Organizations: Fudan University · Institute of Information Engineering, Chinese Academy of Sciences · School of Cyber Security, University of Chinese Academy of Sciences · East China Normal University · Alibaba Group
Flow matching enables high-quality image and video generation, but its iterative sampling process requires repeated evaluations of large neural networks. Existing few-step methods can distort the output and reduce its agreement with a many-step reference. We propose TrajCal, a training-free calibration framework for few-step sampling. Based on scatter search, a well-established metaheuristic for difficult optimization problems, TrajCal starts with diverse seed schedules, deforms and combines these candidates, and selects one fixed schedule by final output fidelity for reuse on new inputs. For both image and video generation, TrajCal achieves higher reference fidelity than all compared baselines at low evaluation budgets while maintaining competitive visual quality. The search is effective even with randomly generated seeds, while trajectory-based seeds further improve fidelity. For cross-style generalization, photography-calibrated schedules also improve fidelity on unseen image styles.
Figures & tables
Figure 2: Schedule calibration and reuse in TrajCal. (a) Three seed schedules provide distinct time allocations. Mixing and deformation expand the candidate pool, followed by preliminary RMSE screening. The retained candidates are evaluated against paired many-step references, and a normalized combination of RMSE, LPIPS, and SSIM losses selects one schedule for each update rule and budget. (b) The selected schedule and update rule are reused on new inputs with B velocity evaluations. For image sampling, the displacement correction uses consecutive velocity predictions without additional model evaluations.
Method
Budget
Reference fidelity
Quality
RMSE ↓
LPIPS ↓
SSIM ↑
FID↓
HPSv3 ↑
CLIP ↑
FLUX-12
12
48.48
0.4720
0.6403
57.24
11.86
0.2633
GITS
52.76
0.5380
0.6193
78.26
8.18
0.2584
SharpEuler-Z
32.81
0.3234
0.7338
41.43
13.14
0.2620
CAT-OV
35.86
0.3739
0.7065
52.26
11.91
0.2602
OptimalSteps
35.51
0.3664
0.7169
48.74
12.52
0.2621
Table 1: Image generation at 12 and 16 evaluations. Results on 500 FLUX.1-dev inputs, with reference fidelity measured against FLUX-50 outputs from the same prompts and base noise. RMSE, LPIPS, and SSIM are averaged over paired outputs; FID compares the two 500-image sets. HPSv3 and CLIP assess human preference and prompt alignment. B denotes the evaluation budget; CAT-OV averages 11.57 and 15.504 evaluations. Bold and underlined values mark the best and second-best results; shading highlights TrajCal-Correction.
Figure 3: Visual comparison at 12 and 16 evaluations. Each row uses the same prompt and base noise across methods, with the FLUX-50 reference in the first column. The upper two rows use 12 evaluations and the lower two use 16. The examples compare scene layout, subject pose, object arrangement, and appearance; TrajCal-Correction more closely preserves these features of the reference outputs.
Method
Budget
Reference fidelity
VBench dimensions ( ×100 )
PSNR ↑
LPIPS ↓
SSIM ↑
OC ↑
SC ↑
BC ↑
MS ↑
IQ ↑
DD ↑
UniPC-10
10
10.58
0.6648
0.3994
18.61
95.81
96.72
98.11
60.11
33
StableVS
11.16
0.6304
0.4167
20.01
96.28
96.92
98.04
64.34
42
OptimalSteps
13.44
0.4911
0.4985
21.78
96.62
96.98
98.28
65.01
49
SharpEuler-Wan
12.26
0.5542
0.4562
21.37
96.80
96.92
98.25
65.27
44
TrajCal-Euler
14.88
0.4073
0.5392
21.63
96.53
97.03
98.27
66.92
50
Table 2: Video generation at 10 and 20 evaluations. Results on 100 Wan2.2-TI2V-5B inputs, using paired UniPC-50 references and B guided velocity evaluations. PSNR uses all 81 frames per video; LPIPS and SSIM use 16 uniformly sampled frames. The VBench dimensions are overall consistency (OC), subject consistency (SC), background consistency (BC), motion smoothness (MS), imaging quality (IQ), and dynamic degree (DD). Bold and underlined values mark the best and second-best few-step results within each budget; shading highlights TrajCal-UniPC.
