{"id":"5bd527de-9e48-46fa-bbb2-ca7f246c9b37","arxiv_id":"2509.05297","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"FlowSeek integrates a frozen depth foundation model and classical motion bases into a SEA-RAFT-style flow network, achieving state-of-the-art zero-shot generalization when trained on a single GPU.","lead":"FlowSeek is an optical flow model that borrows features from a pretrained depth-estimation network and a classical six-degree-of-freedom motion model, trained on a single consumer GPU. It reports larger zero-shot accuracy gains than the previous state-of-the-art SEA-RAFT on Sintel, KITTI, Spring, and LayeredFlow.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Depth-prior attribution is not isolated: no control with a non-depth strong feature extractor, so the central claim that depth foundation model priors drive FlowSeek's gains is unproven.","rationale":"The reader's conditional verdict is appropriate. I considered Table 3 baseline inconsistencies, but the paper explicitly reports one-GPU retrained SEA-RAFT numbers and explains the hardware effect; even though the abstract's 10%/15% do not match single entries, the central empirical comparison is not destroyed. The most load-bearing gap is the missing control for feature quality. DAv2's encoder is DINOv2, so the natural control is cheap and decisive. The paper deserves credit for ablating depth-model choice, BaseNet input, and backbone transfer, and for reproducible single-GPU training claims; this concern is about interpretation, not fabrication. No verdict change.","tokens_in":21136,"tokens_out":7598,"duration_ms":81585,"concrete_test":"Train FlowSeek (T) under the Table 1 protocol with DAv2's Φ replaced by DINOv2-S/Base patch features passed through the same BottleNeck, keeping DAv2's D0 only for BasesNet/ContextNet. Compare TartanAir EPE, KITTI 2012 Fl-EPE, and Sintel/KITTI zero-shot numbers to FlowSeek (T). If the DINOv2 variant matches FlowSeek within ~0.05 EPE (or reproduces most of the gain over SEA-RAFT), the depth-specific prior is not the driver; if it degrades substantially, the depth attribution is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"FlowSeek's central scientific claim is that the gains come from injecting geometric priors from a depth foundation model plus a classical motion basis. The evidence does not isolate those priors. In §3.2 (Eqs. 6–8), the model concatenates the last decoder features Φ from Depth Anything v2 into the correlation volume; DAv2 is built on DINOv2 features, so Φ may simply be a strong dense visual feature extractor. Table 1 varies the depth foundation model (DPT, DAv1, DAv2) and the flow backbone, but never replaces Φ with an equally strong non-depth feature extractor such as DINOv2. Similarly, Table 2 compares BasesNet on B_motion versus depth map D0 only; it does not test whether any informative 8-channel prior (e.g., image gradients or a generic feature embedding) would give the same boost. Consequently, the observed cross-dataset improvement could be due to additional feature capacity/quality rather than to depth-specific or motion-basis-specific knowledge. The empirical gain over SEA-RAFT is plausible and supported by the tables, but the attribution central to the paper's framing is unproven. This is addressable and does not by itself falsify the core result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FlowSeek, an optical flow architecture built on SEA-RAFT. It enriches the correlation features with the last decoder features of a monocular depth foundation model (Depth Anything v2), optionally feeds the estimated depth maps to the context network, and injects a classical rigid-motion basis (Heeger & Jepson) through a small BasesNet. Training is performed on a single RTX 3090 with batch sizes 6/4, i.e. at roughly an 8x smaller hardware budget than the multi-GPU SEA-RAFT baselines. The paper reports ablations on TartanAir/KITTI 2012 and zero-shot generalization results on Sintel, KITTI 2015, Spring and LayeredFlow, claiming superior accuracy over SEA-RAFT variants and, in the abstract, a relative improvement of 10% and 15% over SEA-RAFT on Sintel Final and KITTI.","tokens_in":21450,"tokens_out":10546,"duration_ms":97870,"significance":"If the empirical results hold, the paper makes a useful practical contribution: it shows that competitive optical flow models can be trained on a single consumer GPU by reusing pre-trained geometric foundation models instead of training from scratch on a large cluster. The evaluation is comparatively broad: it includes several depth foundation models, two flow backbones (SEA-RAFT, CRAFT, FlowFormer), multiple model sizes, and four zero-shot generalization benchmarks. The one-GPU retrained SEA-RAFT baselines are an appropriate comparison point, and the paper honestly states its reliance on large pre-trained models. The main weakness is that the mechanism attributed to the improvement—depth-specific and motion-basis-specific priors—is not isolated from a generic improvement in feature quality or capacity; the headline numbers in the abstract also need to be reconciled with