acceptodds
Under review as a conference paper at ICLR 2027

TDFlow: Efficient and Accurate Optical Flow Estimation using Truncated Diffusion Model

Abstract

Diffusion models have recently emerged as a powerful generative technique for diverse tasks in computer vision and robotic policy learning. Their extension to optical flow estimation represents a promising direction, enabling more accurate motion estimation for robot localization. Nevertheless, the heavy computational cost arising from the large number of denoising steps in vanilla diffusion policies poses a major obstacle to practical efficiency and real-world applications. To address this challenge, we propose a novel truncated diffusion policy, named TDFlow, that leverages initial flow estimates as anchors and shortens the diffusion steps by learning the denoising process directly from an anchored Gaussian distribution to the flow output, rather than the original pure Gaussian distributions. Furthermore, we design a Conditional Feature-Space Warping Diffusion Decoder for efficient flow refinement, conditioned by the scene context and flow anchors. Compared with existing approaches, our TDFlow achieves state-of-the-art performance on three optical flow datasets: MPI Sintel, Spring, and KITTI benchmarks, and further demonstrates a superior zero-shot generalization property on KITTI. Due to our designed truncated policy, our model requires only one twentieth of the denoising steps compared to the vanilla diffusion policy, while running up to faster than other diffusion-based methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Optical Flow, Diffusion Model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_f4GxOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose TDFlow, a truncated diffusion model for optical flow that uses an initial flow estimate as an anchor and denoising in a single step to reduce computation, paired with a feature-space warping decoder to maintain accuracy. Author report state-of-the-art results with significantly lower latency than prior diffusion-based methods.

Critical strengths

Diffusion-based flow methods are known to be slow, and This paper makes the whole thing single-step with large measured speedups in latency and memory. That directly addresses the main blocker for deploying generative flow estimators in practice, and the accuracy is kept at a competitive level.

The truncated diffusion policy is integrated on top of several existing anchors from different method families, and it consistently improves all of them. This shows the idea is not tied to one specific backbone, which makes easier for others tasks to adopt.

Critical weaknesses

Q1:The paper tests five anchors for compatibility, but never analyzes how anchor quality affects the final gain or where the quality floor is. If the flow anchor is bad enough that the true flow falls outside the noise perturbation range, yet this is not tested.

Q2:All compared methods are from 2025 or earlier. For an ICLR 2027 submission, how do we know TDFlow still beats the latest 2026 methods? Please note that WAFT is ICLR2026, but it was published earlier on the internet.

Q3:The no-noise ablation already reaches KITTI EPE 3.02 vs. 2.87 with noise,the difference is not significant. This suggests the CFSW residual decoder is the real workhorse while diffusion adds only marginal benefit. No comparison against flow matching or deterministic regression is provided. Is the choice of diffusion necessary?

Questions for authors

Please answer the question, carefully, the weak acceptance is not fixed.

9 comments
AuthorsOpenReview ↗

Response to Reviewer f4Gx (R3)

Q1. How does anchor quality affect the gain, and what happens with no informative anchor?

The submitted compatibility study spans several anchor qualities. Table R7 reproduces the KITTI values and mean relative reduction over EPE and Fl-all, following the manuscript's Eq. (6).

Table R7. Refinement across anchor families after C+T training. Fl-all and PG are percentages.

AnchorBase EPERefined EPEBase Fl-allRefined Fl-allPG
GMA4.694.2917.1015.209.8
KPA-Flow4.463.9215.9014.5010.5
SKFlow4.273.7715.5013.7011.7
DPFlow3.373.1411.1010.605.7

The three less accurate baselines obtain larger mean relative improvements than DPFlow. Their corresponding Sintel improvements are 5.8%, 7.8%, and 4.7%, compared with 3.1% for DPFlow. This supports effectiveness beyond highly accurate anchors. The relationship is not strictly monotonic: initial aggregate error does not describe the spatial structure or correctability of errors.

We additionally trained a pure-noise-initialized configuration, replacing the informative flow anchor with pure noise while retaining image conditioning:

Table R8. Final predictions of separately trained initialization configurations. These are not errors of the initial noise.