Method
Budget
Reference fidelity
RMSE ↓
LPIPS ↓
SSIM ↑
FID↓
FLUX-12
12
56.05
0.4813
0.6879
125.27
GITS
59.57
0.5402
0.6844
169.76
SharpEuler-Z
41.27
0.3374
0.7437
82.38
CAT-OV
40.97
0.3557
0.7326
91.34
OptimalSteps
42.50
0.3787
0.7232
91.40
Table 3: Cross-style generalization. Schedules calibrated on photography are reused unchanged on 30 scenes in four styles: photography, watercolor, anime, and flat-vector art. FID is computed from all 120 generated images and their references; the other three metrics are averaged over these outputs. Bold, underlining, and shading follow Table 1 .
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Seed schedules and schedules selected by TrajCal. Each row shows the B model evaluation times and the final endpoint, where no evaluation is made. Left: the three FLUX seeds and the schedules selected for TrajCal-Euler and TrajCal-Correction at B=12 . Right: native UniPC and the schedules selected for TrajCal-Euler, TrajCal-Correction, and TrajCal-UniPC on Wan at B=10 . Orange marks the first three query times shared by TrajCal-Euler and TrajCal-Correction; TrajCal-UniPC is calibrated separately. The horizontal axis uses a square-root scale to make early times easier to see; tick labels give the actual sampling time t . The FLUX seeds are shown before prefix alignment.
Parameter
Value
Seed schedules
3
Pairwise averages
3, with weight 1/2
Candidates per base schedule
7: original schedule + 6 deformations
Late-interval deformations
2, with α∈{−0.35,0.35}
Suffix deformations
4, with (a,b)∈{−0.15,0.15}2
Candidates after expansion
42; 43 for TrajCal-UniPC, before merging duplicates
Appendix
Table 4: Search settings. Counts are for one model at one evaluation budget. TrajCal-UniPC also includes the native UniPC schedule.
Study
∣Dc∣
∣Ds∣
Shared
Evaluation
Main images, B=12,16
64
128
64
500
Additional images, B=8
64
128
64
200
Additional images, B=20
64
64
64
200
Cross-style, B=12,16
64
128
64
4×30
Random seeds, B=12
8
128
8
200
Videos, B=10,20
24
24
0
100
Appendix
Table 5: Number of calibration and evaluation inputs. We use Dc to construct seeds and screen candidates, and Ds to select the final schedules. “Shared” counts inputs included in both calibration sets. Evaluation inputs are separate from both sets.
Setting
Images
Videos
Model
FLUX.1-dev
Wan2.2-TI2V-5B
Resolution
1024×1024
832×480
Frames
1
81 at 24 fps
Guidance scale
3.5
5
Flow shift
Native FLUX
5
Weight precision
BF16
BF16
Appendix
Table 6: Image and video generation settings. All runs use a batch size of one.
12 evaluations
16 evaluations
Setting
RMSE ↓
LPIPS ↓
SSIM ↑
RMSE ↓
LPIPS ↓
SSIM ↑
OptimalSteps: initial-noise scaling (100 inputs; time nodes fixed)
Initial state (1−t0)z
37.86
0.3804
0.7050
38.21
0.3542
0.7145
Initial state z
37.01
0.3737
0.7111
37.16
0.3421
0.7236
SharpEuler: schedule settings (50 main images)
SharpEuler FLUX script
34.28
0.5391
0.6455
23.74
0.3809
0.7426
Appendix
Table 7: Baseline settings and reference choice. The OptimalSteps and SharpEuler comparisons change only the named setting and use FLUX-50 as the reference. In the last block, we keep the generated outputs fixed and change the reference to iPNDM-60.
