betaFeedback

WorksComputer Science

ForeSay: Tell the Future in Words, Branch the Details in Noise for Precipitation Nowcasting

Abstract

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.

Introduction

Precipitation nowcasting supports disaster mitigation, urban flood control and severe weather warning [33], and is commonly formulated in computer vision as video prediction of radar echo sequences. Unlike natural video generation, which pursues overall visual realism, extreme precipitation is sparse in space: it covers only a tiny fraction of pixels yet decides whether a warning succeeds, so optimizing average fidelity tends to smooth these cores away. Driven by chaotic, nonlinear convection, future precipitation is also intrinsically uncertain at two levels: the macroscopic evolution of the system, i.e. where it moves, whether it intensifies or decays, and whether it splits or reorganizes; and the local detail that remains uncertain once the evolution is fixed, such as the exact location, shape and intensity of convective cells.

Deterministic models [10, 11, 37, 46] regress toward the conditional mean and blur heavy echoes at long lead times. Generative models [12, 13, 33, 43, 49] instead sample from the conditional distribution, preserving fine details, and are widely used to build ensemble nowcasts. Such ensembles, however, face two problems. First, an NN-member ensemble costs NN times a single forecast. Second, members place heavy echoes at different locations, so pixel-wise averaging smooths them out and the aggregate loses the extremes. Both problems share one cause. Each member entangles the macroscopic evolution with the local detail in its own noise. Although this evolution can largely be inferred from the radar history, it is never separated out for members to share, so the early, high-noise steps that settle it are repeated across members, and the evolutions they settle diverge. LangPrecip [22] introduced natural-language motion descriptions as an explicit condition, but its text summarizes past observations rather than forecasting how the system will evolve, so the future evolution is still left to noise. The question is thus how ensemble members can share a single forecast of the storm system’s overall evolution, so that ensembles preserving local heavy-rain echoes can be generated efficiently.

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.

We therefore decouple the prediction of the macroscopic evolution from the generation of local stochastic detail, and build ForeSay on this idea (Fig. 1). A vision–language model first predicts the future evolution in text, which conditions a Rectified Flow [23, 24] generator. Conditioned on this same text, all ensemble members share early trunk sampling steps along the forward trajectory and inject fresh noise only to explore local details. Differing mainly in local details, members are decoded by self-anchored consensus voting instead of pixel averaging, preserving sharp heavy echoes.

Our contributions are summarized as follows:

  • Shared forecast of macroscopic evolution: Existing generative ensembles leave the storm system’s evolution to each member’s noise, or to text that only describes the past, with no explicit forecast of that evolution. We propose ForeSay, which uses a vision–language model to predict a textual scenario of the future evolution directly from radar history, giving all ensemble members a shared macroscopic prior.

  • Efficient branched sampling and consensus decoding: Independent ensemble sampling is costly, and pixel averaging smooths out heavy echoes. We propose training-free branched Rectified Flow sampling, in which members share early trunk steps and branch with fresh noise only in later steps, and self-anchored support-consensus decoding, which replaces pixel averaging with pixel-wise voting to avoid averaging blur. The branched ensemble matches the skill and diversity of 8 independent members at 5.2× compute.

  • Comprehensive benchmarks and validation: On six datasets, ForeSay achieves the best CSI at the highest intensity threshold, improving over the strongest baseline by 11%–84% (Table 1 and Table 2). Ablations, hyperparameter analyses, and qualitative results verify each component.

