Predicting DINO Features Instead of Denoising Speeds Up Flow Models
The paper argues that flow matching models for image generation benefit from being conditioned on pretrained DINO representations, but prior approaches awkwardly co-denoise these representations alongside the image via a second ODE trajectory with custom schedules. The authors propose simply predicting the DINO representation directly (since it's deterministic given the image) and conditioning generation on that prediction, avoiding the extra denoising process entirely. They report this yields faster convergence and better FID scores on ImageNet, beating prior state-of-the-art latent-space methods at half the training epochs and improving pixel-space FID by over 20%. The authors themselves framed the core idea concisely on Twitter: don't denoise information you can just predict outright, since DINO features are a deterministic function of the clean image rather than something requiring a stochastic denoising trajectory. Engagement so far is from the authors promoting the paper rather than independent critical discussion.
Discussion: 2 tweets from 2 authors · @sohonjitghosh, @lucasdegeorge