acceptodds
Under review as a conference paper at ICLR 2027

CorrFlow: Self-Correcting Flow Matching

18 views

Abstract

Flow-matching models generate samples through iterative velocity predictions, but typically do not revisit these predictions using the intermediate representations that produced them. We introduce CorrFlow, a self-correcting flow-matching framework that reuses this information through a lightweight refiner. Under rectified-flow interpolation, the error of the one-step clean prediction equals the realized velocity residual scaled by the noise level. Motivated by this identity, the refiner learns to predict residual corrections from the base model's clean prediction, conditioning, noise level and hidden states from the same forward pass. The correction is added to the base velocity before each sampling update, so subsequent predictions are made from the corrected state. This requires no additional base-model evaluations and only one lightweight refiner evaluation per step. Our theoretical analyses establish conditions under which feature reuse outperforms raw-input affine predictors and refiner training achieves greater residual-risk reduction than continued base-model training under a matched training compute budget. Experiments on class-conditional image generation, text-to-video generation and image-to-video generation show consistent improvements over the corresponding base models, including gains in visual quality and physical consistency.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Flow Matching

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_k4GGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CorrFlow adds a learned refiner to a flow-matching model to correct its velocity predictions. The refiner uses the base model’s predictions and intermediate features, and is motivated by a connection between clean-sample prediction errors and velocity residuals. The paper provides theoretical analyses and evaluates the method on image generation, text-to-video generation, and image-to-video generation.

Critical strengths
  • The paper is generally well written and easy to follow. The motivation and proposed method are clearly explained, and the appendix provides detailed training and evaluation settings that help clarify the implementation.

  • The qualitative video results are promising. The examples presented for both text-to-video and image-to-video generation look visually convincing, and the comparisons with the baselines suggest improvements in object consistency and the plausibility of the depicted interactions.

Critical weaknesses
  • Limited methodological contribution. The method learns velocity residuals from the base model's clean prediction and reused hidden states, then applies these corrections during sampling. This is a coherent design, but I find the methodological advance incremental. The theoretical analyses provide conditional support for refinement, while offering, in my view, limited additional insight into its practical advantages. The modest ImageNet gains in Table 7 also make the overall contribution less compelling to me.

  • Unclear benefit relative to the additional cost. My main concern is whether the improvements justify the extra computation. Reusing the base model's prediction and intermediate features avoids additional base-model evaluations, but the method still requires an extra refiner evaluation at active sampling steps. The image comparisons with equal training epochs and the small-scale video ablation are useful, but they do not establish an empirical advantage at a comparable computational budget. Although the appendix reports absolute training times, it does not quantify the incremental training cost or end-to-end inference overhead relative to the corresponding base model alone.

Questions for authors
  1. What is the main methodological insight gained by predicting corrections from the clean-sample estimate and reused hidden states? Which theoretical or experimental results best demonstrate this insight?

  2. Could you report the incremental training cost and end-to-end generation time relative to the corresponding base model alone, under the same hardware and settings? An empirical comparison with continued base-model training at comparable training cost would also help assess whether the quality improvements justify the additional computation.

4 comments

We thank the reviewer for the constructive review.

[W1] Methodological Novelty and Broad Validation

Firstly, we respectfully disagree that CorrFlow's contribution is incremental. To our knowledge, prior work has not introduced a dedicated refiner that jointly uses a flow model's clean prediction and hidden states from the same forward pass to learn velocity residuals and correct the velocity within the sampling loop. The clean-prediction error identity provides principled supervision. CorrFlow enables the generator to correct its own predictions at each sampling step without additional base-model evaluations.

Secondly, our revised theory explains why this mechanism helps. It identifies learnable residual directions near base-model stationarity, establishes finite-data advantages over direct velocity prediction, and extends these guarantees to moving checkpoints. Our nonlinear AdamW analysis provides sufficient conditions for outperforming continued full-base training at matched compute, while the endpoint analysis connects training risk to generated distributions. An explicit nonlinear construction using our architecture and AdamW verifies these conditions. Experimentally, residual prediction achieves a Quality Score of 0.8029 versus 0.6956 for direct prediction, and CorrFlow outperforms capacity- and compute-matched baselines on all six mean metrics.

Thirdly, we added ImageNet-256 experiments on the original RAE using the RAEv2 protocol. Each Off/On pair shares the same adapted base checkpoint, initial noise and class labels. Enabling the refiner improves all five metrics across both training budgets, with and without AutoGuidance (AG). FID and representation distances decrease while IS increases, demonstrating consistent improvements.

