Optical flow estimation — computing pixel-level motion between video frames — works well on clean data and poorly on degraded data. Blur, noise, and compression artifacts confuse the appearance-matching that flow algorithms rely on.
Note (2026-11-12; read against the paper's full text): the paper says the intermediate representations of image-restoration diffusion models are corruption-aware but lack temporal awareness, and that its contribution is to lift the model to attend across adjacent frames with full spatio-temporal attention, after which the features show zero-shot correspondence; the essay's statements that the restoration model already knows how a blurred moving edge differs from a stationary one, and that the diffusion features are corruption-invariant, are not the paper's description. The paper takes the fine-grained local detail from the convolutional encoder of RAFT, with ablations of the encoder and the upsampling in its appendix, and reports lower endpoint error at nearly every step in one comparison, with a single step where the two methods are virtually tied; the essay's claim that neither approach alone achieves this rests on the encoder ablation in the appendix, whose results I did not read.
The authors (arXiv:2603.23499) discover that image restoration diffusion models, which already understand corruption types, naturally capture temporal correspondence when extended to adjacent frames. Adding spatio-temporal attention to a diffusion denoiser produces features that encode both corruption-awareness and motion information.
DA-Flow merges these diffusion-based features with traditional convolutional features through iterative refinement. The convolutional features provide fine spatial detail; the diffusion features provide corruption-invariant matching. The hybrid substantially outperforms existing flow methods under severe degradation.
The through-claim: the degradation model and the motion model are not independent problems — they share latent structure. A diffusion model trained to remove blur already knows that the blurred appearance of a moving edge differs from a stationary edge. This implicit motion knowledge, extracted from the diffusion process and combined with explicit flow estimation, produces robust correspondence that neither approach achieves alone.