InitializationSintel Clean EPESintel Final EPEKITTI EPEKITTI Fl-all (%)
Pure-noise flow input1.572.934.7917.87
Image-conditioned anchor0.872.092.879.30

The pure-noise configuration still learns optical-flow prediction from image evidence, but is substantially less accurate. This supports the anchor's role as an informative motion prior.

The decoder is supervised against ground-truth flow, so its predicted residual is not bounded by the injected perturbation magnitude. The pure-noise result supports the benefit of an informative anchor in the tested training configuration.

AuthorsOpenReview ↗

Response to Reviewer f4Gx (R3)

Q2. Comparison with recent methods from 2025–2026.

We have updated Table 3(a,b) (p. 7) of the revised manuscript to include these recent methods, with benchmark and cross-dataset results kept separate, and added brief discussion and the corresponding references. We broaden the comparison to FlowSeek (ICCV 2025), OFM (CVPR 2026), FreeFlow and MegaFlow (ECCV 2026), FlowIt (BMVC 2026), and the CVPR 2026 ViSCALE Workshop method of Chanda and Kumar. Table R9 compares hidden-test results after target-benchmark training; Table R10 examines cross-dataset generalization without target-domain flow fine-tuning. These are published results with method-specific pretraining and training budgets.

Table R9. Published hidden-test results after target-benchmark training. Sintel columns are EPE; KITTI is Fl-all (%). Lower is better. All listed methods use two input frames.

MethodVenueCleanFinalKITTI Fl-all
OFMCVPR 20260.941.853.78
FlowItBMVC 20261.012.454.50
FreeFlowECCV 20261.031.994.06
TDFlowOurs0.961.843.34

Sources: OFM, Table 1 [3]; FlowIt v2, Table 1 [6]; and FreeFlow, Table 4 [7]. FlowIt and FreeFlow use the published S configurations. The FlowIt row uses its TartanAir-pretrained setting; FreeFlow uses additional cross-view pretraining.

TDFlow improves all three reported metrics over the listed FlowIt and FreeFlow results. OFM is slightly better on Clean, while TDFlow has lower Final EPE and KITTI Fl-all. These results position TDFlow as competitive across the reported benchmarks.

Table R10. Cross-dataset evaluation on Sintel (train) and KITTI-2015 (train), without target-domain flow fine-tuning. C: FlyingChairs; T: FlyingThings3D; TA: TartanAir. Pretrained visual/depth representations and training budgets remain method-specific.

MethodVenueFlow-training dataCleanFinalKITTI EPEKITTI Fl-all
FlowSeek-LICCV 2025C+T1.072.213.8212.5
FlowSeek-SICCV 2025TA+C+T1.052.373.3211.0
FlowSeek-LICCV 2025TA+C+T1.032.183.3111.2
Chanda and KumarCVPRW 2026C+T1.462.81––
FlowItBMVC 2026TA+C+T0.902.184.8117.3
MegaFlowECCV 2026C+TA+T0.892.073.0010.6
TDFlowOursC+T0.872.092.879.30

Sources: FlowSeek, Table 3 [5]; Chanda and Kumar, Table 1 [4]; FlowIt v2, Table 2 [6]; and MegaFlow, Table 1 [8]. CVPRW denotes the CVPR 2026 ViSCALE Workshop. Dashes mean that no value is taken from the cited table. The target datasets' training splits are used for evaluation because their ground truth is public; their flow labels are not used for target-domain fine-tuning in this comparison. MegaFlow (ECCV 2026) additionally uses pretrained VGGT/DINOv2 representations. We use its main comparison Table 1, not the slightly different values in its ablation Table 6.

Under C+T optical-flow training without target-domain flow fine-tuning, TDFlow achieves competitive cross-dataset performance against the listed recent configurations, including methods using additional TartanAir training data. It has lower errors than the listed FlowSeek and FlowIt results, and lower Sintel EPE than Chanda and Kumar. Relative to MegaFlow, TDFlow has lower Clean EPE and KITTI errors, while MegaFlow has slightly lower Final EPE (2.07 vs. 2.09).

