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. Project Page: https://fogradio.github.io/corrflow-project
Figures & tables
Figure 1: CorrFlow learns from its own imperfect predictions. The base model’s clean prediction (blue) deviates from the clean reference (orange arrow). The resulting correction shifts the prediction closer to the clean reference by −σκr^ (green arrow), where σ is the noise level and κ controls the correction strength. The initial noise is obtained by adding Gaussian noise to the clean reference.
Figure 2: CorrFlow overview for video generation. (a) At noise level σ , the refiner predicts a residual r^ from the base DiT’s clean prediction x^0=xσ−σvθ , selected hidden states hL , conditioning tokens c and noise-level embedding eσ . Sampling corrects the base velocity as vκ=vθ+κr^ with strength κ . (b) The refiner alternates self-attention over clean-prediction tokens with cross-attention to projected conditioning and fused hidden-state tokens. Noise conditioning modulates its blocks and sample-space output head, while 3D rotary embeddings encode video token coordinates. Frames show outputs and cubes denote sample tensors.
Method
Epochs
Guidance
IS ↑
Representation Fréchet Distance ( FDr ) ↓
ConvNeXt ( Liu et al., 2022 )
DINOv2 ( Oquab et al., 2023 )
CLIP ( Radford et al., 2021 )
RAEv2 ( Singh et al., 2026 )
160
No
233.9
1.28
2.22
7.18
SelfFlow ( Chefer et al., 2026 )
160
No
146.4
4.81
7.08
12.13
Ours
80+80
No
234.0
1.24
2.21
7.14
SiT-XL/2 ( Ma et al., 2024 )
800
Yes
256.7
2.02
7.89
17.69
RAE-XL ( Zheng et al., 2026a )
800
Yes
261.0
1.79
2.11
7.87
Table 1: Comparison on ImageNet-256 ( Deng et al., 2009 ) generation. Representation Fréchet Distance ( FDr ) is reported in three feature spaces.
Method
VideoPhy2
FinePhyEval
ALL
HARD
Object
Action
Physics
Overall
PC
SA
Joint
PC
SA
Joint
Wan 2.1-14B ( Wan et al., 2025 )
60.5
29.0
24.7
45.6
8.3
3.9
0.86
0.53
0.46
0.59
PhysRAG-5B ( Cheng et al., 2026 )
65.6
24.7
21.0
59.4
7.2
4.4
0.88
0.49
0.37
0.55
VideoREPA-5B ( Zhang et al., 2025c )
73.6
25.8
23.6
62.8
6.7
5.0
0.82
0.37
0.30
0.47
Wan 2.1-1.3B ( Wan et al., 2025 )
57.8
30.0
24.8
36.7
11.7
5.6
0.84
0.49
0.36
0.54
Table 2: Physics-consistent text-to-video generation on VideoPhy2 ( Bansal et al., 2026 ) and FinePhyEval ( Pothiraj et al., 2026 ) . Higher is better for all metrics.
Method
SC
TF
MS
AQ
IQ
OC
SR
TS
QS
Total
VideoREPA-5B ( Zhang et al., 2025c )
95.47
98.23
98.54
57.19
59.80
67.56
51.85
22.71
79.96
76.67
PhysRAG-5B ( Cheng et al., 2026 )
96.31
98.71
96.93
56.80
65.56
80.50
77.33
23.18
83.07
80.08
Wan 2.1-1.3B ( Wan et al., 2025 )
91.83
99.17
96.51
57.24
59.14
76.09
62.37
22.23
79.81
76.93
VideoJAM-1.3B ( Chefer et al., 2025 )
91.51
99.13
96.05
58.41
62.68
79.22
66.17
24.03
81.18
78.76
DreamWorld-1.3B ( Tan et al., 2026 )
93.62
98.81
98.07
59.59
66.76
81.32
70.47
23.82
83.49
80.97
VPT-1.3B ( Zheng et al., 2026b )
92.65
98.64
97.85
58.09
63.68
73.43
73.35
22.52
82.26
78.79
Table 3: VBench ( Huang et al., 2024 ) comparison on selected metrics (0–100, higher is better).
Figure 3: Qualitative comparison on text-to-video generation. Each scenario compares VideoREPA-5B, VPT-1.3B and CorrFlow (Ours) under the same text prompt, with frames sampled at 0% , 25% , 50% , 75% and 100% of each generated video. The prompts are drawn from VBench (amusement park), FinePhyEval (a domino dropped into a mug) and VideoPhy2 (a somersault on a bouncy castle, a coin caught in a metal cup).