EpochsGuidanceRefinerFID ↓IS ↑FD_r DINOv2 ↓FD_r ConvNeXt ↓FD_r CLIP ↓
120NoneOff2.0325215.203.7082.72010.098
80+40NoneOn2.0106215.823.6682.69510.055
120AGOff1.5550232.702.6122.1888.538
80+40AGOn1.5430233.252.5772.1658.479
160NoneOff1.9204216.653.5512.6169.892
80+80NoneOn1.9102217.183.5352.6059.870
160AGOff1.4677233.352.4852.1088.317
80+80AGOn1.4579233.492.4732.0968.293

These image improvements are complemented by larger gains in video generation, where coherent motion, object interactions and physical consistency pose greater challenges, demonstrating CorrFlow's value beyond single-image generation.

[Q1] Theoretical and Empirical Evidence for Feature Reuse

Our central insight is that a model can improve its prediction by revisiting both its current answer and the intermediate representations that produced it. Under rectified flow, x^0−x0=σr\hat{x}_0-x_0=\sigma r links clean-prediction error to velocity residual rr. The clean estimate provides a prediction to review, while reused hidden states offer contextual features for estimating corrections without additional base-model evaluations.

Theoretical evidence: Our frozen-feature theorem constructs an example where a one-parameter readout of reused features exactly recovers the conditional residual, while raw-input affine predictors retain positive excess risk. Revised finite-data and moving-checkpoint theorems quantify statistical, optimization and feature-drift penalties.

Experimental evidence: Removing text conditioning while retaining hidden-state cross-attention achieves 0.7958, above the no-refiner baseline (0.7894). Removing cross-attention, eliminating hidden-states, reduces the score to 0.7819. Replacing nonlinear hidden-state fusion with weighted averaging lowers it to 0.7914. These comparisons support the value of reused representations and cross-attention.

Related work demonstrates diffusion model's hidden-state feature reuse: Gen4U[1] uses diffusion features for understanding, Next Forcing[2] fuses multilayer features for future prediction, DiT4DiT[3] conditions an action DiT on video features, and VPP[4] uses SVD features for action diffusion. CorrFlow uniquely applies this principle to correcting the generator's own velocity within the sampling loop.

[1] Gen4U: Unifying Video Generation and Understanding via Diffusion, arXiv 2026.

[2] Next Forcing: Causal World Modeling with Multi-Chunk Prediction, NeurIPS 2026.

[3] DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control, arXiv 2026.

[4] Video Prediction Policy: A Generalist Robot Policy with Predictive Visual Representations, ICML 2025.

[Q2 & W2] Matched-Compute Performance and Inference Efficiency

We agree that equal training epochs alone do not establish computational efficiency. First, we now quantify the incremental training cost. Including forward, backward and gradient-checkpoint recomputation, the base DiT costs 1.2324 PFLOPs per training example, while the refiner adds approximately 0.181 PFLOPs (+14.7%). Across five epochs on 66,572 clips, this increases adaptation computation from 410.2 to 470.4 EFLOPs, an additional 60.2 EFLOPs.

Second, we have added a full-scale Wan2.1-1.3B-FT-Long baseline that fine-tunes the same pretrained base model on our complete training dataset with the same base objective and optimization settings. FT-Long allocates the refiner’s training budget to 3,054 additional base-model updates, extending training from 20,805 to 23,859 steps and matching CorrFlow’s total adaptation cost of approximately 470.4 EFLOPs. The results are below, with higher values indicating better performance on all metrics.

BenchmarkMetricWan2.1-1.3BFT-LongOurs
VideoPhy2 ALLPhysical Commonsense57.861.965.7
Semantic Adherence30.028.031.3
Joint24.824.227.0
FinePhyEvalObject0.840.880.90
Action0.490.500.55
Physics0.360.390.44
Overall0.540.560.60
VBenchSubject Consistency91.8396.4296.68
Temporal Flickering99.1799.6199.81
Motion Smoothness96.5198.7998.85
Aesthetic Quality57.2459.2561.65
Imaging Quality59.1466.9668.66
Object Class76.0976.5882.81
Spatial Relationship62.3772.3877.88
Temporal Style22.2323.4324.05
Quality Score79.8182.8384.69
Overall Score76.9379.4881.86

At matched training computation, CorrFlow exceeds FT-Long on every reported metric, including gains of 2.8 percentage points in VideoPhy2 Joint, 0.04 in FinePhyEval Overall and 2.38 points in VBench Overall. Longer base-model training improves visual quality but does not improve VideoPhy2 Joint over the pretrained baseline. These results support allocating the additional training budget to residual refinement.

Third, we have added per-video computation, generation latency and peak GPU memory comparisons for both fp32 and automatic mixed-precision (autocast) refiners, keeping base-model precision and sampling settings fixed. For text-to-video generation at 81×480×83281\times480\times832 (frames × height × width), CorrFlow adds approximately 6% computation and 5% latency with autocast, compared with 53% latency overhead in fp32. Relative costs below are normalized to Wan2.1-1.3B.

