cs.CVOct 8, 2026

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

Authors: Arijit Ghosh, Lucas Degeorge, Paul Couairon, Alexei A Efros, Vicky Kalogeiton, David Picard

Organizations: ENPC, IP Paris · Ecole Polytechnique, IP Paris · AMIAD · UC Berkeley

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

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. We propose a simpler alternative: predict the pretrained representation directly, then condition the model on its own prediction. This removes the need for a second ODE and any representation-specific denoising schedules, while retaining the benefits of representation guidance. Our approach converges substantially faster and achieves better generation quality as measured by FID score. On ImageNet, it outperforms the state of the art in latent space at 2x fewer epochs than prior methods; in pixel space, it improves FID over comparable prior methods by more than 20%. These results support a simple principle: do not denoise what you can predict. Our code is openly available at https://github.com/arijit-hub/dino_forcing.

Figures & tables

Figure 1: Left: Prior co-denoising methods generate both the image and the pretrained representation from noise by jointly solving their individual ODEs through iterative modification during sampling. The model predicts the velocity for both modalities at different time schedules ( top ) which leads to 2 trajectories starting from independent noise samples and arriving at the same pair ( bottom ). Right: Our proposed Dino Forcing ( DF ) method keeps the ODE only for the image generation since the representation is deterministically tied to the image. The model predicts the target representation in addition to the image velocity ( top ), providing a self-condition to guide the image ODE ( bottom ).
Figure 2: Left: Qualitative samples. More examples are given in Figures 8 to 13 . Right: Convergence speed-up (FID-50K vs. training epochs).
Figure 3: Overview of our proposed training framework . A first forward pass produces a representation look-ahead (no gradient), which is then used to condition a second pass that optimizes both the flow matching and representation alignment objectives.
Figure 4: Overview of our sampling framework. At each step the model outputs a ODE step , and a prediction z^\hat{z} used as the self-condition at the next step. Note that z^\hat{z} is not a denoised feature map but the model’s prediction of the DINO features of the denoised image x^\hat{x} .
Figure 5: Evolution of the image and the PCA of the predicted DINO features throughout sampling.
Table 1: Model-size ablation with DINO as the frozen representation.
Table 2: State-of-the-art comparison of latent space models on ImageNet 2562256^{2} . Bold is best, underline is second best, grayed methods use more than 200 epochs and are not considered for the comparison.
Table 3: State-of-the-art comparison of pixel-space models on ImageNet 2562256^{2} . Bold is best, underline is second best, grayed methods use more than 200 epochs and are not considered for the comparison.
Figure 6: Reconstructions from inverted DINOv2-B features. Top row: originals; bottom row: reconstructions.
Table 10
Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

 Latent spacePixel space
Backbone
Base implementationSiT ( Ma et al., 2024 )JiT ( Li and He, 2026 )
Input spacesd-vae-ft-ema latentsRGB pixels
Resolution256×256256\times 256 , 512×512512\times 512256×256256\times 256
Patch size216
Dropout (XL only)0.1, middle third of blocks0.1, middle third of blocks
Appendix
Table 8: Training and sampling configuration of DF in latent and pixel space on ImageNet.
Figure 7: Layer index and dropping probability ablation on SiT-L model.
Figure 8: Uncurated samples from DF-SiT-XL/2 on ImageNet 256×256256{\times}256 , without guidance.
Figure 9: Uncurated samples from DF-SiT-XL/2 on ImageNet 256×256256{\times}256 , with cfg=4.0.
Figure 10: Uncurated samples from DF-SiT-XL/2 on ImageNet 512×512512{\times}512 without guidance.
Figure 11: Uncurated samples from DF-SiT-XL/2 on ImageNet 512×512512{\times}512 , with cfg=4.0.
Figure 12: Uncurated samples from DF-JiT-XL/16 on ImageNet 256×256256{\times}256 without guidance.
Figure 13: Uncurated samples from DF-JiT-XL/16 on ImageNet 256×256256{\times}256 with cfg=4.0.
Figure 14: Evolution of the image and the PCA of the predicted DINO features throughout sampling ( t=0t{=}0 : noise, t=1t{=}1 : data), for DF-SiT-XL/2 at 256×256256{\times}256 and 512×512512{\times}512 and DF-JiT at 256×256256{\times}256 . The PCA projection is fit once per sample on the final features and reused for all timesteps.

Explore similar work

CardsList

  1. RiT: Vanilla Diffusion Transformers Suffice in Representation Space

    May 21, 2026Le Zhang, Ning Mang, Aishwarya AgrawalFlow MatchingVisual Representation Learning

    Mila – Québec AI Institute, UdeM · Utrecht University · Canada CIFAR AI Chair

  2. Rethinking Pixel Mean Flows via Interval Denoiser

    Aug 5, 2026Alexander Zaytsev, Dmitry Baranchuk, Alexander Korotin +1Flow MatchingImage Generation

    HSE University · Yandex Research · Applied AI Institute +2

  3. Training Flow Matching: The Role of Weighting and Parameterization

    Mar 6, 2026Anne Gagneux, Ségolène Martin, Rémi Gribonval +1Flow MatchingAdaptive Loss Weighting

    ENS de Lyon, CNRS, Universit´e Claude Bernard Lyon 1, Inria, LIP, UMR 5668, France · Technische Universität Berlin, Germany · Inria, ENS de Lyon, CNRS, Universit´e Claude Bernard Lyon 1, LIP, UMR 5668, France