Method
Spatial
Spatio-temporal
Weighted Spatial
MSE Score
Verified Score
P-Video-19B ( Pruna AI, 2026 )
38.60
16.40
22.90
23.30
25.30
Wan2.2-5B ( Wan et al., 2025 )
35.89
26.44
22.96
25.54
27.71
WMReward-5B ( Yuan et al., 2026 )
36.47
22.38
19.79
27.26
26.48
VLIPP-5B ( Yang et al., 2025 )
40.51
19.09
23.66
26.48
27.43
PhysRAG-5B ( Cheng et al., 2026 )
36.46
22.07
20.22
23.41
25.54
CorrFlow (Ours)-5B
43.43
26.37
24.74
29.24
30.95
Table 4: Physics-IQ Verified ( Rädsch et al., 2026 ) results on the image-to-video track. All evaluations are without LLM usage. All scores are percentages, and higher is better for every metric.
Figure 4: Qualitative comparison on the Physics-IQ Verified image-to-video track. Each scenario compares WMReward-5B, PhysRAG-5B and CorrFlow (Ours), which share the Wan2.2-TI2V-5B base model and receive the same settings. Temporal order is from left to right. The left scenario shows a statuette’s shadow under a moving light, and the right scenario shows a knife slicing a mandarin.
Method
AQ
BC
IQ
OC
SC
QS
w/o time-step
51.5
92.4
63.7
58.5
91.2
79.7
w/o self-attention
48.9
92.1
60.4
58.1
90.0
78.5
w/o cross-attention
48.5
91.4
60.7
58.2
90.2
78.5
w/o hidden-state MLP
49.3
92.4
63.6
58.2
90.9
79.3
w/o refiner
51.3
92.5
62.0
59.1
91.3
78.6
v input
50.4
92.4
63.7
60.2
91.7
79.8
Table 5: Component ablations (%, higher is better). Colors rank the four-decimal scores in Table 10 , which also expands the metric names.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Image representation
Steps
Guidance
SiT-XL/2 ( Ma et al., 2024 )
SD-VAE
250
CFG
DDT-XL ( Wang et al., 2026b )
SD-VAE
250
CFG
SiT-XL/2-REPA ( Yu et al., 2025 )
SD-VAE
250
CFG
LightningDiT ( Yao et al., 2025 )
VA-VAE
250
CFG
REG ( Wu et al., 2026 )
SD-VAE
250
CFG
REPA-E ( Leng et al., 2025 )
Jointly tuned VAE
250
CFG
Appendix
Table 6: Evaluation configurations for all seven published image-generation baselines in Table 7 . Steps count sampler updates. CFG combines conditional and unconditional predictions, while AutoGuidance combines a main model with a weaker model.
Method
Epochs
Guidance
IS ↑
FID ↓
Representation Fréchet Distance ( FDr ) ↓
ConvNeXt
DINOv2
CLIP
Published baselines
SiT-XL/2 ( Ma et al., 2024 )
800
Yes
256.7
2.12
2.02
7.89
17.69
DDT-XL ( Wang et al., 2026b )
800
Yes
309.3
1.26
1.02
4.26
13.86
SiT-XL/2-REPA ( Yu et al., 2025 )
800
Yes
306.1
1.42
1.22
4.27
12.65
LightningDiT ( Yao et al., 2025 )
800
Yes
294.3
1.42
1.09
3.76
10.21
Appendix
Table 7: Comparison on ImageNet-256 ( Deng et al., 2009 ) generation. For published baselines, the FID column reports guided FID (gFID) as in the original papers. Models are evaluated with and without guidance as indicated. For our method, 80+40 and 80+80 indicate that we start from the 80-epoch RAEv2 checkpoint and train our refiner for an additional 40 and 80 epochs, respectively. “Ours (base only)” evaluates the corresponding base model without the refiner at inference and reports the total base-model training epochs. Representation Fréchet Distance ( FDr ) is reported in the feature spaces of ConvNeXt ( Liu et al., 2022 ) , DINOv2 ( Oquab et al., 2023 ) and CLIP ( Radford et al., 2021 ) .