the tables.","major_comments":[{"comment":"The central attribution claim is not isolated. Φ in Eq. (6) is the last decoder feature of Depth Anything v2, a model built on DINOv2 features; after the BottleNeck it is concatenated to the correlation features. Table 1 varies the depth foundation model (DPT, Depth Anything v1, v2) but never replaces Φ with an equally strong non-depth feature extractor (e.g., DINOv2 or a generic dense ViT feature) at the same resolution. The observed gains from +Φ and the selected best configuration could therefore be due to stronger visual features rather than depth-specific geometric knowledge. Please add such a control, or explicitly rephrase the paper's attribution so that it claims improved features from a foundation model rather than depth priors specifically.","section":"§3.2, Eq. (6)–(7); Table 1"},{"comment":"The abstract's 'relative improvement of 10 and 15%' does not match any single comparison in Table 3. With Tartan pretraining, FlowSeek (L) vs SEA-RAFT (L) as reported by the authors gives about 35% on Sintel Final (3.37→2.18) and about 12% on KITTI Fl-all (12.7→11.2); against the one-GPU retrained SEA-RAFT (L) the values are about 20% and 17%. The claimed 10%/15% figures need to be corrected or explicitly tied to a stated baseline, training schedule, and metric. It should also be clarified whether the comparison is within the Tartan-pretrained block (apples-to-apples) or against C→T-only baselines.","section":"Abstract; Table 3"},{"comment":"The value of the motion-basis prior is not established against a non-motion control. Table 2 only compares B_motion with the depth map D0 as the BaseNet input, both of which are geometry-based; it does not test whether the improvement comes from the rigid-motion subspace structure rather than from any arbitrary extra 8-channel input. Moreover, on KITTI 2015 dynamic regions B_motion is worse than D0 (Fl-EPE 2.60 vs 2.49; Fl-All 9.07 vs 8.61), so the benefit is limited to static regions. A control with a generic prior of matched channel count (e.g., coordinate grids or learned embeddings) would clarify the role of the motion bases.","section":"Table 2; §3.3"}],"minor_comments":[{"comment":"The notation Δ_i^F is used both for a flow residual and for the density of a mixture of Laplace distributions. The equation appears to define a probability density over x, not the residual itself; please clarify the distinction between the predicted distribution parameters and the actual residual used in Eq. (4).","section":"Eq. (14)"},{"comment":"Typo: 'dowsampling' should be 'downsampling' in the Spring paragraph.","section":"§4.3"},{"comment":"The one-GPU SEA-RAFT numbers appear to be from a single retraining run; given the unusual result that the one-GPU SEA-RAFT (S) is better on Sintel Final than the authors' multi-GPU model (3.02 vs 4.32), it would be helpful to report the variance or number of seeds used for these baselines.","section":"Table 3"},{"comment":"The text describes the model as 'compact, yet accurate', but the MAC counts in Table 1 show that adding the BaseNet and Φ roughly doubles the cost of the small model (e.g., 694.7G vs 284.7G for SEA-RAFT (S)). This is not a problem for the central claim, but the wording should be qualified.","section":"Table 1; §4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is an empirical architecture contribution with a defensible headline result, but the depth-prior attribution is currently underdetermined because no strong non-depth feature extractor is used as a control. The abstract's 10%/15% numbers also need to be reconciled with Table 3. These issues are fixable within the paper's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a solid empirical architecture paper. It combines a frozen depth foundation model (Depth Anything v2) with classical motion bases inside a SEA-RAFT-style backbone, and shows you can get state-of-the-art zero-shot generalization on Sintel Final, KITTI, Spring, and LayeredFlow while training on a single RTX 3090 with a much smaller batch than the multi-GPU baselines. That's a real practical contribution, especially for groups without big hardware.\n\nWhat's genuinely new: no prior optical flow work I know of injects the intermediate features of a single-image depth foundation model into the correlation volume and context, and pairs it with a classical motion basis. The paper also demonstrates the recipe transfers to CRAFT and FlowFormer, which makes the idea more credible than a one-off trick. The ablations in Table 1 are thorough—prior combinations, model sizes, depth models, backbones—and Table 2 shows motion bases beat raw depth maps as the BaseNet input. The writing is clear and the single-GPU training schedule is a meaningful practical detail.\n\nThe soft spots are real but addressable. The main one: the paper attributes the gains to 'depth priors,' but Depth Anything v2 is built on DINOv2, so the features you're concatenating might just be strong dense visual features. There is no control with a non-depth strong feature extractor (e.g., DINOv2 alone) to isolate depth-specific knowledge. That doesn't kill the result—the architecture still works—but it means the paper's framing is overreach. The motion basis is similarly justified for rigid motion and only partially tested against a generic informative prior; a minor concern given the gains on non-rigid scenes.