MethodSEVIR (8096) CSI↑ 160HSS↑ 181SSIM↑ 219LPIPS↓ 18MeteoNet (1310) CSI↑ 24HSS↑ 32SSIM↑ 30LPIPS↓ 35Shanghai (526) CSI↑ 30HSS↑ 35SSIM↑ 40LPIPS↓ 30CIKM (4000) CSI↑ 30HSS↑ 35SSIM↑ 40LPIPS↓
MIMO-VP [29].1492.1102.0280.3792.6822.3576.3838.2469.0505.4016.8212.2732.3722.2586.1171.4332.7886.2532
SwinLSTM [41].1585.1188.0373.3892.6775.3309.4558.3631.1762.5129.8203.2404.4409.3460.2141.5212.7708.2831
DiffCast [49].1740.1435.0750.4017.6536.1983.4427.3496.1767.5004.8148.2364.4072.3254.2286.4986.7648.1658
AlphaPre [20].1754.1332.0535.4100.6818.3520.4315.3461.1807.4936.8065.2336.4514.3591.2287.5267.7463.3195
PercpCast [8].2203.1929.1022.4538.6852.1884.4737.3933.2207.5436.8266.1238.4459.3628.2577.5404.7942.1918
exPreCast [39].2133.1649.0716.4355.6388.3237.4704.3833.2090.5346.8051.2375.4399.3471.2323.5250.7144.3224
FlowCast [34].2149.1661.0638.4454.7254.2631.4705.3715.1803.5177.8345.1856.4815.3927.2622.5642.8003.2474
SDIR [52].2201.1828.0934.4578.7408.2869.5037.4220.2349.5665.8423.1843.4919.4040.2797.5768.8213.2097
PDRF [26].2408.2110.1199.4678.6870.2068.5053.4285.2679.5746.8385.1408.4856.4066.3032.5772.7938.1827
Ours.2477.2173.1458.4800.6978.1670.5058.4369.2971.5860.8451.1000.5166.4369.3387.6098.8222.1302

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.

MethodSwedishMRMS
CSI0.06_{0.06} ↑CSI6.3_{6.3} ↑SSIM↑LPIPS↓CSI1_1 ↑CSI8_8 ↑SSIM↑LPIPS↓
PySteps [31]0.4730.0490.5890.2840.3380.0240.8710.187
DGMR [33]0.4330.0390.7020.2040.3670.0220.8830.137
DiffCast [49]0.4390.0290.6600.2920.3280.0180.8690.183
SSLDM-ISI [21]0.5580.0400.7180.1910.4160.0240.8970.120
AlphaPre [20]0.5410.0250.6470.3330.3600.0240.8910.197
DTCA [16]0.5720.0460.7320.1800.4230.0260.8990.118
RainHCNet [45]0.4990.0250.3320.3200.3200.0020.8770.207
SSRF-Net [25]0.4850.0450.5720.3100.3360.0090.8440.213
exPreCast [39]0.4890.0280.5200.3480.3940.0150.8280.203
FlowCast [34]0.4820.0090.6800.3240.3930.0130.9000.192
SDIR [52]0.4240.0190.7180.3070.3490.0130.9070.189
PDRF [26]0.4740.0370.5710.3140.3830.0200.8750.190
LangPrecip [22]0.5860.0740.7320.1930.4350.0310.9050.142
Ours0.5930.1360.7470.1900.4350.0360.8980.131

Table 2. Comparison on the Swedish precipitation dataset and MRMS (80-minute lead time, averaged over 16 frames). Ours is the decoded ensemble (s=4s = 4, N=8N = 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.

Related work

Precipitation nowcasting as video prediction. Deterministic nowcasters [10, 11, 20, 37, 39, 40, 46, 52] regress toward the conditional mean and blur heavy echoes at long lead times. Following generative video models [3, 5, 14, 43], adversarial [33], diffusion [12, 13, 49] and flow-based [8, 26, 34] nowcasters preserve sharp details and express uncertainty with multi-member ensembles, but each member entangles the macroscopic evolution and the local detail in one noise draw. Every member therefore recomputes the full trajectory, and when members diverge, averaging blurs the heavy echoes again. ForeSay predicts the evolution once as a shared text condition and leaves only the local detail to the noise.

Language conditions for video and weather. Text-conditioned video generators [6, 18, 19, 38] follow user prompts rather than physical dynamics, while meteorological language models mostly act as backbones [9, 36], foundation models [51] or event predictors [17, 42]. LangPrecip [22] first conditioned a precipitation generator on motion text, yet its text only summarizes past observations and does not constrain future advection, intensity or reorganization. Omni-Weather [53] reasons via chain-of-thought in a unified backbone, improving LPIPS at the cost of CSI. ForeSay instead predicts the future evolution from radar history, checked against the ground truth (Tab. 4).