References

[3] Ao Luo et al. Optical Flow Matching: Reframing Optical Flow as Continuous Transport Dynamics. CVPR, 2026.

[4] Praroop Chanda and Suryansh Kumar. Rethinking Dense Optical Flow without Test-Time Scaling. CVPR Workshops, ViSCALE, 2026.

[5] Matteo Poggi and Fabio Tosi. FlowSeek: Optical Flow Made Easier with Depth Foundation Models and Motion Bases. ICCV, 2025.

[6] Sadra Safadoust et al. FlowIt: Global Matching via Hierarchical Transformers and Optimal Transport for Optical Flow. BMVC, 2026.

[7] Vladislav Bargatin et al. FreeFlow: A Bias-Free Hierarchical Transformer for Optical Flow Estimation. ECCV, 2026.

[8] Dingxi Zhang, Fangjinhua Wang, Marc Pollefeys, and Haofei Xu. MegaFlow: Zero-Shot Large Displacement Optical Flow. ECCV, 2026.

AuthorsOpenReview ↗

Response to Reviewer f4Gx (R3)

Q3. Is the gain mainly from CFSW, given the cited 3.02 EPE?

The 3.02 KITTI EPE row in Table 5 is “Without CFSW,” not the no-noise control. Anchor only omits the final refinement update; w/o TDP retains that additional WAFT update and CFSW decoder but disables TDP. The independently retrained w/o TDP model therefore provides the direct additional-update control:

Table R11. Components and the independently retrained w/o TDP control, which retains the extra refinement update. C+T optical-flow training.

ConfigurationSintel Clean EPESintel Final EPEKITTI EPEKITTI Fl-all (%)
Anchor only1.032.323.2310.75
w/o TDP (refinement retained)1.012.293.2110.63
Without CFSW0.932.143.029.87
Full TDFlow0.872.092.879.30

Relative to the retrained w/o TDP control, TDFlow reduces Sintel Clean/Final EPE by 13.9%/8.7%, KITTI EPE by 10.6%, and KITTI Fl-all by 1.33 percentage points (12.5% relative). Since both configurations retain the final refinement update, the gain goes beyond adding that update alone. Removing CFSW also hurts accuracy, supporting complementary contributions from feature alignment and TDP.

The fine-tuned comparison in Table R1 shows the same direction, with Sintel EPE decreasing from 1.08/2.36 to 0.96/1.84; the fine-tuned KITTI gains are smaller. The ten-seed experiment measures inference-noise variation at a fixed checkpoint, so the reductions above are empirical effect sizes rather than training-seed significance estimates.

The OFM [3] comparison in Q2 provides context from a recent flow-matching method, while the independently retrained deterministic control isolates the benefit of our formulation within the retained refinement architecture. These results support its effectiveness, rather than establish diffusion as uniquely necessary.

References

[3] Ao Luo et al. Optical Flow Matching: Reframing Optical Flow as Continuous Transport Dynamics. CVPR, 2026.

Reviewer_CCEhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces a diffusion-based optical flow method. Given an initial estimate of optical flow from an off-the-shelf or lightweight model, a following diffusion-model (as a refinement model) refines the initial estimate via diffusion process. The diffusion-based refinement starts from a truncated noise schedule and needs only a few steps (eg, even one step) as the initial estimation already has a good starting point.

The proposed method shows good empirical results (state-of-the-art or competitive accuracy) on public benchmark datasets.

Critical strengths
  • Practical use of diffusion as a refinement module

    Instead of starting from a full-noise image, the method shows that the diffusion-based method can be used as a refinement module that follows after any off-the-shelf methods (ie, initial estimate). This is a practical design that bridges a regression-based method and a generative (eg, diffusion) method with low computational addition.

  • Competitive/state-of-the-art accuracy on public benchmark

    Optical flow is a well-studied problem, and benchmarks tend to be saturated. The proposed method shows very competitive accuracy (or even state-of-the-art on some metrics) on public benchmark datatsets.