\n\nTwo smaller issues: the abstract's 10% and 15% improvement figures don't match any single pair in Table 3 exactly—they're presumably averages or a different comparison, but the reader shouldn't have to guess. And the retrained one-GPU SEA-RAFT baseline numbers in Table 3 look internally inconsistent: the one-GPU retrained model sometimes beats the original multi-GPU model on the same metric, which is confusing and needs a footnote or corrected numbers. These are presentation problems, not fatal flaws.\n\nOverall: the central empirical claim is supported, the attribution is unproven but fixable with one control experiment, and the paper deserves a serious referee. I'd send it to review, and I'd cite it for the architecture recipe.","headline":"Solid empirical architecture paper with a real single-GPU training contribution, but the depth-prior attribution is unproven and the baseline numbers need cleaning up.","tokens_in":21894,"tokens_out":2531,"would_cite":true,"duration_ms":24859,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FlowSeek shows that a compact optical flow network, boosted by frozen depth-model features and classical motion bases, matches or beats state-of-the-art flow methods while training on a single consumer GPU.","keywords":["optical flow","depth foundation models","motion bases","cross-dataset generalization","single-GPU training","SEA-RAFT","zero-shot transfer","rigid motion subspace"],"falsifier":"Train FlowSeek with the depth foundation model replaced by a frozen generic dense feature extractor (e.g., a self-supervised ViT) and identical BottleNeck and BaseNet modules; if Sintel Final and KITTI errors stay within the reported margins, depth-specific priors are not the cause. Also evaluate on a clearly non-rigid sequence and compare FlowSeek with and without the BaseNet to see whether the rigid-motion initial guess hurts.","tokens_in":21030,"feed_emoji":"🎥","tokens_out":4960,"duration_ms":45919,"temperature":0.7,"pith_summary":"FlowSeek argues that state-of-the-art optical flow does not require a large cluster of GPUs. By injecting priors from a frozen depth foundation model (Depth Anything v2) and classical low-dimensional motion bases into a SEA-RAFT-style architecture, the authors train a compact model on a single consumer GPU and report a relative improvement of 10% on Sintel Final and 15% on KITTI over the previous state of the art, with strong zero-shot results on Spring and LayeredFlow. The paper claims the depth model's inverse-depth maps and decoder features, combined with an eight-dimensional rigid-motion basis, give the flow network a strong geometric starting point that generalizes across datasets. If right, this lowers the hardware barrier for flow research and suggests that recycling large pretrained geometric models can substitute for brute-force training scale.","feed_headline":"One GPU now beats multi-GPU optical flow models","feed_subtitle":"Frozen depth features plus classical motion bases: Sintel Final error 10% and KITTI 15% lower than SEA-RAFT.","key_machinery":"Two priors carry the argument. First, a frozen depth foundation model (Depth Anything v2) supplies inverse-depth maps and decoder features that are correlated with depth; the features are adapted by a three-convolution BottleNeck and concatenated into the correlation-volume features, and the depth maps can be added to the context network. Second, from the classical result that a rigid scene's optical flow lies in a six-dimensional subspace, the paper builds eight motion bases from inverse depth, removes focal-length dependence by rescaling, and runs them through a BaseNet so the network starts from a geometrically plausible motion subspace and refines it iteratively.","core_discovery":"The central claim is that an optical flow network trained on one consumer GPU can outperform models trained on eight GPUs when it is seeded with geometric priors from a depth foundation model. Inverse-depth maps D0, D1 and last-decoder features from Depth Anything v2 are fused into the backbone's features and context; a set of eight motion bases — three translational and five rotational components derived from D0 with focal-length terms scaled away — is processed by a small BaseNet and injected as an initial motion prior. With the SEA-RAFT training recipe and mixture-of-Laplace supervision, FlowSeek variants beat their SEA-RAFT counterparts on TartanAir, Sintel, KITTI, Spring, and LayeredFlo","pith_inferences":["The paper does not ablate whether the gains come from depth-specific information or from any dense feature extractor; swapping in a generic self-supervised visual feature model would test whether depth is the causal ingredient.","The rigid-motion basis is theoretically justified only for static scenes; testing on heavily non-rigid motion would reveal whether the 'initial guess' helps or