Shared sampling and ensemble aggregation. Existing trajectory-sharing and re-noising methods accelerate or edit single samples: DeepCache [27] reuses features across adjacent steps, SAGE [50] shares early steps across similar prompts, expert denoisers [1, 44] split the noise stages across networks, and re-noising [28, 48] serves editing or refinement. None is designed for ensemble forecasting, which needs members that share the macroscopic evolution and differ only in local details. Branched sampling shares a trunk under predicted evolution and replaces only each member’s noise endpoint. For aggregation, probability matching (PMM) [4, 7] maps member amplitudes onto the blurry ensemble-mean pattern, whereas support-consensus decoding avoids the mean field and corrects one anchor member by a pixel-wise vote (Sec. 5.3).

Method

A generative nowcaster draws each forecast from a random noise sample, and this sample entangles the macroscopic storm evolution with local convective details. In an ensemble, every member therefore iterates the full sampling chain from scratch to trace its own evolution path, so an NN-member ensemble costs NN times a single forecast; members with divergent evolutions also misalign heavy echoes in space, which simple pixel averaging blurs away. ForeSay addresses these two problems with three coordinated components (Fig. 2): (i) future scenario condition learning (Sec. 3.1) uses a vision–language model (VLM) to predict the macroscopic evolution in text as a shared prior cc; (ii) branched ensemble sampling (Sec. 3.2) shares early trunk sampling steps across all members to cut compute; and (iii) self-anchored support-consensus decoding (Sec. 3.3) conducts pixel-wise consensus voting across members to avoid averaging blur.

Problem setup. Given LL observed radar frames hh, nowcasting predicts the next FF frames xx. A frozen VAE [2, 35] with encoder E\mathcal{E} and decoder D\mathcal{D} maps each 256×256256 \times256 frame to a 32×32×3232 \times32 \times32 latent, so each frame ff of x0=E(x)x_0 = \mathcal{E}(x) has df=32768d_f = 32768 coordinates. The generator is a Rectified Flow [24] in this latent space. Its velocity field vθ(zτ,τ;h,c)v_\theta(z_\tau,\tau;h,c) with parameters θ\theta is evaluated at the latent state zτz_\tau at time τ∈[0,1]\tau\in[0,1], is conditioned on the history latent E(h)\mathcal{E}(h) and a text condition cc (Sec. 3.1), and transports Gaussian noise at τ=1\tau= 1 to the data at τ=0\tau= 0. Sampling takes TT Euler steps on a grid 1=τ0>τ1>⋯>τT=01 = \tau_0 > \tau_1 > \cdots> \tau_T = 0, so the state after ss steps lies at time τs\tau_s, and D\mathcal{D} decodes the final state.

Future Scenario Condition Learning

To give all members a shared forecast of the macroscopic storm evolution, the future scenario condition learning module turns radar history into a textual scenario condition cc. Specifically, a vision–language model (VLM) reads the observed sequence hh to predict a text scenario describing how radar echoes will evolve, covering advection direction, intensity trends and structural reorganization. The hidden states of this text condition the Rectified Flow generator. Branched ensemble sampling (§ 3.2) relies on cc to share the early sampling steps across members, and support-consensus decoding (§ 3.3) relies on the resulting members differing mainly locally. Unlike prior approaches such as LangPrecip [22] that only describe observed historical frames and leave future evolution unconstrained, ForeSay directly forecasts the future evolution itself, providing explicit guidance on the macroscopic evolution for downstream generation.

Training. Training has three stages (top row of Figure 2). (i) A teacher VLM (Qwen-VL-Max) sees hh and the true future xx and writes a reference description c∗c^* offline, used only as supervision. (ii) A student VLM, Qwen3.5-9B [32] with LoRA [15] on all linear layers, learns to generate c^\hat{c} from hh alone by minimizing the token-level cross-entropy

ForeSay overview

Figure 2. ForeSay overview. Training (top): (a) a teacher VLM that sees the true future writes the description c∗c^*; (b) Qwen3.5-9B is LoRA-tuned to predict c∗c^* from the history hh alone, and cc is read from its penultimate layer; (c) with the VLM frozen, the DiT velocity field vθ(zτ,τ;h,c)v_\theta(z_\tau,\tau;h,c) is trained with Rectified Flow (hh concatenated along time, cc via cross-attention, τ\tau via adaLN). Inference (bottom): (d) the tuned VLM predicts cc (c^\hat{c} for inspection only); (e) after a shared trunk of s∗=4s^* = 4 steps, N=8N = 8 members keep the shared x^0\hat{x}_0 and differ only in the noise endpoint, costing 5.2×5.2\times instead of 8×8\times; (f) the anchor’s precipitation support is corrected by a kk-of-NN vote.

