Precipitation nowcasting involves two distinct scales of uncertainty: macroscopic storm evolution and fine-scale local convective variations. Current generative ensemble methods entangle both scales within independent random noise draws. This coupling causes two practical limitations: every ensemble member must execute the full sampling trajectory from scratch, so an N-member ensemble costs N times a single forecast; moreover, divergent trajectories spatially misalign heavy-rain echoes across members, causing standard pixel-wise averaging to blur critical high-intensity extremes. We present ForeSay, a generative nowcasting framework that decouples macroscopic evolution prediction from local stochastic detail generation. Given historical radar frames, a vision-language model first predicts a textual scenario of future storm dynamics to condition a Rectified Flow generator. Guided by this shared macroscopic prior, ensemble members share early trunk sampling steps to establish the storm trajectory and branch with fresh noise only in later steps to explore local convective details. Because members then differ mainly in local details, self-anchored support-consensus decoding replaces pixel averaging with discrete support voting to preserve sharp heavy-rain cores. Across six datasets, ForeSay achieves the best CSI at the highest precipitation thresholds, with relative improvements over the strongest baseline on each dataset ranging from over 10% to 84%. Our 8-member ensemble retains the forecasting skill and diversity of eight independently sampled runs while reducing computational cost from 8x to 5.2x that of a single forecast.
Figures & tables
Figure 1: Describing the past vs . predicting the future. Conditioned only on the observed history, LangPrecip [ 22 ] leaves the evolution to each member’s noise (paths illustrative) and repeats every step. ForeSay conditions on the predicted evolution, so members share the early steps and then branch ( 5.2× vs . 8× cost). Bottom: forecasts at +70 min, with red boxes marking the same heavy-echo region.
Figure 2: ForeSay overview. Training (top): (a) a teacher VLM that sees the true future writes the description c∗ ; (b) Qwen3.5-9B is LoRA-tuned to predict c∗ from the history h alone, and c is read from its penultimate layer; (c) with the VLM frozen, the DiT velocity field vθ(zτ,τ;h,c) is trained with Rectified Flow ( h concatenated along time, c via cross-attention, τ via adaLN). Inference (bottom): (d) the tuned VLM predicts c ( c^ for inspection only); (e) after a shared trunk of s∗=4 steps, N=8 members keep the shared x^0 and differ only in the noise endpoint, costing 5.2× instead of 8× ; (f) the anchor’s precipitation support is corrected by a k -of- N vote.
Figure 3: Branched sampling on the Swedish case of Fig. 6 ( +70 min, s∗=4 , N=8 ). Each panel is the endpoint estimate x^0 at that step. The trunk fixes the rain band early; after branching, the members differ in the location and intensity of the heavy cores (dashed box), which the k=2 vote on members ≥35 dBZ ( ≈6.3 mm/h) resolves by adding (ADD) or removing (REM) anchor pixels (right column). All eight members: App. Fig. 10 .
Figure 4: Selection of the branching step s on the held-out set. (a) Peak-normalized posterior width of early and late frames. Below the axis, τs is the noise level of the sampler at step s ( 1 is pure noise, 0 is data). (b) Highest-threshold CSI when branching at step s ( N=8 , λ=1 , k=2 ), with the cost below the axis and the plateau shaded. Dashed lines: ens8 ( 8.0× ). Vertical line: s∗=4 . Numbers in App. Tab. 7 .
SEVIR ( 8096 )
MeteoNet ( 1310 )
Shanghai ( 526 )
CIKM ( 4000 )
CSI ↑
HSS ↑
SSIM ↑
LPIPS ↓
CSI ↑
HSS ↑
SSIM ↑
LPIPS ↓
CSI ↑
HSS ↑
SSIM ↑
LPIPS ↓
CSI ↑
HSS ↑
SSIM ↑
LPIPS ↓
Method
160
181
219
18
24
32
30
35
40
30
35
40
MIMO-VP [ 29 ]
.1492
.1102
.0280
.3792
.6822
.3576
.3838
.2469
.0505
.4016
.8212
.2732
.3722
.2586
.1171
.4332
.7886
.2532
.1933
.1351
.0752
.3198
.6624
.3853
SwinLSTM [ 41 ]
.1585
.1188
.0373
.3892
.6775
.3309
.4558
.3631
.1762
.5129
.8203
.2404
.4409
.3460
.2141
.5212
.7708
.2831
.2990
.2002
.0799
.3885
.6058
.3485
DiffCast [ 49 ]
.1740
.1435
.0750
.4017
.6536
.1983
.4427
.3496
.1767
.5004
.8148
.2364
.4072
.3254
.2286
.4986
.7648
.1658
.2792
.2161
.1261
.4002
.6676
.3091
AlphaPre [ 20 ]
.1754
.1332
.0535
.4100
.6818
.3520
.4315
.3461
.1807
.4936
.8065
.2336
.4514
.3591
.2287
.5267
.7463
.3195
.2604
.1810
.1070
.3674
.6359
.2936
Table 1: Comparison on the four radar extrapolation benchmarks (official test sets, sample counts in parentheses, 128×128 protocol). CSI is given at the three highest thresholds of each benchmark (SEVIR in VIL 0 – 255 , the others in dBZ). ↑ / ↓ : higher/lower is better. Best in bold, second best underlined. Ours is the decoded ensemble with thr at the highest tier. CSI-M and CSI at all thresholds are in App. Tab. 11 .