T2V methodPFLOPs/videoRelative computeLatency (s)Relative latencyPeak memory (GiB)
Wan2.1-1.3B28.91.00×89.3 ± 2.41.00×15.4
CorrFlow-1.3B (fp32 refiner)30.71.06×136.3 ± 0.21.53×17.8
CorrFlow-1.3B (autocast refiner)30.71.06×94.2 ± 0.21.05×17.6
DreamWorld-1.3B70.92.45×238.2 ± 0.52.67×21.2
VPT-1.3B56.91.97×193.4 ± 0.12.17×19.9
VideoREPA-5B88.43.06×294.0 ± 1.13.29×25.1
PhysRAG-5B10.80.37×36.1 ± 0.10.40×34.5
Wan2.1-14B168.15.82×475.2 ± 0.15.32×46.6

CorrFlow requires less computation, lower latency and less peak memory than DreamWorld, VPT, VideoREPA and Wan2.1-14B. PhysRAG is faster and uses fewer FLOPs but requires more memory.

For image-to-video generation at 81×704×124881\times704\times1248, CorrFlow adds approximately 7% computation and 8% latency with autocast, compared with 59% latency overhead in fp32. The autocast variant has lower latency and peak memory than all compared I2V adaptation methods. Relative costs below are normalized to Wan2.2-5B.

I2V methodPFLOPs/videoRelative computeLatency (s)Relative latencyPeak memory (GiB)
Wan2.2-5B30.61.00×83.6 ± 2.41.00×25.9
CorrFlow-5B (fp32 refiner)32.61.07×132.7 ± 0.11.59×29.2
CorrFlow-5B (autocast refiner)32.61.07×90.0 ± 0.21.08×29.2
WMReward-5B (guidance only)62.72.05×230.7 ± 0.72.76×115.2
PhysRAG-5B30.81.01×98.5 ± 0.01.18×48.9
VLIPP-5B (30 steps)223.47.30×713.2 ± 0.18.53×45.4

These results demonstrate improved quality at matched training computation, with approximately 5–8% additional generation latency when using an autocast refiner.

Dear Reviewer,

Thank you for your valuable feedback. In response to your concerns, we have provided further clarifications and additional experimental analyses in our rebuttal.

We would greatly appreciate it if you could revisit our responses and let us know whether they have adequately addressed your concerns. If your concerns have been resolved, we sincerely hope you would consider raising your rating.

Once again, we sincerely thank you for your time, effort, and thoughtful review of our work.

Best,

Authors

Dear Reviewer,

With about 20 hours remaining before the discussion period closes, we would greatly appreciate it if you could revisit our rebuttal and additional experimental results. If your concerns have been adequately addressed, we sincerely hope you would consider raising your rating.

Thank you again for your time and thoughtful consideration.

Best,
Authors

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

The paper adds a small refiner to a flow-matching model. At each step, it uses the base model’s clean prediction and hidden states to predict the velocity error, then adds this fix to the base velocity. Tests on ImageNet, text-to-video, and image-to-video show gains over the base models.

Critical strengths

1.Simple and general idea. The refiner is a small add-on that needs no extra base-model passes and works for both image and video models. This makes it easy to use in practice.

2.Gains on many benchmarks. The video results improve on VBench, VideoPhy2, FinePhyEval, and Physics-IQ over the base models. This shows the idea may help physical consistency, if the gains hold under fair baselines.

Critical weaknesses

1.Missing key baseline. In the video tests, the base model is fine-tuned on new physics data (WISA-80K, OpenVidHD), but it is compared to the original Wan models. The gains may come from the data, not the refiner. A full-scale “same data, no refiner” baseline is needed. No new information. The refiner only sees what the base model already computed, so it is learning the same target as the base model. It is basically a second network with 10% more parameters. A baseline with a slightly bigger base model or more training is needed to show the refiner itself helps.

2.Small and mixed ImageNet gains. The gains over RAEv2 are tiny, and FID gets worse (1.034 to 1.059). FID is left out of the main Table 1 and only shown in the appendix. This makes the image results unconvincing.

3.Weak theory. The theorems use hand-built toy examples, for example a 1-D XOR case, or a base model given a weaker Jacobian by design. They do not show the method should beat simply training the base model longer in real settings.

4.Weak ablation and no error bars. The ablation uses only 1,000 training clips and 50 test clips, and the gaps are small. No results report variance across runs.

5.Unclear setup. The tables mix published numbers with local runs under different settings, and WMReward is run without Best-of-N. The gain κ and the σ range are hand-picked with no sensitivity study, and inference cost is not reported.