Figure 5: Random schedules and the three selected seeds. Each column shows four schedules sampled from one ρ range at B=12 . Top: interval lengths hj as percentages of the full sampling interval. Bottom: hj+1/hj , the ratio of the next interval length to the current one; the dashed line at one marks equal lengths. Intervals grow throughout in the first two columns, but their growth ratios increase on the left and decrease in the middle. In the right column, interval lengths peak near the end and then decrease. Colors mark the selected seeds S1–S3; the remaining schedules are gray. The schedules are shown before mixing, deformation, and prefix alignment. Table 8 gives all twelve parameter pairs.
ρ range
Schedule
a
ρ
ea
ρ∈[−0.5,0]
R1-1
1.8410
−0.4501
6.30
R1-2
2.0849
−0.0573
8.04
R1-3
2.4894
−0.3352
12.05
R1-4
3.1125
−0.1748
22.48
ρ∈[0,1]
R2-1
1.4529
0.3120
4.28
R2-2
1.9562
0.4006
7.07
Appendix
Table 8: Parameters of the twelve random schedules. The two indices in each name identify the ρ range and a subinterval. The last column gives the ratio of the last interval length to the first, ea=h11/h0 . Values are rounded for display.
Method
Seed type
Reference fidelity
Quality
RMSE ↓
LPIPS ↓
SSIM ↑
FID ↓
HPSv3 ↑
CLIP ↑
FLUX-12
—
49.38
0.4754
0.6367
84.59
11.76
0.2664
GITS
—
53.50
0.5381
0.6169
112.98
7.83
0.2611
SharpEuler-Z
—
33.70
0.3286
0.7275
58.80
13.17
0.2659
CAT-OV
—
36.34
0.3759
0.7017
76.67
11.68
0.2629
OptimalSteps
—
35.67
0.3615
0.7176
68.65
12.44
0.2652
Appendix
Table 9: Random and trajectory-based seeds at 12 evaluations. All methods use the same 200 image inputs and paired FLUX-50 references. Both seed constructions use the same candidate expansion and selection procedure. CAT-OV uses an adaptive number of evaluations. Bold and underlined values mark the best and second-best results, excluding the reference. Blue rows show TrajCal-Correction. HPSv3 and CLIP measure human preference and prompt alignment, respectively.
Setting
RMSE ↓
LPIPS ↓
SSIM ↑
HPSv3 ↑
CLIP ↑
Selection objective
TrajCal-Euler, RMSE only
20.13
0.2171
0.8111
12.51
0.2702
TrajCal-Euler, combined objective
19.83
0.2061
0.8204
12.01
0.2696
TrajCal-Correction, RMSE only
20.23
0.1779
0.8356
12.19
0.2665
TrajCal-Correction, combined objective
20.23
0.1779
0.8356
12.19
0.2665
Correction strength in TrajCal-Correction
Appendix
Table 10: Selection objective, correction strength, and shared prefix. We report mean scores on the same 100 image inputs at 12 evaluations. The objective comparison uses the same retained candidates. We select a schedule separately for each η and repeat screening and selection when prefix alignment is removed.
12 evaluations
16 evaluations
Schedule selected for
Update rule
RMSE ↓
LPIPS ↓
SSIM ↑
RMSE ↓
LPIPS ↓
SSIM ↑
Euler
Euler
19.83
0.2061
0.8204
21.52
0.1909
0.8246
Euler
Correction
20.88
0.1920
0.8256
17.87
0.1411
0.8615
Correction
Euler
20.69
0.2047
0.8204
22.23
0.1883
0.8253
Correction
Correction
20.23
0.1779
0.8356
18.47
0.1382
0.8640
Appendix
Table 11: Euler and correction on fixed schedules. We apply both updates to each selected schedule at 12 and 16 evaluations, using the same 100 image inputs. Time nodes are fixed within each comparison; correction uses η=0.5 . Bold and underlined values mark the best and second-best results at each budget.
