v1