Questions for authors

1.Can you report a full-scale baseline that fine-tunes Wan on the same data with no refiner? This is a clarification, and it is key to my score. 2.How does CorrFlow compare to a base model with about 10% more parameters, or one trained longer at the same total compute? This is a clarification. 3.How were κ and the σ range chosen, and how do results change for other values? This is a clarification.

5 comments

Thanks for the reviewer's constructive review.

[W1 & Q1] Matched-Compute and Capacity Baseline Comparisons

We added the requested full-scale Wan2.1-1.3B-FT-Long baseline, which fine-tunes Wan without a refiner on exactly the same WISA-80K and OpenVidHD training data. It retains the base objective and optimization settings and trains longer to match the total training FLOPs of our base-plus-refiner adaptation. Results on VideoPhy2 ALL, FinePhyEval and VBench follow (higher is better).

BenchmarkMetricWan2.1-1.3BFT-LongOurs
VideoPhy2 ALLPhysical Commonsense57.861.965.7
Semantic Adherence30.028.031.3
Joint24.824.227.0
FinePhyEvalObject0.840.880.90
Action0.490.500.55
Physics0.360.390.44
Overall0.540.560.60
VBenchSubject Consistency91.8396.4296.68
Temporal Flickering99.1799.6199.81
Motion Smoothness96.5198.7998.85
Aesthetic Quality57.2459.2561.65
Imaging Quality59.1466.9668.66
Object Class76.0976.5882.81
Spatial Relationship62.3772.3877.88
Temporal Style22.2323.4324.05
Quality Score79.8182.8384.69
Total76.9379.4881.86

Fine-tuning contributes to the gains, but Ours exceeds FT-Long on every reported metric at matched training compute, including +2.8 points on VideoPhy2 Joint, +0.04 on FinePhyEval Overall and +2.38 points on VBench Total. Thus, additional training on the same data cannot fully explain our gains. We also added two controls in the 1,000-sample ablation. Capacity-matched DiT widens the backbone to match total parameters, while Compute-matched DiT trains the original backbone longer to match FLOPs. Both remove the refiner, and Ours leads on all six mean scores (see W4). These comparisons support the benefit of residual correction beyond additional data, capacity and compute.

[W2] Improvements on RAE's ImageNet-256

RAEv2 is already a strong baseline with limited headroom. We acknowledge its FID regression, despite improvements in complementary metrics. As generative models become sufficiently strong and performance approaches saturation, FID can disagree with human judgments and fail to track progressive visual improvements, as documented by [1]. To provide stronger evidence, we added ImageNet-256 experiments on the original RAE under the same protocol. Each Off/On pair shares the same adapted base checkpoint, initial noise and class labels. Across both training budgets, enabling the refiner improves every reported metric, including FID, IS and all three representation distances, both with and without AutoGuidance (AG). These controlled comparisons isolate the benefit of deploying the refiner and show that it complements guidance.

EpochsGuidanceRefinerFID ↓IS ↑FD_r DINOv2 ↓FD_r ConvNeXt ↓FD_r CLIP ↓
120NoneOff2.0325215.203.7082.72010.098
80+40NoneOn2.0106215.823.6682.69510.055
120AGOff1.5550232.702.6122.1888.538
80+40AGOn1.5430233.252.5772.1658.479
160NoneOff1.9204216.653.5512.6169.892
80+80NoneOn1.9102217.183.5352.6059.870
160AGOff1.4677233.352.4852.1088.317
80+80AGOn1.4579233.492.4732.0968.293

[1] Rethinking FID: Towards a Better Evaluation Metric for Image Generation, CVPR 2024

[W3] Generalized Theory Beyond Constructed Examples

We extended the theory beyond the XOR illustration, now moved to the appendix. First, our finite-sample analysis compares residual and direct-velocity heads with identical nonlinear features, parameters, data and update budgets. It establishes high-probability improvement conditions accounting for estimation, regularization and optimization, while a moving-checkpoint theorem controls feature and target drift. Second, an exact identity relates residual alignment along each base-parameter direction to the risk gradient, explaining how learnable residual signals can persist near base-model stationarity, with explicit bounds for training-weight mismatch. Third, our nonlinear AdamW theorem compares joint training against continued training of all base parameters at matched compute. Under gradient detachment and coupled updates, both methods share identical base states through their common training prefix. We certify lower risk when the correction gain exceeds the upper bound on gains from further base training, accounting for sampling uncertainty, nonlinear drift and optimizer dynamics. Finally, we connect training risk to finite-grid endpoint distributions through a conditional Wasserstein comparison that accounts for guidance and discretization errors. Together, these results establish explicit sufficient conditions for outperforming longer base training under the stated assumptions and finite first-pass regime.