LLM(ϕ)=−∑j=1∣c∗∣log⁡pϕ(cj∗∣c<j∗,h)(1)\mathcal{L}_{\mathrm{LM}}(\phi)=-\sum_{j=1}^{|c^*|}\log p_{\phi}\left(c_j^* \mid c_{<j}^*,h\right) \tag*{(1)}

where ∣c∗∣|c^*| is the number of tokens in c∗c^*, cj∗c_j^* is its jj-th token, c<j∗c_{<j}^* are the tokens before it, pϕp_{\phi} is the next-token distribution of the student and ϕ\phi are its LoRA parameters. From the tokens c^\hat{c} that the student generates autoregressively given hh, we take the penultimate-layer hidden states, padded to 256 tokens, as c∈R256×4096c \in\mathbb{R}^{256 \times4096}. (iii) With the weights of the student VLM and the VAE frozen, only the parameters θ\theta of the DiT [30] velocity field are trained, with the Rectified Flow loss

LRF(θ)=Ex0,ε,τ∥vθ(zτ,τ;h,c)−(x0−ε)∥2(2)\mathcal{L}_{\mathrm{RF}}(\theta)=\mathbb{E}_{x_0,\varepsilon,\tau}\left\|v_{\theta}(z_{\tau},\tau;h,c)-(x_0-\varepsilon)\right\|^2 \tag*{(2)}

where x0=E(x)x_0=\mathcal{E}(x) is the latent of the true future, ε∼N(0,I)\varepsilon\sim\mathcal{N}(0,I) is Gaussian noise of the same shape, zτ=(1−τ)x0+τεz_{\tau}=(1-\tau)x_0+\tau\varepsilon is the point at time τ\tau on the straight path from x0x_0 to ε\varepsilon, and cc is the condition the student predicts for hh, as at inference. The target x0−εx_0-\varepsilon is the constant velocity of this path toward the data. E(h)\mathcal{E}(h) is concatenated with zτz_{\tau} along the time axis, cc enters through cross-attention and τ\tau through adaLN. The generator never sees the teacher’s c∗c^*, and the text c^\hat{c} serves only for inspection.

Branched Ensemble Sampling

Given the evolution condition cc, drawing NN members by independent sampling would cost NN times a single forecast. However, the early high-noise steps settle the coarse structure, as the trunk in the top row of Figure 3 shows. Under a shared scenario of the macroscopic evolution, independent members repeat nearly the same computation in these steps. To reduce this redundancy, we propose training-free branched ensemble sampling. Rather than rerunning the full trajectory, all members share the early steps as a single “trunk” to settle the macroscopic evolution, and then branch at step ss by injecting fresh random noise to explore local convective details independently. This strategy lowers the computational cost of an 8-member ensemble from 8×8\times to 5.2×5.2\times that of a single forecast, and the resulting members differ mainly in local details, as the subsequent consensus decoding requires (§ 3.3). Below, we first formalize the branching operator, and then address its two critical design choices: which evolution the trunk should convey, and at which step ss branching should occur.

Figure 3. Branched sampling on the Swedish case of Figure 6 (+70 min, s∗=4s^* = 4, N=8N = 8). Each panel is the endpoint estimate x^0\hat{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=2k = 2 vote on members ≥35\ge35 dBZ (≈6.3\approx6.3 mm/h) resolves by adding (ADD) or removing (REM) anchor pixels (right column). All eight members: App. Fig. 10.

Branching operator. Let zz be the trunk state after ss shared steps, τs∈(0,1)\tau_s\in(0,1) the corresponding noise level (on the decreasing sampling grid 1=τ0>τ1>⋯>τT=01=\tau_0>\tau_1>\cdots>\tau_T=0, where τ=1\tau=1 denotes pure noise and τ=0\tau=0 data), and vv the velocity already computed at the last trunk step. The estimates of the data endpoint x0x_0 and the noise endpoint ε\varepsilon of the path through x^0=z+τsv\hat{x}_0=z+\tau_s v and ε^=z−(1−τs)v\hat{\varepsilon}=z-(1-\tau_s)v, satisfy z=(1−τs)x^0+τsε^z=(1-\tau_s)\hat{x}_0+\tau_s\hat{\varepsilon} exactly. Each of the NN members b=0,…,N−1b=0,\ldots,N-1 keeps x^0\hat{x}_0 and mixes ε^\hat{\varepsilon} with fresh noise εb∼N(0,I)\varepsilon_b\sim\mathcal{N}(0,I) at strength λ∈[0,1]\lambda\in[0,1],