Figure 5: Qualitative results on ImageNet at 256×256 resolution. Columns show shoji, plate rack, hatchet and marmoset. Row groups show Self-Flow (Selfflow), RAEv2 and CorrFlow (Ours). Each method–category block contains four samples in a 2×2 grid, giving 48 images in total. All three base DiTs are trained for 160 epochs. The CorrFlow refiner is trained for 80 epochs.
Figure 6: Qualitative results on ImageNet at 256×256 resolution. Columns show axolotl, Papillon, Miniature Poodle and Komondor. Row groups show Self-Flow (Selfflow), RAEv2 and CorrFlow (Ours). Each method–category block contains four samples in a 2×2 grid, giving 48 images in total. All three base DiTs are trained for 160 epochs. The CorrFlow refiner is trained for 80 epochs.
Method
Subject Consistency
Background Consistency
Temporal Flickering
Motion Smoothness
Dynamic Degree
Aesthetic Quality
Imaging Quality
Object Class
Multiple Objects
Human Action
VideoREPA-5B ( Zhang et al., 2025c )
95.47
95.10
98.23
98.54
47.22
57.19
59.80
67.56
40.32
83.00
PhysRAG-5B ( Cheng et al., 2026 )
96.31
95.66
98.71
96.93
81.94
56.80
65.56
80.50
75.00
63.00
Wan 2.1-1.3B ( Wan et al., 2025 )
91.83
94.71
99.17
96.51
65.00
57.24
59.14
76.09
53.66
74.60
VideoJAM-1.3B ( Chefer et al., 2025 )
91.51
96.01
99.13
96.05
73.88
58.41
62.68
79.22
59.66
79.20
DreamWorld-1.3B ( Tan et al., 2026 )
93.62
94.95
98.81
98.07
79.16
59.59
66.76
81.32
65.03
81.20
VPT-1.3B ( Zheng et al., 2026b )
92.65
97.27
98.64
97.85
70.83
58.09
63.68
73.43
51.92
74.80
Appendix
Table 8: VBench results (Part 1/2): consistency, motion and object-level metrics. All scores are reported on a 0–100 scale.
Method
Color
Spatial Relationship
Scene
Appearance Style
Temporal Style
Overall Consistency
Quality Score
Semantic Score
Overall Score
VideoREPA-5B ( Zhang et al., 2025c )
86.12
51.85
32.27
21.37
22.71
24.20
79.96
63.53
76.67
PhysRAG-5B ( Cheng et al., 2026 )
93.95
77.33
15.55
22.00
23.18
23.18
83.07
68.12
80.08
Wan 2.1-1.3B ( Wan et al., 2025 )
89.93
62.37
20.03
22.91
22.23
24.25
79.81
65.43
76.93
VideoJAM-1.3B ( Chefer et al., 2025 )
88.92
66.17
28.34
22.67
24.03
25.01
81.18
69.08
78.76
DreamWorld-1.3B ( Tan et al., 2026 )
92.61
70.47
29.71
22.14
23.82
24.87
83.49
70.89
80.97
VPT-1.3B ( Zheng et al., 2026b )
86.99
73.35
20.13
20.87
22.52
23.42
82.26
64.91
78.79
Appendix
Table 9: VBench results (Part 2/2): appearance, style and overall scores. All scores are reported on a 0–100 scale.
Figure 7: Qualitative results on text-to-video generation. The upper case shows a rubber duck among dominoes, and the lower case shows a skateboard and a surfboard. Within each case, rows show VideoREPA, VPT, PhysRAG and Ours from top to bottom. Each row contains five frames sampled at evenly spaced relative positions across the video, ordered from left to right.
Figure 8: Qualitative results on text-to-video generation. The upper case shows a canoeist paddling across a lake, and the lower case shows two jetskis racing side by side. Within each case, rows show VideoREPA, VPT, PhysRAG and Ours from top to bottom. Each row contains five frames sampled at evenly spaced relative positions across the video, ordered from left to right.
Figure 9: Qualitative results on text-to-video generation. The prompt describes a parasailer with an orange parachute being pulled across the ocean by a speedboat. Rows show VideoREPA, VPT, PhysRAG and Ours from top to bottom. Each row contains five frames sampled at evenly spaced relative positions across the video, ordered from left to right.
Figure 10: Additional image-to-video comparison on Physics-IQ Verified: paper falling into water. A grabber releases crumpled paper into a bowl of water. Rows show GT, WAN, VLIPP, WMReward, PhysRAG and CorrFlow (Ours). Each row contains five video frames in temporal order.