[W4] Expanded Ablations across Runs and Matched-Baseline Comparisons

We now report mean ± sample standard deviation across four inference seeds for every ablation. Each variant uses a fixed trained checkpoint, so these statistics quantify inference variability. We add three controls: Direct velocity trains the same refiner architecture to predict the full target velocity, replacing the base velocity at inference instead of adding a residual correction. Capacity-matched DiT widens the base feed-forward layers to match CorrFlow’s total parameter count without a refiner. Compute-matched DiT trains the original base DiT longer without a refiner to match CorrFlow’s estimated total training FLOPs. CorrFlow achieves the highest mean on five of six metrics. Removing cross-attention, self-attention or the refiner lowers mean Quality Score by 0.0210, 0.0195 and 0.0135, respectively, each exceeding the reported standard deviations for both the corresponding variant and CorrFlow.

MethodAesthetic Quality (AQ)Background Consistency (BC)Imaging Quality (IQ)Overall Consistency (OC)Subject Consistency (SC)Quality Score (QS)
w/o noise-level conditioning0.5055 ± 0.00280.9260 ± 0.00230.6458 ± 0.00650.5850 ± 0.01080.9165 ± 0.00280.7982 ± 0.0014
w/o self-attention0.4830 ± 0.00650.9211 ± 0.00130.6088 ± 0.00480.5853 ± 0.00930.9081 ± 0.00360.7834 ± 0.0026
w/o cross-attention0.4775 ± 0.00460.9156 ± 0.00320.6033 ± 0.00630.5823 ± 0.00520.9071 ± 0.00270.7819 ± 0.0032
w/o hidden-state MLP0.4888 ± 0.00130.9234 ± 0.00310.6432 ± 0.00920.5765 ± 0.01340.9135 ± 0.00230.7914 ± 0.0045
w/o refiner0.5101 ± 0.00530.9289 ± 0.00060.6379 ± 0.00870.5832 ± 0.01500.9168 ± 0.00050.7894 ± 0.0016
Velocity input0.5080 ± 0.00180.9262 ± 0.00240.6414 ± 0.00380.5802 ± 0.00900.9160 ± 0.00420.8002 ± 0.0012
w/o text conditioning0.5052 ± 0.00060.9279 ± 0.00180.6215 ± 0.00710.5812 ± 0.01410.9201 ± 0.00220.7958 ± 0.0011
Direct velocity0.3834 ± 0.00180.8809 ± 0.00160.4461 ± 0.00470.4261 ± 0.00410.8448 ± 0.00300.6956 ± 0.0020
Capacity-matched DiT0.5156 ± 0.00180.9262 ± 0.00270.6413 ± 0.00740.5853 ± 0.00980.9180 ± 0.00410.8014 ± 0.0024
Compute-matched DiT0.5146 ± 0.00310.9273 ± 0.00230.6383 ± 0.00810.5844 ± 0.01420.9172 ± 0.00320.8009 ± 0.0048
Ours0.5268 ± 0.00190.9363 ± 0.00480.6455 ± 0.01220.5886 ± 0.01060.9264 ± 0.00250.8029 ± 0.0006

[Q2] Capacity- and Compute-Matched Comparisons

For a base model with approximately 10% more parameters, our Capacity-matched DiT (response to W4) widens the base feed-forward layers to match CorrFlow’s total parameter count while retaining the same training schedule. CorrFlow achieves higher mean scores on all six ablation metrics. For longer training at the same compute, we evaluate both FT-Long on the full training set (response to W1&Q1) and Compute-matched DiT in the new ablations. Both allocate the refiner’s training budget to additional base-model updates, matching CorrFlow’s estimated total training FLOPs. CorrFlow exceeds Compute-matched DiT on all six ablation means and outperforms FT-Long. These results support the benefit of residual correction beyond additional parameters or longer training.

[Q3 & W5] Evaluation Fairness, Sensitivity, and Inference Efficiency

First, evaluation fairness. Published and locally reproduced results follow the same benchmark protocols, with method-specific inference settings documented in the appendix. WMReward uses reward guidance with one candidate per condition to match the candidate budget. Since Best-of-N requires generating and scoring N videos at additional cost, our comparison evaluates guidance-only performance rather than the full Best-of-N pipeline.

Second, hyperparameter sensitivity. We added sensitivity studies reporting mean ± standard deviation across four inference seeds with fixed checkpoints. With σ∈[0.15,0.92]\sigma\in[0.15,0.92], Quality Score remains stable (0.8011–0.8032) across κ∈[0.1,1.0]\kappa\in[0.1,1.0], confirming robustness to gain selection. Larger gains degrade quality, consistent with overcorrection under CFG scale 5. Our default κ=0.2\kappa=0.2 provides a moderate correction strength.