zb=(1−τs)x^0+τs(1−λ2 ε^+λεb)(3)z_b=(1-\tau_s)\hat{x}_0+\tau_s\left(\sqrt{1-\lambda^2}\,\hat{\varepsilon}+\lambda\varepsilon_b\right) \tag*{(3)}

where the two weights keep the noise term at unit variance. With λ=1\lambda=1, used throughout (§ 5.2),

zb=(1−τs)x^0+τsεb=z+τs(εb−ε^).(4)z_b=(1-\tau_s)\hat{x}_0+\tau_s\varepsilon_b=z+\tau_s(\varepsilon_b-\hat{\varepsilon}). \tag*{(4)}

The branching operator replaces the noise component of the trunk and leaves its data component unchanged. Every member thus starts from the same x^0\hat{x}_0 mixed with its own noise. The members x^0(b)\hat{x}_0^{(b)} finish the remaining T−sT-s steps independently, and are decoded by D\mathcal{D} into x(b)x^{(b)}. One ensemble costs s+N(T−s)s+N(T-s) network evaluations, 5252 for s=4s=4, N=8N=8, T=10T=10, i.e. 5.2×5.2\times a single forecast against 8×8\times for 8 independent samples (ens8). The two design choices appear in Eq. (4) as the shared x^0\hat{x}_0, setting the common evolution, and the weight τs\tau_s, setting the fresh noise level.

Trunk evolution guidance. Since every member keeps the same x^0\hat{x}_0, the whole ensemble inherits the evolution that the trunk has settled by step ss. Without a future condition, this evolution comes from the noise drawn at the start of the trunk, and all members inherit one random choice. Under the predicted scenario cc of § 3.1, the trunk is guided by the predicted evolution, and the members spend their fresh noise on local details. § 5.1 ablates this effect.

Optimal branching timing. With the evolution guided by cc, the step ss sets the balance in Eq. (4) between the shared x^0\hat{x}_0, weighted by 1−τs1-\tau_s, and the fresh noise, weighted by τs\tau_s. If the members branch after only a few shared steps (ss small, τs→1\tau_s \to1), each member restarts from nearly pure noise and stays close to an independent sample. Skill is then preserved, but the trunk saves little compute. If they branch after many shared steps (ss close to TT, τs→0\tau_s \to0), x^0\hat{x}_0 has already fixed local details that the few remaining steps cannot change. The members then collapse toward copies of one forecast, and the vote has little to act on. Branching too early thus saves little compute, whereas branching too late costs both diversity and skill, so skill stays on a plateau for early-to-mid ss and drops sharply once local details freeze (App. § 7).

The best branching step is therefore the latest one at which the macroscopic evolution is settled while local details remain unconstrained. Rather than running full ensembles at every candidate step against ground truth, we locate it without labels from the unresolved posterior variance of each frame along a single sampling trajectory. Restricting the flow-matching Tweedie identity Cov⁡(x0∣zτ)=τ21−τ[I+τJv]\operatorname{Cov}(x_0 \mid z_\tau) = \frac{\tau^2}{1-\tau}[I+\tau J_v], where Jv=∂v/∂zJ_v = \partial v/\partial z [47], to frame ff yields its posterior width:

Uf(τ)=τ21−τ[df+τdiv⁡fv],(5)U_f(\tau) = \frac{\tau^2}{1-\tau}\left[d_f + \tau\operatorname{div}_f v\right], \tag*{(5)}

where div⁡fv\operatorname{div}_f v is the trace of the frame-ff block of JvJ_v. Since Uf(τs)U_f(\tau_s) quantifies the remaining stochastic exploration capacity, branching beyond the peak of UfU_f rapidly reduces member diversity. We therefore branch where the mean width of the late frames (f=8f = 8–16) peaks:

s∗=arg max⁡s U‾late(τs).(6)s^{*} = \underset{s}{\operatorname{arg\,max}}\ \overline{U}_{\mathrm{late}}(\tau_s). \tag*{(6)}