Method
RMSE ↓
LPIPS ↓
SSIM ↑
FID ↓
HPSv3 ↑
CLIP ↑
12 evaluations
FLUX-12
49.47
0.4967
0.6335
131.82
13.12
0.2836
GITS
52.29
0.5616
0.6144
152.10
10.37
0.2892
SharpEuler-Z
33.51
0.3320
0.7338
78.31
13.86
0.2874
CAT-OV
35.54
0.3874
0.7056
98.05
12.99
0.2881
OptimalSteps
30.62
0.3327
0.7449
83.12
12.86
0.2912
Appendix
Table 12: Results for photography. Calibration uses photography. Each budget uses the same 30 scenes and paired FLUX-50 references. Bold and underlined values mark the best and second-best results, excluding the reference.
Method
RMSE ↓
LPIPS ↓
SSIM ↑
FID ↓
HPSv3 ↑
CLIP ↑
12 evaluations
FLUX-12
59.74
0.5003
0.6642
171.69
9.49
0.2771
GITS
63.12
0.5650
0.6671
251.06
1.00
0.2696
SharpEuler-Z
46.12
0.3936
0.6883
114.81
12.57
0.2776
CAT-OV
43.52
0.3864
0.6981
122.91
11.41
0.2850
OptimalSteps
43.13
0.4211
0.6742
113.60
12.93
0.2849
Appendix
Table 13: Results for watercolor. Schedules calibrated on photography are used unchanged. Each budget uses the same 30 scenes and paired FLUX-50 references. Bold and underlined values mark the best and second-best results, excluding the reference.
Method
RMSE ↓
LPIPS ↓
SSIM ↑
FID ↓
HPSv3 ↑
CLIP ↑
12 evaluations
FLUX-12
62.26
0.4896
0.6487
139.16
12.08
0.3275
GITS
66.73
0.5483
0.6466
194.30
7.80
0.3236
SharpEuler-Z
43.78
0.3284
0.7263
92.11
13.02
0.3263
CAT-OV
43.94
0.3482
0.7061
109.24
13.11
0.3253
OptimalSteps
49.98
0.4131
0.6787
129.61
13.37
0.3161
Appendix
Table 14: Results for anime. Schedules calibrated on photography are used unchanged. Each budget uses the same 30 scenes and paired FLUX-50 references. Bold and underlined values mark the best and second-best results, excluding the reference.
Method
RMSE ↓
LPIPS ↓
SSIM ↑
FID ↓
HPSv3 ↑
CLIP ↑
12 evaluations
FLUX-12
52.72
0.4387
0.8050
230.12
3.72
0.3041
GITS
56.15
0.4861
0.8096
313.28
-4.03
0.2637
SharpEuler-Z
41.67
0.2956
0.8264
127.47
11.57
0.3197
CAT-OV
40.87
0.3007
0.8206
130.43
11.67
0.3187
OptimalSteps
46.26
0.3480
0.7948
148.11
12.58
0.3227
Appendix
Table 15: Results for flat-vector art. Schedules calibrated on photography are used unchanged. Each budget uses the same 30 scenes and paired FLUX-50 references. Bold and underlined values mark the best and second-best results, excluding the reference.
Method
RMSE ↓
LPIPS ↓
SSIM ↑
FID ↓
12 evaluations
Native Euler
58.24
0.4762
0.7060
143.04
GITS
62.00
0.5331
0.7078
200.20
SharpEuler-Z
43.86
0.3392
0.7470
93.73
CAT-OV
42.78
0.3451
0.7416
101.02
OptimalSteps
46.46
0.3941
0.7159
105.84
Appendix
Table 16: Combined results for the three unseen styles. Schedules calibrated on photography are applied unchanged to watercolor, anime, and flat-vector art. Each budget uses 30 scenes in three styles, giving 90 images per method. RMSE, LPIPS, and SSIM are averaged over these images; FID is computed jointly from all 90 outputs and their paired FLUX-50 references. Bold and underlined values mark the best and second-best results at each budget.