Figure 11: Additional image-to-video comparisons on Physics-IQ Verified: motion and reflections. The upper case shows a suspended tennis ball being lowered in front of a mirror, and the lower case shows a tennis ball rolling toward a kettlebell beside a mirror. Within each case, rows show GT, WAN, VLIPP, WMReward, PhysRAG and CorrFlow (Ours). Each row contains five video frames in temporal order.
Figure 12: Additional image-to-video comparisons on Physics-IQ Verified: ball motion. The upper case shows a bouncing tennis ball near a ramp, and the lower case shows a blue ball rolling from a table onto glass. Within each case, rows show GT, WAN, VLIPP, WMReward, PhysRAG and CorrFlow (Ours). Each row contains five video frames in temporal order. Red boxes highlight visible balls and adjacent reflections in the four baselines after the initial frame.
Method
Aesthetic Quality (AQ)
Background Consistency (BC)
Imaging Quality (IQ)
Overall Consistency (OC)
Subject Consistency (SC)
Quality Score (QS)
w/o time-step
0.5147
0.9236
0.6369
0.5850
0.9120
0.7970
w/o self-attention
0.4892
0.9213
0.6043
0.5812
0.9000
0.7854
w/o cross-attention
0.4848
0.9144
0.6066
0.5821
0.9022
0.7846
w/o hidden-state MLP
0.4928
0.9238
0.6359
0.5817
0.9086
0.7928
w/o refiner
0.5129
0.9249
0.6204
0.5914
0.9126
0.7863
v input
0.5038
0.9238
0.6368
0.6021
0.9174
0.7980
Appendix
Table 10: Component ablations at four-decimal precision (0–1, higher is better). Column abbreviations match Table 5 . Colors indicate the top three scores in each column.
Flow Matching enables high-quality visual generation via continuous-time dynamics, but inference remains costly due to multiple sequential function evaluations. Existing acceleration methods reduce the number of function evaluations but often introduce additional training overhead, degrade quality, or fail to account for input-dependent variability. We propose COFLOW, an inference-time method that adaptively selects the step counts each generation based on the prompt features. Our context-aware COFLOW is trained online with an unsupervised reward that balances inference efficiency and generation fidelity. Our method is plug-and-play, requiring no retraining of the underlying generative model. It generalizes to image and video generation, achieving over 2.5x speedup while preserving perceptual and semantic quality. We further provide a theoretical analysis establishing an O(1/K) forward-Euler discretization error bound under standard regularity conditions.
In rectified-flow-based generative models, the neural network can be trained to predict two different targets, such as the instantaneous velocity or the data endpoint, to perform denoising. Although prior work shows that these parameterizations lead to different empirical behaviors, the mechanisms underlying their respective advantages remain to be underexplored, and how to combine them effectively is still unclear. In this work, we analyze how learning errors from different parameterizations affect the generation performance. We show that predicting the data endpoint has a clear training signal that stabilizes training, whereas predicting the velocity maintains stable sampling dynamics near the data manifold. Motivated by these insights, we propose Self-Consistent Flow (SC-Flow), a new method that unifies the benefits of both parameterizations. By employing a lightweight consistency loss, SC-Flow jointly trains a single network to predict both the local velocity and the data endpoint, and the consistency between the two predictions improves the model's performance. The method requires no major architectural changes and adds minimal computational overhead. Extensive experiments on image generation tasks demonstrate that SC-Flow substantially stabilizes optimization and improves the straightness of generation paths, leading to significant gains in generation quality over standard rectified-flow baselines.
We introduce CNP-Flow, a flow matching framework for temporal generation that learns conditional source distributions through flow reversal. Whereas standard conditional flow matching (FM) incorporates conditioning through the vector field and draws source samples from a standard Gaussian, CNP-Flow uses a conditional noise predictor (CNP) to produce an isotropic Gaussian source for each temporal condition. The CNP is supervised by source samples obtained through flow reversal, which maps observed targets backward through a pretrained FM model. A three-stage pipeline pretrains the FM model, trains the CNP, and fine-tunes the FM model using the learned source distribution, while preserving the FM backbone architecture. Across video prediction, video interpolation, and 7-DoF Franka robot motion planning, CNP-Flow consistently improves generation quality. It also matches baseline performance with fewer function evaluations. Project page: https://embodiedai-ntu.github.io/cnpflow