Posterior-width probe. We estimate div⁡f\operatorname{div} f in Eq. (5) with a Hutchinson trace estimator using a Rademacher probe on frame ff, which costs one Jacobian–vector product (JVP). Locating s∗s^{*} thus takes only F=16F = 16 JVPs per step along a single trajectory, with neither ground truth nor branched members (App. § 8). On the Swed­ish held-out set (Sec. 4.1), this posterior width peaks at s∗=4s^{*} = 4 and falls to 0.43 of its peak by s=8s = 8 (Figure 4a). A supervised sweep over all ss (Figure 4b) confirms the choice: CSI stays close to ens8 for s∈[2,6]s \in[2,6] and drops from s=7s = 7, and s∗=4s^{*} = 4 captures the full 8-member gain over all frames (94% on frames 8–16) at 5.2× compute. The cheaper s=6s = 6 (3.8×) suits compute-limited deployments (App. § 10.§ 1).

Figure 4. Selection of the branching step ss on the held-out set. (a) Peak-normalized posterior width of early and late frames. Below the axis, τs\tau_s is the noise level of the sampler at step ss (1 is pure noise, 0 is data). (b) Highest-threshold CSI when branching at step ss (N=8N = 8, λ=1\lambda= 1, k=2k = 2), with the cost below the axis and the plateau shaded. Dashed lines: ens8 (8.0×). Vertical line: s∗=4s^* = 4. Numbers in App. Tab. 7.

Self-Anchored Support-Consensus Decoding

Averaging the NN members suppresses variance but smooths out their spatially misaligned heavy echoes, reintroducing the blur of deterministic models (§ 5.3). To avoid this averaging blur, we propose self-anchored support-consensus decoding, which uses no pixel averaging. It takes one complete TT-step member as the anchor, at no extra sampling cost, and adds or removes heavy precipitation by pixel-wise support voting across all members. Because all members share the scenario condition cc (§ 3.1) and branch from a shared trunk (§ 3.2), their convective cores are already well aligned in space and differ primarily in local extent and intensity (Figure 3). Resolving such local disagreements via consensus voting harnesses the diversity benefits of the ensemble while preserving the peak intensity and sharp boundaries of extreme echoes.

Let a=x(0)a = x^{(0)} denote the anchor. For a pixel pp of any forecast frame and a given heavy-precipitation intensity threshold thr, let P^(p)=1N#{b:x(b)(p)≥thr}\hat{P}(p) = \frac{1}{N}\#\{b : x^{(b)}(p) \ge\mathrm{thr}\} denote the fraction of members at or above thr. The delivered forecast is

y(p)={max⁡bx(b)(p),P^(p)≥k/N, a(p)<thr,thr−δ,P^(p)<k/N, a(p)≥thr,a(p),otherwise,(7)y(p) = \begin{cases} \max_b x^{(b)}(p), & \hat{P}(p) \ge k/N,\ a(p) < \mathrm{thr},\\ \mathrm{thr} - \delta, & \hat{P}(p) < k/N,\ a(p) \ge\mathrm{thr},\\ a(p), & \text{otherwise}, \end{cases} \tag*{(7)}

where kk is the required number of votes. The first case fills heavy rain that the anchor misses and at least kk members predict. The second removes an anchor false alarm supported by fewer than kk members by setting the pixel to $\mathrm{thr} - \delta (δ=10−4\delta= 10^{-4}), removing the unsupported extreme-tier value with a minimal change to the anchor. All other pixels keep the anchor’s values, so no pixel is averaged. We set thr\mathit{thr} to the highest intensity tier of each dataset and k=2k = 2, the smallest vote that forms a consensus (Table 6). The vote edits only pixels where the anchor and the ensemble disagree about this tier (effect on other tiers and metrics in § 5.3), and neither branching nor decoding changes the trained generator.

Number of votes kk (top, N=8N = 8) and ensemble size NN (bottom, k=2k = 2) on the held-out set. CSI is at ≥6.3\ge6.3 mm/h and FB is the frequency bias (1 is unbiased). N=1N = 1 is the anchor. Full sweeps are in App. Tabs. 8 and 9.

kk12345678
CSI↑\uparrow.1760.1740.1518.1266.1036.0821.0608.0420
FB→\to11.921.050.700.500.370.270.19
NN1246816
CSI↑\uparrow.1239.0893.1399.1649.1740.1867
FB→\to10.660.330.650.881.051.53

Table 6.