κ\kappaAQBCIQOCSCQS
0.100.5152±0.00280.9407±0.00230.6365±0.00880.5821±0.01070.9247±0.00240.8022±0.0012
0.200.5268±0.00190.9363±0.00480.6455±0.01220.5886±0.01060.9264±0.00250.8029±0.0006
0.300.5146±0.00240.9400±0.00290.6356±0.00700.5824±0.00840.9242±0.00300.8032±0.0012
0.500.5139±0.00160.9413±0.00270.6353±0.00610.5832±0.00800.9240±0.00210.8015±0.0012
0.800.5132±0.00240.9417±0.00240.6350±0.00970.5838±0.01430.9234±0.00240.8011±0.0020
1.000.5116±0.00290.9401±0.00210.6341±0.00770.5849±0.00990.9226±0.00290.8015±0.0012
2.000.5099±0.00250.9391±0.00290.6295±0.00680.5846±0.01090.9189±0.00340.8001±0.0017
3.000.5069±0.00250.9340±0.00330.6224±0.00950.5838±0.01240.9118±0.00250.7949±0.0013
4.000.5024±0.00300.9252±0.00290.6120±0.00620.5835±0.00760.9029±0.00400.7877±0.0022
5.000.4993±0.00300.9218±0.00220.6001±0.00620.5830±0.01170.8911±0.00240.7809±0.0012
6.000.4951±0.00340.9163±0.00360.5869±0.00920.5835±0.01080.8771±0.00370.7695±0.0015

For the noise-interval sweep, we fix κ=0.3\kappa=0.3 and use the same 50-step sampler. Broad intervals maintain similar Quality Scores (0.8029–0.8035), while narrow intervals reduce quality to approximately 0.79. This supports applying moderate corrections across a sufficiently broad noise range without precise boundary tuning.

σ\sigma intervalAQBCIQOCSCQS
[0.00,1.00]0.5144±0.00300.9406±0.00200.6351±0.00620.5835±0.00740.9234±0.00370.8029±0.0017
[0.15,1.00]0.5144±0.00190.9424±0.00230.6354±0.00450.5835±0.01210.9235±0.00320.8034±0.0020
[0.00,0.92]0.5145±0.00230.9406±0.00280.6365±0.00640.5830±0.01220.9245±0.00310.8035±0.0013
[0.30,0.92]0.5145±0.00170.9408±0.00160.6364±0.00510.5824±0.00930.9244±0.00210.8035±0.0017
[0.50,0.92]0.5021±0.00320.9354±0.00210.6222±0.00500.5827±0.01530.9134±0.00290.7935±0.0018
[0.15,0.80]0.5029±0.00170.9343±0.00240.6212±0.00720.5830±0.01330.9148±0.00310.7936±0.0014
[0.30,0.70]0.4985±0.00220.9344±0.00360.6184±0.00840.5830±0.00940.9128±0.00310.7915±0.0017
[0.40,0.60]0.4980±0.00300.9318±0.00200.6155±0.00630.5819±0.01210.9109±0.00270.7897±0.0018
[0.50,0.60]0.4991±0.00180.9312±0.00240.6135±0.00700.5824±0.01180.9114±0.00220.7894±0.0015

Third, inference efficiency. We added per-video FLOPs, latency and peak GPU memory comparisons for fp32 and mixed-precision (autocast) refiners, keeping base precision and sampling settings fixed. For text-to-video (81×480×83281\times480\times832), autocast adds only 6% compute and 5% latency, versus 53% latency overhead in fp32. CorrFlow is more efficient than DreamWorld, VPT, VideoREPA and Wan2.1-14B across all three measures. PhysRAG is faster but uses more memory. Costs are normalized to Wan2.1-1.3B.

T2V methodPFLOPsCompute×Latency(s)Latency×Memory(GiB)
Wan2.1-1.3B28.91.00×89.3±2.41.00×15.4
CorrFlow-1.3B (fp32)30.71.06×136.3±0.21.53×17.8
CorrFlow-1.3B (autocast)30.71.06×94.2±0.21.05×17.6
DreamWorld-1.3B70.92.45×238.2±0.52.67×21.2
VPT-1.3B56.91.97×193.4±0.12.17×19.9
VideoREPA-5B88.43.06×294.0±1.13.29×25.1
PhysRAG-5B10.80.37×36.1±0.10.40×34.5
Wan2.1-14B168.15.82×475.2±0.15.32×46.6

For image-to-video (81×704×124881\times704\times1248), autocast adds 7% compute and 8% latency, versus 59% latency overhead in fp32. It also achieves lower latency and memory than all compared adaptation methods. Costs are normalized to Wan2.2-5B.