hurts when the prior is wrong.","The success suggests a broader recipe: freeze a large geometric foundation model and inject classical low-dimensional structure to train compact models for other dense prediction tasks on modest hardware.","The 8x smaller batch size changes optimization dynamics as well as memory; part of the generalization gain may come from the regularization effect of small batches, not only from the priors."],"forward_implications":["State-of-the-art cross-dataset generalization on Sintel Final, KITTI, Spring, and LayeredFlow becomes reachable with a single consumer GPU, lowering the hardware barrier for flow research.","Depth foundation models can be recycled as frozen priors for motion estimation rather than trained from scratch, avoiding prohibitive compute.","The combination of depth features and motion bases outperforms either prior alone, and the bases beat feeding raw depth to the same network.","The recipe transfers across backbones: CRAFT and FlowFormer also improve when given the same priors.","Newer or stronger depth foundation models should translate directly into further flow accuracy gains."],"supporting_citations":[{"why":"Supplies the SEA-RAFT backbone, training schedule, and mixture-of-Laplace supervision that FlowSeek builds on.","marker":"[81]"},{"why":"The Depth Anything v2 foundation model that provides inverse-depth maps and decoder features used as priors.","marker":"[93]"},{"why":"The classical result that rigid-motion flow fields live in a six-dimensional subspace, from which the eight motion bases are derived.","marker":"[21]"},{"why":"Establishes all-pair correlation volumes, iterative refinement, and convex upsampling that the flow backbone uses.","marker":"[76]"},{"why":"Shows the BottleNeck adapter design for injecting foundation-model features into geometric matching networks.","marker":"[83]"},{"why":"Provides evidence that last decoder features correlate strongly with depth, motivating the feature-injection strategy.","marker":"[35]"},{"why":"TartanAir is the primary synthetic training and evaluation dataset for the ablation and generalization studies.","marker":"[80]"}],"fun_headline_variants":["Depth priors let one-GPU flow beat multi-GPU","Single GPU optical flow beats 8-GPU rivals","FlowSeek: one GPU, better flow than SEA-RAFT","Depth foundation models boost optical flow on one GPU","One GPU flow model tops multi-GPU with depth priors"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The improvement is attributed to depth priors, but no experiment swaps the depth model for a different strong feature extractor, and the motion-basis prior is assumed to help on non-rigid scenes even though it is derived for rigid motion.","fun_headline_variants_meta":{"raw":{"variants":["Depth priors let one-GPU flow beat multi-GPU","Single GPU optical flow beats 8-GPU rivals","FlowSeek: one GPU, better flow than SEA-RAFT","Depth foundation models boost optical flow on one GPU","One GPU flow model tops multi-GPU with depth priors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000322,"raw_usage":{"total_tokens":1601,"prompt_tokens":648,"completion_tokens":953,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":392,"completion_tokens_details":{"reasoning_tokens":881}},"tokens_in":392,"tokens_out":953,"duration_ms":7277,"temperature":1.0,"reasoning_tokens":881,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:23:56.734270+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train FlowSeek with the depth foundation model replaced by a frozen generic dense feature extractor (e.g., a self-supervised ViT) and identical BottleNeck and BaseNet modules; if Sintel Final and KITTI errors stay within the reported margins, depth-specific priors are not the cause. Also evaluate on a clearly non-rigid sequence and compare FlowSeek with and without the BaseNet to see whether the rigid-motion initial guess hurts.","supporting_citations":[{"cited_title":"Sea-raft: Simple, efficient, accurate raft for optical flow","cited_arxiv_id":null,"evidence_quote":"Supplies the SEA-RAFT backbone, training schedule, and mixture-of-Laplace supervision that FlowSeek builds on."},{"cited_title":"Depth any- thing v2.Advances in Neural Information Processing Sys- tems, 37:21875–21911, 2024","cited_arxiv_id":null,"evidence_quote":"The Depth Anything v2 foundation model that provides inverse-depth maps and decoder features used as priors."},{"cited_title":"Foundationstereo: Zero- shot stereo matching.arXiv, 2025","cited_arxiv_id":null,"evidence_quote":"Shows the BottleNeck adapter design for injecting foundation-model features into geometric matching networks."},{"cited_title":"Defom-stereo: Depth foundation model based stereo matching","cited_arxiv_id":null,"evidence_quote":"Provides evidence that last decoder features correlate strongly with depth, motivating the feature-injection strategy."},{"cited_title":"Tartanair: A dataset to push the limits of visual slam","cited_arxiv_id":null,"evidence_quote":"TartanAir is the primary synthetic training and evaluation dataset for the ablation and generalization studies."}],"review_version":1}