Experiments

Experimental Setup

Datasets. We evaluate ForeSay on six datasets. The Swedish precipitation dataset and MRMS follow the evaluation setup of LangPrecip [22] (256 × 256, 5-minute interval), forecasting 16 future frames (80 min) from 4 historical observations on 1000 test sequences each. SEVIR (8096 test sequences), MeteoNet (1310), Shanghai (526), and CIKM (4000) serve as four public radar benchmarks (128 × 128 protocol), on which models predict 20 future frames from 5 inputs (10 frames for CIKM, App. § 11). All inference hyperparameters (s∗=4s^* = 4, λ=1\lambda= 1, k=2k = 2, N=8N = 8) are selected once on a 500-sample held-out set of the Swedish dataset, disjoint from the training and test splits, and kept fixed on all six benchmarks.

Metrics. On the Swedish precipitation dataset and MRMS we follow LangPrecip [22] and SSDLDM-ISI [21] and report CSI at a weak and a heavy threshold (0.06/6.3 mm/h for Sweden, 1/8 mm/h for MRMS), FSS, CRPS, SSIM and LPIPS. On the four radar benchmarks we follow the protocol of DiffCast [49], AlphaPre [20] and PercpCast [8] with CSI-M, per-threshold CSI, HSS, SSIM and LPIPS.

Results

Quantitative evaluation and trade-off analysis. On the four radar benchmarks (Table 1), ForeSay ranks first on all three high-threshold CSIs and HSS. At the highest tier it outperforms the strongest baseline PDRF by +21.6% (SEVIR), +10.9% (MeteoNet), +11.7% (Shanghai) and +12.8% (CIKM), with the best multi-scale CSI-M (App. Tab. 11). On the Swedish and MRMS benchmarks (Table 2), heavy-rain CSI exceeds the text baseline LangPrecip by +84% (0.074 → 0.136) and +16% (0.031 → 0.036), and ours stays ahead at every lead time (Figure 5). Even our single-sample anchor (1×) matches or beats the strongest baseline on all datasets but MRMS (Table 5).

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. Figure 8, 11 and 13–19.

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.

Effect of support-consensus decoding. Anchor (1.0×) vs. decoded ensemble (5.2×), paired on the same anchor. Base.: the strongest baseline (Table 1 and Table 2). Δ\Delta: 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.

DatasetBase.AnchorOursΔ\Delta (%)CSI2nd_{2\mathrm{nd}}SSIMLPIPSCRPS
Swedish.074.0984.1357+37.90.00.0+0.1−6.8
MRMS.031.0278.0360+29.50.00.0−0.2−0.3
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+0.40.0+0.1−0.1
CIKM.1720.1744.1941+11.3+0.9−0.1+0.1−0.3

Table 5.

These results also expose the double-penalty trade-off against metrics that favor blur (App. § 14). Pixel-wise metrics (CRPS, MAE) penalize a displaced echo twice, as a false alarm and a miss, and the neighborhood metric FSS also tends to favor smooth rain fields, so smooth baselines (e.g. FlowCast, SDIR) score well while omitting heavy echoes (CSI down to 0.009). When displacement is negligible (≤ 10 min), ForeSay has the lowest CRPS of all methods (App. Fig. 8). Over the full horizon, consensus decoding raises the Swedish CRPS by 6.8% (at most 0.3% on the other datasets, App. Tabs. 13 and 14) in exchange for a +37.9% gain in heavy-rain CSI.

Qualitative results. In Figure 6, at +70 and +80 min, PySteps and DTCA smooth the heavy echoes inside the dashed box, DiffCast nearly loses them, DGMR breaks them into speckles and LangPrecip weakens the core above 35 dBZ (≈ 6.3 mm/h). Guided by c^\hat{c} (bottom), which describes the diagonal band and its rapid upward motion, our anchor keeps part of the heavy core, and decoding brings its extent and intensity closer to the ground truth without blurring. App. § 14 shows more examples on all six datasets.

Case study of a precipitation event with radar composite and forecast comparison

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^\hat{c} for this case is shown at the bottom.