I2V methodPFLOPsCompute×Latency(s)Latency×Memory(GiB)
Wan2.2-5B30.61.00×83.6±2.41.00×25.9
CorrFlow-5B (fp32)32.61.07×132.7±0.11.59×29.2
CorrFlow-5B (autocast)32.61.07×90.0±0.21.08×29.2
WMReward-5B (guidance only)62.72.05×230.7±0.72.76×115.2
PhysRAG-5B30.81.01×98.5±0.01.18×48.9
VLIPP-5B (30 steps)223.47.30×713.2±0.18.53×45.4
Reviewer_GFDROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CorrFlow, which augments a pretrained flow-matching generator with a lightweight transformer refiner. The refiner receives the base model's one-step clean prediction, intermediate hidden states, conditioning information, and noise level. It predicts a residual correction that is added to the velocity field during sampling. They further provide theoretical analyses of feature reuse, compute-matched training, and deployment-time velocity improvement. The empirical evaluation spans ImageNet-256, text-to-video generation, and image-to-video generation, reporting improvements on several visual-quality and physical-consistency benchmarks.

Critical strengths

Simple and well-motivated formulation: The identity x^0−x0=σr\hat{x}_0-x_0=\sigma r provides a clear mathematical motivation for learning velocity residuals. Applying corrections within the sampling loop is a natural and practical design.

Efficient feature reuse: The refiner leverages intermediate representations from the same base-model forward pass, avoiding additional backbone evaluations. Its modular architecture makes it applicable to existing flow-matching models.

Broad empirical effectiveness: The method demonstrates improvements across image generation, text-to-video, and image-to-video tasks, including visual quality, temporal consistency, and physical plausibility.

Critical weaknesses

1. Limited relevance of theoretical guarantees to practical generative improvement: Theorem 2 establishes an advantage over affine predictors using a constructed example where the hidden representation explicitly contains the required nonlinear interaction. Theorem 3 assumes a favorable tangent-risk separation, while Theorem 4 guarantees only conditional velocity-error improvement, not better generated distributions. Appendix C.4 even demonstrates that lower velocity error can produce worse endpoint distributions. Consequently, the theory does not convincingly explain why CorrFlow improves generation quality in realistic nonlinear sampling settings.

2. The proposed self-correction mechanism is not fundamentally distinguished from auxiliary velocity prediction: Since the refiner's inputs are deterministic functions of the base model's current computation, it receives no independent information about prediction errors. Under squared-error supervision, its optimal correction is simply v∗−vθv^*-v_\theta, making the corrected velocity equivalent to another approximation of the Bayes-optimal velocity field. Thus, the approach resembles learning a larger velocity predictor through residual parameterization rather than a distinct self-correction mechanism. The paper does not establish why residual prediction should outperform an equally expressive direct velocity predictor.

3. Improvements cannot be conclusively attributed to residual correction: The base model and refiner are trained jointly, so improvements may arise from continued backbone adaptation or additional model capacity. Although base-only and component ablations are provided, the paper lacks decisive comparisons against equally sized direct-velocity heads or capacity-matched alternatives. Therefore, the experiments demonstrate the effectiveness of the combined system without sufficiently isolating the proposed mechanism.

Questions for authors

See weaknesses.

3 comments

Thanks for the reviewer's constructive review.

[W1] Stronger Theoretical Guarantees for Distributional Improvement

We agree that the original theory did not establish improved endpoint distributions. The revision addresses these gaps explicitly. First, revised Theorems 2–3 extend beyond the original constructed example to establish finite-sample, high-probability guarantees for general bounded nonlinear features and moving checkpoints. These bounds account for estimation, regularization, optimization and feature drift without requiring exact target representation. A complementary stationarity result identifies useful residual signals outside the base model’s tangent space. Second, revised Theorem 4 compares nonlinear joint AdamW training against continued training of all base parameters at matched compute. It replaces the assumed tangent-risk separation with explicit bounds on correction gains and additional base-training gains. Third, revised Theorem 5 connects training risk to endpoint distributions under nonlinear finite-grid sampling, accounting for irreducible noise, training-to-sampling transfer, guidance, correction strength, discretization and numerical errors. It combines an upper bound on the corrected sampler’s conditional Wasserstein-2 distance with an independent held-out lower bound for the base, adjusted for further training. A positive margin certifies better endpoint distributions at matched compute. Finally, we retain the Appendix C.4 counterexample and explicitly address temporal cancellation through cross-time error correlations, ordered Jacobian propagation and nonlinear remainder bounds. The revised theory therefore provides sufficient conditions for distributional improvement under stated regularity assumptions, while preserving the distinction between lower velocity error and better generation.