  • Clearly written

    The paper is clearly written, self-contained, and easy to follow. Most questions appearing during reading the paper are answered by reading the paper.

Critical weaknesses
  • Discussion on the number of denoising steps

    It will be good to include an in-depth discussion on why the method works with only a few denoising steps.

    • Unlike generation approaches which may need many denoising steps to get fine details, it's known in the literature that discriminative tasks (eg, optical flow, depth, etc) need only few denoising steps to get better accuracy. Usually more denoising steps hurt accuracy: blurry results give better accuracy on metrics than the results with fine details despite of qualitatively looking better. In Table 5, can the paper provide accuracy with more denoising steps, such as 3, 4, ..., up to TtruncT_{trunc} and include discussions? It will be good to know if the 1-denoising-step benefit is from the strong conditioning signal from the initial flow estimates, or if this is from just properties of diffusion on discriminative tasks.

    • Also, in Table 5, can the paper provide accuracy with more steps on TtruncT_{trunc}, such as 40 and 50?

    • L15: the abstract says the large number of denoising steps is impractical; but FlowDiffuser only uses 4 steps, which does not sound very large.

  • Possible unfair comparison on the runtime

    In Table 1 and 11, does the runtime/latency include run-time that takes to get the initial estimate? It would be unfair if the runtime for the initial estimate was discarded.

  • Possible incorrect figure (Fig. 1)

    • Fig 1(a): why is the sparse GT from KITTI is 90-degree rotated? Also, how does the method output sparse GT from the denoising process? It looks like the denoised GTs (or estimates) in the figure are not the actual denoised output but simulated mock-up images.
    • Fig 1(b), previous diffusion models: the denoised intermediate results look already clean, which looks wrong. It looks like also just mock-up images instead of real experiment results. It is highly recommended to explain how these images are actually generated/obtained.
  • Possible overclaiming novelty: At L161 (... we propose a novel truncated diffusion policy ...)

    It's a bit confusing if it's right to call the proposed method as a novel diffusion policy, like the one commonly used in the robotics literature. The proposed method considered as: given a strong conditioning signal (initial estimate), the paper proposes a diffusion refinement module with its tuned denoising schedule. In this regards, the novelty sounds a bit overclaimed.


Given these weaknesses, I would rate the paper to Borderline Reject (closer to Clear Reject). But, I am happy to increase the rating if the concerns are resolved. The paper shows strong empirical results.

Questions for authors
  • Can the paper provide any qualitative results on where the accuracy improvement comes from? One could visualize both optical flow and error maps of initial estimate (ie, initial flow anchor) and the diffusion-refined output, and discuss where the accuracy gains usually come from.
5 comments
AuthorsOpenReview ↗

Response to Reviewer CCEh (R2)

Q1. Would larger truncation timesteps and more denoising evaluations improve accuracy?

We agree that few-step denoising can be effective for dense prediction and that additional steps need not improve endpoint accuracy; single-step depth prediction provides a related example (DepthMaster [9]). Our contribution lies in the anchor-conditioned truncated formulation and its benefit over independently retrained deterministic refinement.

The maximum forward-noising index is set to Ttrunc=20T_{\mathrm{trunc}}=20. Training samples i∈{1,…,20}i\in\lbrace 1,\ldots,20\rbrace and perturbs the anchor as

τ(i)=αˉi,τanchor+1−αˉi,ϵ,ϵ∼N(0,I).\boldsymbol{\tau}^{(i)} = \sqrt{\bar{\alpha}_i}\\,\boldsymbol{\tau}_{\mathrm{anchor}} + \sqrt{1-\bar{\alpha}_i}\\,\boldsymbol{\epsilon}, \qquad \boldsymbol{\epsilon}\sim\mathcal{N}(0,I).

The noise schedule determines the perturbation magnitude. Separately, K=1K=1 denotes one denoiser evaluation at inference. We extend the original ablation with Ttrunc∈{40,50}T_{\mathrm{trunc}}\in\lbrace 40,50\rbrace and K∈{2,3}K\in\lbrace 2,3\rbrace:

Table R3. Truncation and denoiser-evaluation ablation under C+T training. The last four rows are additional experiments. Lower is better.

