# Dino Forcing Flow Models: Predicting Representations Directly Beats Denoising Them

_Research · published 2026-10-14_

A new arXiv paper from researchers including Arijit Ghosh, Lucas Degeorge, Paul Couairon, Alexei Efros, Vicky Kalogeiton, and David Picard proposes a simpler alternative to co-denoising pretrained representations like DINO in flow matching models. Rather than introducing a second denoising trajectory with its own ODE and schedule, the method predicts the pretrained representation directly and conditions the model on that prediction — removing architectural complexity while retaining the representational benefits.

The approach demonstrates strong results on ImageNet, outperforming the state of the art in latent space at half the training epochs, and improving FID scores in pixel space by more than 20% over comparable prior methods. Code is publicly available, and the authors summarize their finding as a clean guiding principle: do not denoise what you can predict.

## Sources
- [arxiv.org](https://arxiv.org/abs/2610.11751)
---
Canonical: https://genbuzz.news/posts/dino-forcing-flow-models-predicting-representations-directly-beats-denoising-them