Figure 5: Highest-threshold CSI versus forecast frame on all six datasets. Ours is the decoded ensemble ( 5.2× ). The strongest baseline of each dataset is drawn in blue and named in the panel, with our relative gain in mean CSI (Tabs. 1 and 2 ). The shaded band is the margin between the two, and grey lines are the remaining baselines. Curves for every metric are in App. Figs. 8 , 11 and 13 – 19 .
Figure 6: Case study of a precipitation event starting on 21 June 2021 at 18:05 UTC, showing a rain band with embedded heavy cores over central Sweden. Left: the radar composite, with the red box marking the 256×256 crop ( 58.5 – 62∘ N, 13 – 19∘ E), and the Meteosat IR10.8 image at 18:00 UTC. Right: the four input frames, the ground truth and forecasts at +10 to +80 min. Dashed boxes mark a region of heavy echoes at later lead times. The predicted scenario c^ for this case is shown at the bottom.
Swedish
MRMS
Method
CSI 0.06↑
CSI 6.3↑
SSIM ↑
LPIPS ↓
CSI 1↑
CSI 8↑
SSIM ↑
LPIPS ↓
PySteps [ 31 ]
0.473
0.049
0.589
0.284
0.338
0.024
0.871
0.187
DGMR [ 33 ]
0.433
0.039
0.702
0.204
0.367
0.022
0.883
0.137
DiffCast [ 49 ]
0.439
0.029
0.660
0.292
0.328
0.018
0.869
0.183
SSLDM-ISI [ 21 ]
0.558
0.040
0.718
0.191
0.416
0.024
0.897
0.120
AlphaPre [ 20 ]
0.541
0.025
0.647
0.333
0.360
0.024
0.891
0.197
Table 2: Comparison on the Swedish precipitation dataset and MRMS ( 80 -minute lead time, averaged over 16 frames). Ours is the decoded ensemble ( s=4 , N=8 , 5.2× ) with thr at the highest tier. LangPrecip on MRMS uses CFG scale 4 . Best in bold, second best underlined. FSS and CRPS are in App. Tabs. 15 and 16 .
w/o branching ( 1× )
w/ branching ( 5.2× )
Condition
CSI 0.06↑
CSI 6.3↑
SSIM ↑
FSS ↑
CSI 6.3↑
Radar only
0.578
0.0895
0.736
0.910
0.1273
+ history text
0.584
0.0949
0.743
0.914
0.1275
+ future text
0.593
0.0984
0.747
0.915
0.1357
Table 3: Conditioning ablation on the Swedish precipitation dataset. History text describes the observed frames as in LangPrecip, and future text is our predicted scenario. W/o branching: the anchor. W/ branching: the decoded ensemble.
Motion direction
Intensity and structure
Main
L/R
U/D
8-way
Intens.
Decay
Split
Merge
95.1
90.9
90.8
84.2
76.7
73.6
75.4
90.8
Table 4: Agreement (%, ↑ ) of the predicted c^ with the ground truth. Main: primary axis; L/R, U/D: left/right, up/down; 8-way: compass direction.
Figure 7: Branching and aggregation ablations. (a) CSI gain over the anchor and (b) member IoU at frame 12 versus the noise strength λ (held-out set), with ens8 dashed. (c) Heavy-rain CSI per lead time for different aggregations of the same N=8 members (test set, FB not comparable to Tab. 6 ). All-frame CSI (FB): ours 0.1357 ( 1.18 ), PMM 0.1165 ( 0.73 ), consensus k=40.1002 ( 0.62 ), anchor 0.0984 ( 0.73 ), ensemble mean 0.0852 ( 0.50 ).