TtruncT_{\mathrm{trunc}}KKSintel Clean EPESintel Final EPEKITTI EPEKITTI Fl-all (%)
1010.922.142.929.90
2010.872.092.879.30
2020.872.092.889.30
2030.882.102.879.30
3010.882.102.909.60
4020.892.112.939.72
4030.882.132.959.75
5020.902.132.929.74
5030.892.122.979.78

None of the newly tested settings improves upon the default Ttrunc=20,K=1T_{\mathrm{trunc}}=20,K=1. At Ttrunc=40T_{\mathrm{trunc}}=40 or 50, increasing KK from two to three produces mixed changes. Together with the K=1,2,3K=1,2,3 comparison at Ttrunc=20T_{\mathrm{trunc}}=20, these results support the default accuracy–computation trade-off among the tested settings; K=1K=1 at Ttrunc=40,50T_{\mathrm{trunc}}=40,50 was not tested.

We also agree that endpoint accuracy and visual detail are different criteria. The new predictions, error maps, and boundary crops in manuscript Figures 5 and 7 complement the EPE results for the illustrated examples; they do not establish a general relationship between denoising count and boundary fidelity.

Q2. Does one-step recovery work only because the anchor is exceptionally strong?

The compatibility study in the submitted Table 4 evaluates less accurate anchor estimators as well. On KITTI, GMA improves from 4.69/17.10 to 4.29/15.20 EPE/Fl-all, KPA-Flow from 4.46/15.90 to 3.92/14.50, and SKFlow from 4.27/15.50 to 3.77/13.70. Thus, the refinement benefit is not restricted to an exceptionally accurate WAFT-style initialization.

An image-conditioned anchor supplies structured motion evidence, and CFSW aligns features using the perturbed state for residual recovery. These results support refinement across the tested anchor qualities; the separately trained pure-noise control in our response to R3 further examines the anchor's contribution.

Q3. Why is efficiency a concern when FlowDiffuser uses only four steps?

We agree that four steps alone are not a large sampling count. The practical concern is the full-model latency. The submitted Table 1 reports 290 ms for FlowDiffuser and 69 ms for TDFlow on an RTX 4090, batch size 1, with 376×1248376\times1248 inputs. The corresponding reciprocal rates are approximately 3.4 and 14.5 flow estimates per second:

Table R4. Practical interpretation of the reported full-model timings. Rates are computed as 1000 / latency (ms), rather than measured streaming throughput.

MethodLatency (ms)Estimates/sWithin 100 ms
FlowDiffuser2903.4No
TDFlow6914.5Yes

KITTI records synchronized data at 10 Hz [10], corresponding to a 100 ms input interval. TDFlow's model computation fits within this budget under the reported profiling setup. The 69 ms includes feature extraction, all four anchor updates, and final refinement/output generation. Data loading and visualization are excluded, and complete application latency also depends on downstream processing.

The reported 290/69≃4.2×290/69\simeq4.2\times full-model speedup reflects both fewer denoiser evaluations (four versus one) and lower per-evaluation cost (53 versus 9 ms).

References

[9] Ziyang Song et al. DepthMaster: Taming Diffusion Models for Monocular Depth Estimation. arXiv:2501.02576, v2, 2026.

[10] The KITTI Vision Benchmark Suite.

AuthorsOpenReview ↗

Response to Reviewer CCEh (R2)

Q4. Are Figure 1's thumbnails actual denoising outputs?

Figure 1(a) visualizes the anchor flow after adding noise during training. The 90-degree rotation is solely for layout. This panel depicts the perturbed anchor state, not sparse ground truth generated by the denoiser; TDFlow predicts dense optical flow. Figure 1(b) shows actual intermediate outputs from the model's denoising process. Specifically, “(A) Previous diffusion models” shows FlowDiffuser [11] results. A visually clean intermediate prediction does not necessarily imply accurate optical flow.

