Reproduce “Dino Forcing Flow Models: Do not denoise what you can predict” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Arijit Ghosh, Lucas Degeorge, Paul Couairon, Alexei A Efros, Vicky Kalogeiton, David Picard
DINO Forcing uses a pretrained DINO vision encoder to guide image generation without separately denoising a feature stream. From a noisy image, it predicts DINO features and reuses them as conditioning, with feature alignment during training. On ImageNet, it reports unguided FID 2.47 after 100 latent-space epochs and 5.53 after 200 pixel-space epochs; these comparisons are benchmark-specific, not wall-clock evidence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.