CSI ↑ , highest tier
Δ (%), other metrics
Dataset
Base.
Anchor
Ours
Δ (%)
CSI 2nd
SSIM
LPIPS
CRPS
Swedish
.074
.0984
.1357
+37.9
0.0
0.0
+0.1
− 6.8
MRMS
.031
.0278
.0360
+29.5
0.0
0.0
− 0.2
− 0.3
SEVIR
.1199
.1202
.1458
+21.3
+4.1
+0.6
+0.6
+1.9
MeteoNet
.2679
.2747
.2971
+8.2
+0.3
+0.2
+0.1
+0.5
Shanghai
.3032
.3217
.3387
+5.3
+0.4
0.0
+0.1
− 0.1
Table 5: Effect of support-consensus decoding. Anchor ( 1.0× ) vs . decoded ensemble ( 5.2× ), paired on the same anchor. Base.: the strongest baseline (Tabs. 1 and 2 ). Δ : relative change (%) of Ours over the anchor, with the sign set so that positive is better. Radar CSI uses the 128×128 protocol, and the other metrics use the native 256×256 output. Full values in App. Tabs. 13 and 14 .
k
1
2
3
4
5
6
7
8
CSI ↑
.1760
.1740
.1518
.1266
.1036
.0821
.0608
.0420
FB →1
1.92
1.05
0.70
0.50
0.37
0.27
0.19
0.12
N
1
2
4
6
8
16
CSI ↑
.1239
.0893
.1399
.1649
.1740
.1867
FB →1
0.66
0.33
0.65
0.88
1.05
1.53
Table 6: Number of votes k (top, N=8 ) and ensemble size N (bottom, k=2 ) on the held-out set. CSI is at ≥6.3 mm/h and FB is the frequency bias ( 1 is unbiased). N=1 is the anchor. Full sweeps are in App. Tabs. 8 and 9 .
Appendix figures & tables24 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
Steps
Cost
CSI ↑ (all 16 frames) (of 8× )
CSI ↑ ( f=8 – 16 ) (of 8× )
Member IoU ( f=12 )
s=2 , N=8
66
6.6×
0.1743 ( 97% )
0.1366 ( 91% )
0.151
s=3 , N=8
59
5.9×
0.1753 ( 85% )
0.1379 ( 87% )
0.174
s=4 , N=8 (default)
52
5.2×
0.1740 ( 100% )
0.1355 ( 94% )
0.166
s=5 , N=8
45
4.5×
0.1699 ( 102% )
0.1339 ( 103% )
0.190
s=6 , N=8
38
3.8×
0.1688 ( 93% )
0.1275 ( 82% )
0.220
s=7 , N=8
31
3.1×
0.1632 ( 79% )
0.1244 ( 73% )
0.228
Appendix
Table 7: Sweep of the branching step s on the Swedish precipitation held-out set ( N=8 , T=10 , λ=1 , k=2 ). Steps is the number of network evaluations per forecast, s+N(T−s) , and cost is relative to a single sample ( 10 steps). CSI is at the highest tier ( ≥6.3mm/h ), computed per sample and frame and then averaged. “Of 8× ” is the paired CSI gain of each setting over its own anchor (the first branch of the same run), as a fraction of the gain of ens8 over its own first member. The anchor is re-noised at its own branching step, so this fraction need not follow the order of the absolute CSI. Member IoU is the intersection over union of the precipitation supports of all N members in the 12 th forecast frame, and lower means more diverse members.
k
CSI (all 16 frames) ↑
CSI ( f=8 – 16 ) ↑
FB (closer to 1 is better)
1
0.1760
0.1449
1.92
2
0.1740
0.1355
1.05
3
0.1518
0.1099
0.70
4
0.1266
0.0862
0.50
5
0.1036
0.0652
0.37
6
0.0821
0.0455
0.27
Appendix
Table 8: Sweep of the number of votes k (Swedish precipitation held-out set, 500 samples, s=4 , N=8 , λ=1 ). CSI is at the highest tier ( ≥6.3mm/h ). FB is the frequency bias, the number of forecast pixels above the threshold divided by the observed number, so 1 is unbiased, >1 overestimates and <1 underestimates the rain area. k=1 takes the union of all member supports, and k=2 is the smallest vote that forms a consensus.