In the revised manuscript, we have added Figure 5 (p. 9) showing actual Anchor Flow, Refined Flow, GT, and error maps, together with Table 6 (p. 10) reporting EPE for the three visualized Sintel examples. We have also added corresponding boundary crops in Appendix G.1, Figure 7 (p. 25). In this response, Table R5 reproduces the results from the newly added manuscript table, while Table R6 provides additional statistics for 62 saved Anchor/Refined pairs.

Table R5. Per-example EPE of matched Anchor/Refined outputs.

Sintel Clean frame tripletAnchor EPERefined EPERelative reduction
0044–0045–00466.693.3250.4%
0043–0044–00453.101.6945.5%
0048–0049–00505.103.6628.2%

We additionally summarize Anchor/Refined pairs with matching GT renderings: 49 Clean and 13 Final outputs.

Table R6. Paired statistics on Sintel. EPE is in pixels.

SubsetNAnchorRefinedΔEPEReductionImproved
Clean & Final620.8780.6680.21023.9%62/62

Together, the added visualizations and paired statistics demonstrate actual dense-flow refinement, with mean EPE decreasing from 0.878 to 0.668 pixels (23.9%) on the 62 saved Sintel pairs.

Q5. Is “diffusion policy” an overstated or confusing description?

Here, “policy” refers to an anchor-guided truncated denoising strategy, following the terminology of DiffusionDrive [1]. DiffusionDet [2] also demonstrates how a visual perception task can be formulated as conditional denoising. In TDFlow, this strategy is implemented as noise-conditioned residual refinement with an image-conditioned dense anchor and state-conditioned feature warping. The independently retrained additional-update control (Table R2) supports its benefit over deterministic refinement within the retained refinement architecture.

References

[1] Bencheng Liao et al. DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving. CVPR, 2025.

[2] Shoufa Chen et al. DiffusionDet: Diffusion Model for Object Detection. ICCV, 2023.

[11] Ao Luo et al. FlowDiffuser: Advancing Optical Flow Estimation with Diffusion Models. CVPR, 2024.

Dear Reviewer CCEh,

Thank you again for taking the time to review our paper and share your comments. We hope our earlier response has addressed your concerns. Since the ICLR discussion period ends soon, we would greatly appreciate your feedback on our response, and we are happy to answer any further questions.

Best regards,

The Authors

Reviewer_vKu7OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents TDFlow, a supervised optical flow method. The model first produces an initial flow anchor using a network similar to WAFT, and then refines this estimate with a truncated diffusion policy. A Conditional Feature-Space Warping (CFSW) decoder is introduced to improve the diffusion process. The method reduces the high computational cost of diffusion-based flow estimation, requiring only a single denoising step. Experiments show promising results.

Critical strengths
  1. The methodology is clearly presented and generally sound. The model architecture and experimental settings are described in detail, which supports reproducibility.
  2. The experiments are extensive, and the figures are well presented.
Critical weaknesses
  1. Although the paper positions its contribution within the line of diffusion-based methods, the network appears to be more of a simple adaptation of WAFT.
    • The initial flow anchor generation corresponds exactly to WAFT, except that it uses four recurrent updates instead of five. The truncated diffusion stage consists of only one denoising step, and its denoiser is also a DPT decoder, the same structure used in WAFT. Therefore, the method essentially modifies WAFT by replacing its last recurrent update with a diffusion step of the same structure, while adding noise and the CFSW module in between.
    • The stated motivation is to reduce the heavy computational cost of diffusion-based methods. However, the method appears to be derived from WAFT rather than from existing diffusion-based methods. In addition, with only a single denoising step, the model behaves less like a typical diffusion model. As a result, the positioning of the method in the paper is somewhat misaligned with what it actually is.
  2. Some mathematical notation is not rigorous, and the symbols should be made consistent.
    • In Eq. (3), the flow is denoted by both τ\tau and ff, where τ\tau is not bold but ff is.
    • Line 253: in Warp(τ(i))p\text{Warp}(\tau^{(i)})_p, the subscript pp placed after the parentheses is not standard mathematical notation.
    • Line 247 uses PP to denote the pixel grid, while Line 266 uses ΩΩ.
Questions for authors
  1. Section 4.4 examines whether the improvement mainly comes from a stronger initialization or from the proposed anchor-guided truncated diffusion policy. This is an important question, since the anchor is essentially the output of the first four recurrent updates of WAFT. The anchor itself may therefore already be very accurate, and the subsequent single diffusion step may provide only a marginal refinement, or an improvement that is inconsistent or not statistically significant. However, the presented results in the main body do not fully address this question. Specifically:
    • Was TDFlow w/o TDP retrained, or was it obtained by running the trained TDFlow at inference time without TDP?
    • In either case, could the authors provide a direct comparison between TDFlow and TDFlow w/o TDP in the format of Table 3? Table 5 and Table 8 only report results under C+T training, which is not the standard benchmark setting.
4 comments
AuthorsOpenReview ↗