8 evaluations
20 evaluations
Method
RMSE ↓
LPIPS ↓
SSIM ↑
FID↓
RMSE ↓
LPIPS ↓
SSIM ↑
FID↓
Native Euler
53.73
0.5410
0.6011
97.96
41.80
0.3819
0.6944
71.60
GITS
54.52
0.5696
0.6145
130.99
46.23
0.4457
0.6583
87.87
SharpEuler-Z
40.80
0.4258
0.6661
74.68
19.80
0.1634
0.8455
34.34
CAT-OV
39.23
0.4946
0.6552
120.09
31.77
0.2817
0.7597
55.88
OptimalSteps
33.33
0.3881
0.7019
77.19
38.76
0.3533
0.7134
66.40
Appendix
Table 17: Image generation at 8 and 20 evaluations. Both budgets use the same 200 image inputs and paired FLUX-50 references. CAT-OV averages 7.58 and 19.41 evaluations, respectively. Bold and underlined values mark the best and second-best results at each budget.
Figure 6: Video generation at 10 guided evaluations. The two examples use the prompts “a coconut tree by the house” and “road on mountain ridge.” Each example shows frames 1, 41, and 81 of an 81-frame video at 24 fps. The same frames are shown for every method. The top row is the UniPC-50 reference; all other rows use 10 guided evaluations with the same prompt and base noise. TrajCal-UniPC more closely preserves the reference scene layout and appearance, including the pool and house in (a) and the road markings and rock face in (b).
Figure 7: Image generation with EasyCache and TrajCal-Correction. Each row uses one prompt–noise pair. From left to right: the FLUX-50 reference, EasyCache with 13 full model evaluations, and TrajCal-Correction with 12. EasyCache preserves the overall reference composition but leaves visible noise, particularly in the cake background and laptop screen; TrajCal-Correction produces cleaner outputs.
Image generation (FLUX)
Video generation (Wan)
Method
12 evaluations
16 evaluations
Method
10 evaluations
20 evaluations
FLUX
5.115 ± 0.012
6.826 ± 0.016
UniPC
13.959 ± 0.010
27.922 ± 0.018
GITS
5.113 ± 0.027
6.825 ± 0.021
StableVS
13.943 ± 0.013
27.933 ± 0.019
SharpEuler-Z
5.116 ± 0.014
6.828 ± 0.020
OptimalSteps
13.932 ± 0.012
27.905 ± 0.019
CAT-OV
4.952 ± 0.280
6.645 ± 0.339
SharpEuler-Wan
13.928 ± 0.044
27.886 ± 0.091
OptimalSteps
5.114 ± 0.019
6.828 ± 0.012
TrajCal-Euler
13.942 ± 0.017
27.903 ± 0.017
Appendix
Table 18: Image and video generation time. Mean ± sample standard deviation of recorded denoising time in seconds, on the same 500 image inputs or 100 video inputs used in the main comparisons. Budgets count velocity evaluations, with guidance included for video. CAT-OV uses 11.57 and 15.50 evaluations on average at the two image budgets. Decoding is excluded.
Screening
Final selection
Budget
No reuse
With reuse
No reuse
With reuse
8
2,688
1,624
18,432
11,648
12
4,032
2,456
27,648
18,944
16
5,376
3,208
36,864
25,088
20
6,720
4,000
23,040
15,360
Appendix
Table 19: Model calls with and without prefix reuse. Counts are per update rule for image calibration. Screening uses 42 candidates and eight inputs; final selection uses 18 candidates and 128 inputs (64 at budget 20). Each stage starts with an empty cache. Counts exclude reference and anchor generation, decoding, and metric computation.
Flow-based generative modeling learns probability transports from simple noise to data for fast sampling, density estimation, and controllable generation.