N
Cost
CSI ( f=1 – 4 ) ↑
CSI ( f=8 – 16 ) ↑
CSI (all 16 frames) ↑
FB (closer to 1 is better)
1 (anchor, no vote)
1.0×
0.1986
0.0892
0.1239
0.66
2
1.6×
0.1641
0.0553
0.0893
0.33
4
2.8×
0.2186
0.1009
0.1399
0.65
6
4.0×
0.2429
0.1245
0.1649
0.88
8
5.2×
0.2481
0.1355
0.1740
1.05
16
10.0×
0.2547
0.1538
0.1867
1.53
Appendix
Table 9: Sweep of the number of members N with the vote fixed at k=2 (Swedish precipitation held-out set, 500 samples, s=4 , λ=1 ). CSI is at the highest tier ( ≥6.3mm/h ). Cost is s+N(T−s) network evaluations relative to a single sample ( 10 steps). FB is the frequency bias, and 1 is unbiased.
Dataset
Evaluation split
Samples
Condition frames → forecast frames
Intensity thresholds
SEVIR (VIL)
official test
8096
5→20
16/74/133/160/181/219
MeteoNet
test
1310
5→20
12/18/24/32
Shanghai
official test
526
5→20
20/30/35/40
CIKM
test
4000
5→10
20/30/35/40
Appendix
Table 10: Evaluation settings of the four radar extrapolation benchmarks. The sample count is the number of test samples, and each sample has 5 input frames. Thresholds follow the convention of each benchmark, in VIL counts ( 0 – 255 ) for SEVIR and in dBZ for the others.
SEVIR ( 8096 )
MeteoNet ( 1310 )
Shanghai ( 526 )
CIKM ( 4000 )
Method
CSI-M ↑
CSI-16 ↑
CSI-74 ↑
CSI-133 ↑
CSI-160 ↑
CSI-181 ↑
CSI-219 ↑
CSI-M ↑
CSI-12 ↑
CSI-18 ↑
CSI-24 ↑
CSI-32 ↑
CSI-M ↑
CSI-20 ↑
CSI-30 ↑
CSI-35 ↑
CSI-40 ↑
CSI-M ↑
CSI-20 ↑
CSI-30 ↑
CSI-35 ↑
CSI-40 ↑
MIMO-VP [ 29 ]
0.2931
0.6643
0.5562
0.2509
0.1492
0.1102
0.0280
0.2878
0.4701
0.3838
0.2469
0.0505
0.3108
0.4954
0.3722
0.2586
0.1171
0.2492
0.5932
0.1933
0.1351
0.0752
SwinLSTM [ 41 ]
0.3110
0.6761
0.6015
0.2738
0.1585
0.1188
0.0373
0.3778
0.5161
0.4558
0.3631
0.1762
0.3839
0.5347
0.4409
0.3460
0.2141
0.3002
0.6217
0.2990
0.2002
0.0799
DiffCast [ 49 ]
0.3083
0.6586
0.5410
0.2578
0.1740
0.1435
0.0750
0.3693
0.5084
0.4427
0.3496
0.1767
0.3655
0.5010
0.4072
0.3254
0.2286
0.3067
0.6053
0.2792
0.2161
0.1261
AlphaPre [ 20 ]
0.3249
0.6984
0.6007
0.2883
0.1754
0.1332
0.0535
0.3627
0.4926
0.4315
0.3461
0.1807
0.3972
0.5497
0.4514
0.3591
0.2287
0.2853
0.5928
0.2604
0.1810
0.1070
PercpCast [ 8 ]
0.3492
0.6856
0.5844
0.3101
0.2203
0.1929
0.1022
0.4039
0.5279
0.4737
0.3933
0.2207
0.3995
0.5316
0.4459
0.3628
0.2577
0.3163
0.6108
0.2766
0.2223
0.1553
Appendix
Table 11: CSI-M and CSI at every intensity threshold on the four radar extrapolation benchmarks (official test sets, 128×128 protocol, sample counts in parentheses). Best in bold, second best underlined. CSI-M is the CSI averaged over thresholds. The subscripts are the thresholds of each benchmark, in VIL ( 0 – 255 ) for SEVIR and in dBZ for MeteoNet, Shanghai and CIKM.