[W2] Self-Correction vs. Direct Velocity Prediction

CorrFlow differs from auxiliary velocity prediction through its inputs, supervision and deployment. The refiner revisits the base model’s clean prediction and hidden states, predicts a detached residual and corrects the base velocity at each sampling step. Although these features provide no independent information, they can expose predictable errors that the existing predictor has not eliminated. Intuitively, reviewing an answer together with the intermediate work that produced it can improve that answer without introducing new observations. We agree that both methods share the same Bayes-optimal velocity under pure squared-error supervision. However, this does not imply equivalent learning under finite data and compute. Revised Theorem 2 compares residual and direct readouts using identical nonlinear features, head capacity, samples and update budgets, certifying lower risk for residual prediction when the derived margin is positive. Theorem 3 extends these guarantees to moving checkpoints with explicit drift and optimizer penalties. Our matched-head ablation (It is shown in revised paper’s table-5 and response to W3) shows that residual prediction achieves a Quality Score of 0.8029 versus 0.6956 for direct prediction, with improvements across all six metrics. These results support the finite-resource advantage of correcting pretrained predictions, rather than claiming a different Bayes optimum.

[W3] Residual Correction vs. Direct and Capacity-Matched Baselines

To isolate residual correction from additional capacity and backbone adaptation, we evaluate two controls under the same ablation protocol, reporting mean ± standard deviation across four inference seeds per fixed checkpoint. Direct velocity uses the identical refiner architecture to predict the full velocity, replacing the base prediction at inference. Capacity-matched DiT removes the refiner, widens the backbone’s feed-forward layers to match our total parameter count and fine-tunes the backbone with the same data, optimization settings and training schedule.

MethodAQBCIQOCSCQS
Direct velocity0.3834 ± 0.00180.8809 ± 0.00160.4461 ± 0.00470.4261 ± 0.00410.8448 ± 0.00300.6956 ± 0.0020
Capacity-matched DiT0.5156 ± 0.00180.9262 ± 0.00270.6413 ± 0.00740.5853 ± 0.00980.9180 ± 0.00410.8014 ± 0.0024
Ours0.5268 ± 0.00190.9363 ± 0.00480.6455 ± 0.01220.5886 ± 0.01060.9264 ± 0.00250.8029 ± 0.0006

CorrFlow achieves the highest mean on all six metrics, improving Quality Score from 0.6956 to 0.8029 over direct prediction. The capacity-matched model recovers much of the gain but remains slightly behind. Together, these controls support the practical benefit of residual correction beyond direct velocity prediction and increased model capacity.

Dear Reviewer,

Thank you for your valuable feedback. In response to your concerns, we have provided further clarifications and additional experimental analyses in our rebuttal.

We would greatly appreciate it if you could revisit our responses and let us know whether they have adequately addressed your concerns. If your concerns have been resolved, we sincerely hope you would consider raising your rating.

Once again, we sincerely thank you for your time, effort, and thoughtful review of our work.

Best,

Authors

Dear Reviewer,

With about 20 hours remaining before the discussion period closes, we would greatly appreciate it if you could revisit our rebuttal and additional experimental results. If your concerns have been adequately addressed, we sincerely hope you would consider raising your rating.

Thank you again for your time and thoughtful consideration.

Best,
Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response for Manuscript Revision

We sincerely thank the reviewers for their constructive feedback. We have substantially strengthened the manuscript through the following additions, all highlighted in blue:

  • Expanded Theoretical Analysis: We developed rigorous finite-sample guarantees, moving-checkpoint analyses, and nonlinear matched-compute comparisons to establish when residual refinement can outperform continued base-model training.
  • Stronger Sampling Guarantees: We established conditions for reducing velocity errors under the corrected sampler's own state distribution and derived bounds connecting training risk to final generation distributions under explicit assumptions.
  • Fairer Baseline Comparisons: We introduced longer base-model fine-tuning (FT-Long), capacity-matched DiT, and compute-matched DiT baselines to distinguish the benefits of self-correction from those of additional parameters or training computation.
  • Broader Experimental Validation: We extended our ImageNet-256 experiments to the original RAE architecture, demonstrating consistent improvements in FID, Inception Score, and representation-based metrics under both guided and unguided sampling.
  • Comprehensive Ablation Studies: We systematically examined residual prediction, clean-prediction inputs, attention mechanisms, and conditioning components, reporting means and standard deviations across four inference seeds.
  • Inference Efficiency Evaluation: We added detailed comparisons of computational cost, inference latency, and GPU memory usage, showing that mixed-precision refinement incurs only approximately 5–8% additional latency relative to the corresponding video base models.
  • Sensitivity and Reproducibility Analyses: We investigated correction strength and active noise intervals, clarified their quality–efficiency trade-offs, and provided detailed implementation settings and exact training-budget calculations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.