![[2602.21133] SOM-VQ: Topology-Aware Tokenization for Interactive Generative Models](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.21133] SOM-VQ: Topology-Aware Tokenization for Interactive Generative Models
Summary
The paper presents SOM-VQ, a novel tokenization method that enhances interactive generative models by integrating vector quantization with Self-Organizing Maps, allowing for topology-aware control in generative tasks.
Why It Matters
SOM-VQ addresses the limitations of traditional vector quantization in generative models by introducing a topology-aware approach that enhances semantic understanding and human control. This is particularly relevant for applications in human motion generation, music, and other interactive domains, making generative models more intuitive and effective.
Key Takeaways
- SOM-VQ combines vector quantization with Self-Organizing Maps for better tokenization.
- The method preserves semantic structure, enabling intuitive human control.
- Demonstrated effectiveness in human motion generation and other interactive applications.
- Allows for geometric manipulation of latent space for enhanced user interaction.
- Provides a framework applicable to various generative domains, including music and gesture.
Computer Science > Machine Learning arXiv:2602.21133 (cs) [Submitted on 24 Feb 2026] Title:SOM-VQ: Topology-Aware Tokenization for Interactive Generative Models Authors:Alessandro Londei, Denise Lanzieri, Matteo Benati View a PDF of the paper titled SOM-VQ: Topology-Aware Tokenization for Interactive Generative Models, by Alessandro Londei and 2 other authors View PDF HTML (experimental) Abstract:Vector-quantized representations enable powerful discrete generative models but lack semantic structure in token space, limiting interpretable human control. We introduce SOM-VQ, a tokenization method that combines vector quantization with Self-Organizing Maps to learn discrete codebooks with explicit low-dimensional topology. Unlike standard VQ-VAE, SOM-VQ uses topology-aware updates that preserve neighborhood structure: nearby tokens on a learned grid correspond to semantically similar states, enabling direct geometric manipulation of the latent space. We demonstrate that SOM-VQ produces more learnable token sequences in the evaluated domains while providing an explicit navigable geometry in code space. Critically, the topological organization enables intuitive human-in-the-loop control: users can steer generation by manipulating distances in token space, achieving semantic alignment without frame-level constraints. We focus on human motion generation - a domain where kinematic structure, smooth temporal continuity, and interactive use cases (choreography, rehabilitation, HCI) m...