Metric
Protocol
SEVIR
MeteoNet
Shanghai
CIKM
CSI-M ↑
256×256
0.3484→0.3579
0.4081→0.4153
0.4456→0.4509
0.3482→0.3525
128×128
0.3604→0.3706
0.4407→0.4470
0.4668→0.4714
0.3451→0.3507
HSS ↑
256×256
0.4531→0.4659
0.5454→0.5550
0.5835→0.5901
0.4268→0.4329
128×128
0.4664→0.4800
0.5779→0.5860
0.6041→0.6098
0.4468→0.4546
SSIM ↑
256×256
0.7019→0.7061
0.8365→0.8385
0.8346→0.8350
0.7066→0.7057
128×128
0.6926→0.6978
0.8439→0.8451
0.8218→0.8222
0.6757→0.6754
Appendix
Table 12: Our results under the two evaluation resolutions (single model → ensemble). The ensemble setting is that of Table 14 , and the samples and forecast setting are those of Table 10 . 256×256 is the native output and 128×128 the protocol of prior work.
Swedish
MRMS
CSI ↑ , highest
( 6.3 ) 0.0984→0.1357
( 8 ) 0.0278→0.0360
CSI ↑ , weak
( 0.06 ) 0.5926→0.5927
( 1 ) 0.4346→0.4347
FSS ↑
0.9149→0.9148
0.7052→0.7034
SSIM ↑
0.7467→0.7466
0.8985→0.8983
CRPS ↓
0.1159→0.1238
0.1360→0.1364
LPIPS ↓
0.1897→0.1896
0.1306→0.1308
Appendix
Table 13: Anchor (single model, 1.0× ) → decoded ensemble ( 5.2× ) on the Swedish precipitation dataset and MRMS, paired on the same anchor ( 16 frames, 80 -minute lead time). Thresholds in mm/h. The better value of each pair is in bold. This table gives the exact values behind Tab. 5 of the main text.
SEVIR ( 8096 )
MeteoNet ( 1310 )
Shanghai ( 526 )
CIKM ( 4000 )
CSI ↑ by intensity level, highest to lowest ( 128×128 protocol)
Highest
(219) 0.1202→0.1458
(32) 0.2747→0.2971
(40) 0.3217→0.3387
(40) 0.1744→0.1941
2nd highest
(181) 0.2087→0.2173
(24) 0.4354→0.4369
(35) 0.4351→0.4369
(35) 0.2574→0.2598
3rd
(160) 0.2405→0.2477
(18) 0.5049→0.5058
(30) 0.5171→0.5166
(30) 0.3166→0.3172
4th
(133) 0.3258→0.3341
(12) 0.5479→0.5482
(20) 0.5933→0.5933
(20) 0.6321→0.6317
5th
(74) 0.5901→0.5974
—
—
—
Appendix
Table 14: Anchor alone versus decoded ensemble on the official test sets of the four radar benchmarks. Each cell reads “anchor ( 1.0× ) → ensemble ( 5.2× )”, with the better value in bold. CSI and HSS follow the 128×128 protocol of Table 1 of the main text and Table 11 . The threshold of each tier is given in parentheses (VIL 0 – 255 for SEVIR, dBZ for the others), and HSS is averaged over thresholds. MAE, RMSE, CRPS, SSIM and LPIPS are computed on the native 256×256 output, since the table contains no baseline. MAE and RMSE are in the original units of each dataset and are not comparable across datasets. Appendix F compares the two resolutions.
Method
CSI 0.06↑
CSI 6.3↑
FSS ↑
CRPS ↓
SSIM ↑
LPIPS ↓
PySteps [ 31 ]
0.473
0.049
0.872
0.144
0.589
0.284
DGMR [ 33 ]
0.433
0.039
0.882
0.143
0.702
0.204
DiffCast [ 49 ]
0.439
0.029
0.903
0.149
0.660
0.292
SSLDM-ISI [ 21 ]
0.558
0.040
0.906
0.122
0.718
0.191
AlphaPre [ 20 ]
0.541
0.025
0.894
0.149
0.647
0.333
DTCA [ 16 ]
0.572
0.046
0.911
0.115
0.732
0.180
Appendix
Table 15: Comparison on the Swedish precipitation dataset ( 1000 test samples, 4→16 frames, native 256×256 , mean over the 16 forecast frames). CSI thresholds are in mm/h. The CSI, SSIM and LPIPS columns repeat Table 2 . Best in bold, second best underlined.