Response to Reviewer vKu7 (R1)

Q1. Is TDFlow simply a WAFT adaptation with its final update replaced?

Our contribution combines anchor-conditioned truncated denoising with state-conditioned feature alignment for efficient dense-flow recovery. We build on WAFT for the initial estimator and reuse DPT-based components. The methodological change is the final update's learning formulation: an image-conditioned anchor defines an informative starting point, training perturbs this anchor over a truncated range of noise levels, and CFSW aligns features using the perturbed flow state for noise-conditioned residual recovery. This design targets the accuracy–computation trade-off of dense-flow estimation.

The independently retrained w/o TDP model is the additional-WAFT-update control. It retains the fifth refinement update and CFSW decoder but removes TDP and noise injection during both training and inference. Anchor only instead stops after four WAFT updates and omits this final stage. Comparing full TDFlow with w/o TDP therefore isolates the benefit of TDP within the retained refinement architecture.

Under C+T training, the retrained control obtains 1.01/2.29 Sintel Clean/Final EPE and 3.21/10.63 KITTI EPE/Fl-all, compared with 0.87/2.09 and 2.87/9.30 for TDFlow (Table R2). These gains show that the extra deterministic update alone does not account for the improvement. DiffusionDet [2] adopts the detection-decoder design of Sparse R-CNN and retains its six-stage cascade, while reformulating detection as image-conditioned recovery from noisy boxes. This provides a concrete precedent for introducing a diffusion-based learning formulation within an established decoder architecture.

Q2. Why call the method diffusion-based if inference uses one denoising step?

The aim is to combine noise-conditioned learning around an informative flow prior with efficient dense-flow recovery. Training samples Gaussian perturbations at multiple timesteps within the truncated interval and learns image-conditioned recovery from these states. At inference, K=1K=1 denotes one learned denoiser evaluation; the four preceding anchor updates remain part of the complete model.

We describe this formulation as anchor-guided truncated denoising. Its benefit is supported by the independently retrained deterministic control and the denoiser-evaluation ablation in our response to R2. DiffusionDrive [1] trains a truncated diffusion process around prior trajectory anchors and reports effective planning even with a single denoising step. This supports using informative initialization and a truncated noise schedule for efficient denoising. TDFlow adapts this principle to dense optical flow through an image-conditioned flow anchor and state-conditioned feature alignment.

Q3. Was w/o TDP retrained, and does it help after benchmark fine-tuning?

Yes. The w/o TDP control described in Q1 was independently retrained. The tables below compare it with full TDFlow in both evaluation settings. The original submission's earlier w/o-TDP diagnostic is documented separately in the manuscript's legacy-ablation discussion.

Tables R1 and R2 give the direct comparison after dataset-specific fine-tuning on hidden test splits and after C+T optical-flow training without target-dataset flow fine-tuning, respectively.

Table R1. Fine-tuned benchmark comparison. KITTI columns are outlier rates (%), not EPE. Lower is better.

MethodSintel Clean EPESintel Final EPEKITTI Fl-allKITTI Fl-noc
TDFlow w/o TDP (retrained)1.082.363.412.03
TDFlow0.961.843.342.01

Table R2. Zero-shot evaluation after C+T optical-flow training. Lower is better.

