← AI Techniques Radar
Training & fine-tuning

Critic Spectral Adaptation (CSA) for improving motion-tokenizer diffusability

Improve a motion tokenizer used with diffusion by adding a spectral-focused adaptation step during text-conditioned training. A pretrained diffusion model acts as a critic to suppress undesired high-frequency energy in the tokenizer’s latent space, aiming to reduce a generation–reconstruction mismatch attributed to spectral bias and temporal artifacts.

Discovered 2026-11-10 · Sources checked 2026-11-10

This summary was generated with AI assistance.

Why it matters

The approach targets a specific failure mode where scaling tokenizer capacity improves reconstruction but not text-conditioned generation quality, suggesting the latent representation’s frequency characteristics can hinder diffusion-based denoising. By steering latent spectra via a diffusion-based critic, the tokenizer is made more compatible with the diffusion process used downstream.

What changed

Instead of increasing tokenizer capacity or modifying only the diffusion backbone, the method introduces a training component that uses a pretrained diffusion model as a critic to adjust (adapt) the tokenizer’s latent spectral properties, specifically suppressing high-frequency latent energy to improve diffusability.

Where it helps

  • Text-to-motion generation with a VAE-based motion tokenizer for diffusion
  • Improving latent-space compatibility between tokenizers and diffusion denoisers
  • Reducing high-frequency artifacts in motion latent representations

Evidence & reproduction

Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.

Limitations & open questions

The provided description does not specify training loss formulations, implementation details, or failure cases; it only states the motivation (spectral mismatch) and the role of a pretrained diffusion model as a critic to suppress high-frequency latent energy.

Related World Models papers

Sources