Figure 8: Skill versus lead time on the Swedish precipitation dataset ( 4→16 frames, 1000 test samples, native 256×256 ). The horizontal axis of every panel is the lead time in minutes. Ours is the decoded ensemble ( 5.2× ). The FSS panel is cut at 0.8 , so RainHCNet and SSRF-Net, whose FSS stays below this value, are not visible there.
Figure 9: Forecast examples on the Swedish precipitation dataset. Each panel shows the four input frames, the ground truth and the forecast of every compared method at every second forecast frame ( +10 to +80 min). Below each panel is the scenario c^ that the fine-tuned VLM predicts from the input frames alone.
Figure 10: All eight members of the branched ensemble in Fig. 3 of the main text at five lead times ( s∗=4 , N=8 ). The members share the position, orientation and motion of the rain band and differ in the location and extent of the heavy cores. Vote count : number of members at or above 35 dBZ ( ≈6.3 mm/h) at each pixel. Ours (ens.) : the field delivered by support-consensus decoding.
Method
CSI 1↑
CSI 8↑
FSS ↑
CRPS ↓
SSIM ↑
LPIPS ↓
PySteps [ 31 ]
0.338
0.024
0.579
0.167
0.871
0.187
DGMR [ 33 ]
0.367
0.022
0.633
0.154
0.883
0.137
DiffCast [ 49 ]
0.328
0.018
0.613
0.168
0.869
0.183
SSLDM-ISI [ 21 ]
0.416
0.024
0.704
0.135
0.897
0.120
AlphaPre [ 20 ]
0.360
0.024
0.699
0.144
0.891
0.197
DTCA [ 16 ]
0.423
0.026
0.711
0.131
0.899
0.118
Appendix
Table 16: Comparison on MRMS ( 4→16 frames, native 256×256 , mean over the 16 forecast frames). CSI thresholds are in mm/h. LangPrecip uses CFG scale 4 . The CSI, SSIM and LPIPS columns repeat Table 2 . Best in bold, second best underlined.
Figure 11: Skill versus lead time on MRMS ( 4→16 frames, native 256×256 ). The horizontal axis of every panel is the lead time in minutes. Ours is the decoded ensemble ( 5.2× ).
Figure 12: Forecast examples on MRMS. Each panel shows the four input frames, the ground truth and the forecast of every compared method at every second forecast frame ( +10 to +80 min). Below each panel is the scenario c^ that the fine-tuned VLM predicts from the input frames alone.
Figure 13: Skill versus forecast frame on SEVIR ( 5→20 frames, 128×128 protocol). The horizontal axis of every panel is the forecast frame. Thresholds are VIL levels ( 0 – 255 ). FAR is undefined where a method forecasts no pixel above a threshold. Such points are left blank, and the threshold average of FAR skips them. After frame 6 SwinLSTM forecasts almost no pixel above 219 (POD <0.01 ), so its FAR at this threshold rests on very few pixels.
Figure 14: Forecast examples on SEVIR. Each panel shows the five input frames, the ground truth and the forecast of every compared method at every second forecast frame. Below each panel is the scenario c^ that the fine-tuned VLM predicts from the input frames alone.
Figure 15: Skill versus forecast frame on MeteoNet ( 5→20 frames, 128×128 protocol). The horizontal axis of every panel is the forecast frame.
Figure 16: Forecast examples on MeteoNet. Each panel shows the five input frames, the ground truth and the forecast of every compared method at every second forecast frame. Below each panel is the scenario c^ that the fine-tuned VLM predicts from the input frames alone.
Figure 17: Skill versus forecast frame on Shanghai ( 5→20 frames, 128×128 protocol). The horizontal axis of every panel is the forecast frame.
Figure 18: Forecast examples on Shanghai. Each panel shows the five input frames, the ground truth and the forecast of every compared method at every second forecast frame. Below each panel is the scenario c^ that the fine-tuned VLM predicts from the input frames alone.
Figure 19: Skill versus forecast frame on CIKM ( 5→10 frames, 128×128 protocol). The horizontal axis of every panel is the forecast frame.
Figure 20: Forecast examples on CIKM. Each panel shows the five input frames, the ground truth and the forecast of every compared method at every forecast frame. Below each panel is the scenario c^ that the fine-tuned VLM predicts from the input frames alone.
School of Intelligence Science and Technology, Nanjing University, Suzhou, 215163, China. · State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, 210023, China.