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

TL;DR

Dino Forcing takes aim at the co-denoising recipes (RAE, REPA-style parallel-denoising variants) that run a second denoising trajectory on a pretrained representation such as DINO. Instead of jointly denoising the image and its DINO features, the model predicts the clean DINO representation and then conditions on its own prediction while denoising the image. One ODE, no representation-specific noise schedule, same representation-guidance benefit. Reported FID gains on ImageNet 256²: SOTA in latent space at 2× fewer epochs than prior methods, and >20% FID improvement over comparable pixel-space baselines.

Key claims

Co-denoising pretrained representations (DINO) alongside the image improves flow-matching training speed and quality but requires a second denoising trajectory and a representation-specific noise schedule [Abstract].
Replacing the second denoising trajectory with a direct prediction of the pretrained representation, then conditioning the main denoising on that prediction, retains the representation-guidance benefit without the extra ODE or schedule [Abstract].
Converges substantially faster than co-denoising recipes and reaches better final FID [Abstract].
On ImageNet latent-space, outperforms prior SOTA at 2× fewer training epochs [Abstract].
On ImageNet pixel-space, improves FID by more than 20% over comparable prior methods [Abstract].
Code released openly [Abstract].

Method

The paper frames two prior design points: (a) REPA-style auxiliary alignment to a frozen DINO feature, which doesn’t require an extra ODE but constrains intermediate activations; (b) co-denoising recipes that denoise both the image and a noisy DINO feature jointly, requiring a second trajectory and a carefully designed representation-specific noise schedule. Dino Forcing proposes a middle path: a single denoising ODE for the image, with a side-head that predicts the clean DINO representation from the current (noisy) image state; the main denoiser is then conditioned on this self-predicted DINO target. The result is architecturally simpler (one ODE, standard schedule) while reusing the representation-guidance mechanism.

Results

ImageNet 256² latent-space: beats prior SOTA with 2× fewer training epochs than prior methods [Abstract].
ImageNet 256² pixel-space: >20% FID improvement over comparable prior methods [Abstract].
Code openly available [Abstract].

(Quantitative FID numbers and encoder/backbone details are not visible in the fetched abstract.)

Why it’s interesting

Fits directly into the fork documented in Diffusion training efficiency and Representation Autoencoders between “put DINO in the encoder” (RAE / FAE), “distill DINO into intermediate layers” (REPA), and “self-distill without external features” (SRA / Self-Flow). Dino Forcing adds a fourth design point: predict the external representation from the current denoising state and feed the prediction back as conditioning. This complements Diffusion Transformers with Representation Autoencoders (RAE’s co-denoising-with-wide-DDT-head) by removing the second ODE entirely, and contrasts with No Other Representation Component Is Needed: Diffusion Transformers Can Provide Representation Guidance by Themselves (SRA) (SRA) which argues external DINO is unnecessary at all — Dino Forcing keeps the external target but makes it a prediction target rather than a co-denoising target. The pixel-space headline also contributes a new datapoint to Pixel-space diffusion, joining PixelGen: Improving Pixel Diffusion with Perceptual Supervision (PixelGen / P-DINO) and Training Text-to-Image Models 3.6× Faster (JiT-DDT) (JiT-DDT) on the “DINO supervision for pixel diffusion” axis.