MethodSintel Clean EPESintel Final EPEKITTI EPEKITTI Fl-all (%)
TDFlow w/o TDP (retrained)1.012.293.2110.63
TDFlow0.872.092.879.30

TDFlow improves every reported metric in both settings. Fine-tuned Sintel Clean/Final EPE decreases by 11.1%/22.0% relative to the retrained no-TDP control. Under C+T training, KITTI EPE decreases by 10.6%, and Fl-all decreases by 1.33 percentage points (12.5% relative). The fine-tuned KITTI gains are smaller. These results support a benefit beyond adding the same deterministic refinement architecture, without implying the same gain magnitude across datasets.

Q4. Mathematical notation and minimal corrections.

Thank you for pointing out these notation issues. We have revised the manuscript to standardize the notation for optical flow and residuals, clarify the Warp indexing, and explicitly define PP and Ω\Omega.

References

[1] Bencheng Liao et al. DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving. CVPR, 2025.

[2] Shoufa Chen et al. DiffusionDet: Diffusion Model for Object Detection. ICCV, 2023.

Dear Reviewer vKu7,

Thank you again for taking the time to review our paper and share your comments. We hope our earlier response has addressed your concerns. Since the ICLR discussion period ends soon, we would greatly appreciate your feedback on our response, and we are happy to answer any further questions.

Best regards,

The Authors

Thank you for the responses, and sorry for the delay. I was very busy.

Q1: The authors acknowledge that the architecture is largely inherited from WAFT, with the final iteration replaced by a CFSW-based diffusion step. Novelty is therefore not a strong point of this work.

Q2: Although training follows a diffusion formulation, inference is effective only with K=1K=1; more denoising steps bring no benefit. The model effectively learns a one-step refinement rather than step-by-step denoising, so positioning it as diffusion-based is inaccurate. This also conflicts with the stated motivation of reducing the computational cost of diffusion-based methods.

Q3: My question concerns a model that is not retrained. A retrained w/o TDP model naturally performs similarly to WAFT (Table R1 vs. the reported WAFT results), which is uninformative. As mentioned in Question 1, the second bullet point, I would expect an evaluation on the end-to-end trained TDFlow using only its anchor (without the final step) on the benchmark test sets. Table 5 reports “anchor only” results only under the C+T zero-shot setting, which is not the standard setting for this benchmark.

Q4: Resolved.

Q1. Contribution beyond the inherited architecture.

Thank you for carefully reading our paper and rebuttal. Our contribution is the anchor-conditioned truncated diffusion formulation with state-conditioned feature alignment, whose benefit is demonstrated against an independently retrained deterministic control retaining the same refinement architecture.

Q2. Does single-step inference conflict with the diffusion formulation?

Thank you for carefully reading our paper and rebuttal. Two-step denoising remains comparably effective for both anchors (Sintel Clean/Final EPE: 0.87/2.09 for both K=1 and K=2 with the default anchor), while single-step inference delivers the intended efficiency of our scheduled, timestep-conditioned diffusion formulation.

Q3. Does final refinement improve benchmark results without retraining?

Thank you for clarifying. We evaluate the anchor output of the end-to-end trained, benchmark-finetuned TDFlow by bypassing its final refinement stage, without retraining. For each benchmark, both configurations use the same checkpoint and evaluation protocol.

Table 1. Effect of final refinement on benchmark test sets using the same checkpoint without retraining. KITTI columns are outlier rates (%). Lower is better.

Inference configurationSintel Clean EPESintel Final EPEKITTI Fl-all (%)KITTI Fl-noc (%)
TDFlow - Anchor only (final refinement bypassed)1.132.373.732.47
TDFlow - Full model0.961.843.342.01

Final refinement reduces Sintel Clean/Final EPE by 0.17/0.53 pixels and KITTI Fl-all/Fl-noc by 0.39/0.46 percentage points. We have added these results to Appendix E.1, Table 10 (p. 21), of the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.