Yu-Han Wu

Yu-Han Wu 吳雨翰

PhD student in diffusion models

Sorbonne University · Google DeepMind

All talks

Poster EPFL Bernoulli Center · Lausanne, Switzerland

Lemanth 2026

Poster presentation of MIND: Monge Inception Distance for Generative Models Evaluation.

Poster Paper page

Poster

poster page 1 of 1
Download PDF

The paper behind this talk

MIND: Monge Inception Distance for Generative Models Evaluation

NeurIPS 2026

Quentin Berthet, Yu-Han Wu, Clément Crepy, Romuald Elie, Klaus Greff, Michael E. Sander

  • Problem. FID needs tens of thousands of samples and can be fooled by matching moments.
  • Approach. A sliced Wasserstein distance between Inception features, computed by sorting one-dimensional projections, with no high-dimensional means or covariances to estimate.
  • Result. An order of magnitude more sample-efficient, two orders of magnitude faster, more robust; 5k samples match FID at 50k.
Key figure from MIND: Monge Inception Distance for Generative Models Evaluation
Left: MIND during a diffusion model training run on ImageNet-64 (log scale), illustrating how MIND with 5k samples can replace FID with 50k, with a larger range. Right: correlation with the number of training steps, better for MIND with 1k or 5k samples than for FID with 50k.

Read more about the paper

All talks