arxiv.org web signal

Dino Forcing cuts flow-matching training to half the epochs

By Published Updated

TL;DR

  • The method predicts a pretrained DINO representation directly and conditions the flow model on its own prediction, dropping the second denoising ODE.
  • On ImageNet in latent space the authors claim state-of-the-art quality at 2x fewer epochs than prior methods.
  • In pixel space, they report FID improves by more than 20% over comparable prior methods.

"Co-denoising pretrained representations such as DINO can substantially improve the training speed and quality of flow matching models, but it introduces a second denoising trajectory and requires carefully designed schedules." That is the opening line of a new preprint on arXiv from Arijit Ghosh, Lucas Degeorge, Paul Couairon, Alexei A. Efros, Vicky Kalogeiton and David Picard.

Their proposal is narrow and concrete: predict the pretrained DINO representation directly, then condition the model on its own prediction. The paper says this "removes the need for a second ODE and any representation-specific denoising schedules, while retaining the benefits of representation guidance."

The reported wins come in two numbers. On ImageNet in latent space, the method "outperforms the state of the art" at "2x fewer epochs than prior methods." In pixel space, FID improves "by more than 20%" over comparable prior methods. The authors fold the result into a one-line principle: "do not denoise what you can predict."

The abstract publishes no absolute FID numbers, and the identity of the "comparable prior methods" is left to the paper body.

Shared on Bluesky by 2 AI experts

Originally reported by arxiv.org

Read the original article →

Original headline: Dino Forcing Flow Models: Do not denoise what you can predict