We present SEA-RAFT, a more simple, efficient, and accurate method for optical flow estimation built upon the RAFT architecture. Compared to RAFT, SEA-RAFT introduces three key improvements: (1) a new mixture of Laplace loss function that better handles the multi-modal nature of optical flow uncertainty; (2) direct regression of an initial flow estimate that provides a better starting point for iterative refinement, leading to faster convergence; and (3) rigid-motion pre-training on synthetic data to improve generalization to real-world scenes. SEA-RAFT achieves state-of-the-art accuracy on the Spring benchmark with 3.69 endpoint-error (EPE) and 0.36 1-pixel outlier rate, representing 22.9% and 17.8% error reduction from best published results, while being at least 2.3x faster than existing methods.

Optical flow estimation — the task of computing dense per-pixel motion between consecutive video frames — is a fundamental problem in computer vision with applications in video understanding, autonomous driving, video editing, and action recognition. The introduction of RAFT (Recurrent All-Pairs Field Transforms) marked a paradigm shift in the field, establishing an architecture based on iterative refinement of flow through a recurrent update operator. RAFT's design has proven remarkably influential, with subsequent works largely building incremental improvements on top of its framework. However, many of these follow-up works introduce significant computational overhead through larger feature extractors, more complex correlation computations, or additional network components.

The Mixture of Laplace (MoL) loss is our first key contribution. Standard optical flow training uses an L1 or L2 loss over the sequence of iterative predictions, treating flow estimation as a deterministic regression problem. However, optical flow is inherently ambiguous in many regions — at occlusion boundaries, textureless areas, and repetitive patterns, multiple flow values are plausible. The MoL loss models the flow prediction as a mixture of Laplace distributions, where the model outputs both the predicted flow and a confidence map. This allows the model to express uncertainty and allocate its capacity to learn from reliable regions while being robust to ambiguous ones. Empirically, the MoL loss provides consistent improvements across all benchmarks compared to the standard loss.

Our second contribution is direct initial flow regression. In the original RAFT, the iterative refinement starts from a zero-initialized flow field, requiring many iterations to converge to the correct solution, especially for large displacements. We add a lightweight regression head that directly predicts an initial flow estimate from the correlation volume. This provides a much better starting point for the iterative updates, reducing the number of iterations needed for convergence by approximately 50%. Our third contribution, rigid-motion pre-training, uses synthetically generated scenes with known rigid-body transformations to provide the model with a strong prior for real-world motion patterns, significantly improving cross-dataset generalization, particularly on KITTI and Spring.

We evaluate SEA-RAFT on four major optical flow benchmarks: Sintel, KITTI, Spring, and FlyingThings. On the Spring benchmark, SEA-RAFT achieves a new state-of-the-art with 3.69 EPE, surpassing the previous best of 4.79 EPE. On KITTI 2015, we achieve an Fl-all rate of 4.31%, competitive with significantly larger models. Importantly, SEA-RAFT achieves the best cross-dataset generalization performance — when trained only on FlyingThings and evaluated on KITTI and Spring without fine-tuning, our method outperforms all baselines. In terms of efficiency, SEA-RAFT processes 1080p frames at 8.3 FPS, which is 2.3x faster than FlowFormer and 3.1x faster than GMFlow+, while achieving better or comparable accuracy.

We have presented SEA-RAFT, demonstrating that carefully designed, simple modifications to the RAFT framework can yield substantial improvements in accuracy, efficiency, and generalization. Our three contributions — mixture of Laplace loss, direct initial flow regression, and rigid-motion pre-training — are each independently beneficial and combine synergistically. SEA-RAFT achieves state-of-the-art results on Spring with significant error reductions, the best cross-dataset generalization, and at least 2.3x speedup over comparable methods. Our work demonstrates the value of revisiting fundamental design choices rather than pursuing ever-more-complex architectures.