Ablation Studies

Conditioning Components

We first evaluate the role of predicted scenario text in guiding macroscopic storm evolution (Table 3). Under the unbranched setting, history text raises the heavy-rain CSI of radar alone from 0.0895 to 0.0949, and the predicted future scenario advances it to 0.0984. With branching and consensus decoding, the two conditions diverge: history text gains almost nothing over radar alone (0.1275 vs. 0.1273), whereas the future scenario reaches 0.1357 (+6.6%). While history text merely summarizes past observations, the predicted scenario guides the trunk toward the true evolution (Table 4: 84.2%–95.1% on motion direction, 73.6%–90.8% on intensity and structure changes, App. § 9). Members thus share the storm evolution and differ mainly locally, enabling consensus voting to recover heavy echoes (§ 3.3).

Conditionw/o branching (1×) CSI0.06_{0.06} ↑w/ branching (5.2×) CSI6.3_{6.3} ↑
CSI0.06_{0.06} ↑CSI6.3_{6.3} ↑SSIM↑FSS↑CSI6.3_{6.3} ↑
Radar only0.5780.08950.7360.9100.1273
+ history text0.5840.09490.7430.9140.1275
+ future text0.5930.09840.7470.9150.1357

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 directionIntensity and structure
MainL/RU/D8-wayIntens.DecaySplitMerge
95.190.990.884.276.773.675.490.8

Table 4. Agreement (%, ↑) of the predicted c^\hat{c} with the ground truth. Main: primary axis; L/R, U/D: left/right, up/down; 8-way: compass direction.

Branched Sampling and Noise Strength

The noise strength λ\lambda in Eq. (3) balances member diversity and forecast skill (Figure 7a, Figure 7b). At λ=0\lambda= 0, members collapse to the anchor. As λ\lambda increases, support IoU drops from 1.000 to 0.166 and paired CSI gains rise monotonically. At λ=1\lambda= 1, the ensemble matches ens8 in both skill (+0.0501 vs. +0.0503) and diversity (IoU 0.166 vs. 0.165). Thus, late-stage branching retains skill and diversity while preserving macroscopic evolution (Figure 3).

Figure 7. Branching and aggregation ablations. (a) CSI gain over the anchor and (b) member IoU at frame 12 versus the noise strength λ\lambda (held-out set), with ens8 dashed. (c) Heavy-rain CSI per lead time for different aggregations of the same N=8N = 8 members (test set, FB not comparable to Table 6). All-frame CSI (FB): ours 0.1357 (1.18), PMM 0.1165 (0.73), consensus k=4k = 4 0.1002 (0.62), anchor 0.0984 (0.73), ensemble mean 0.0852 (0.50).

Support-Consensus Decoding and Aggregation

Aggregating ensemble members without blurring convective extremes is critical (Figure 7c). Pixel averaging drops heavy-rain CSI to 0.0852 (FB 0.50), below the anchor’s 0.0984. While PMM restores amplitudes to 0.1165, it keeps the mean pattern. Support-consensus decoding avoids the mean field entirely, reaching 0.1357 (+16.5% over PMM) and leading at every lead time. Because members differ mainly in heavy cores (Figure 3), gains concentrate at the highest tier (Table 5, +5.3% to +37.9%), while second-tier CSI, SSIM, LPIPS, and CRPS mostly stay within ±1% (exceptions: SEVIR second-tier CSI +4.1% and CRPS +1.9%, Swedish CRPS −6.8%), as voting only edits heavy-rain regions at thr. Across the four radar benchmarks, 19 of 22 CSIs and all four HSS improve (App. § 13).

Vote count and ensemble size. In Table 6 (top), union voting k=1k = 1 gives the highest raw CSI but doubles the rain area with an excessive FB of 1.92. In contrast, k=2k = 2, the minimal consensus requirement, preserves nearly the same CSI at a near-unbiased FB of 1.05. With k=2k = 2 (bottom), N=2N = 2 reduces to the intersection of two members and under-forecasts (FB 0.33). Scaling to N=16N = 16 adds 0.0127 CSI at 1.9× compute cost, mainly inflating rain area (FB 1.53). We therefore use k=2k = 2 and N=8N = 8.

Conclusion

ForeSay separates the two scales of nowcasting uncertainty: the macroscopic evolution is forecast as scenario text, and only the local details are sampled. Members no longer need their own early sampling steps, so they share a trunk and match eight independent samples in skill and diversity at 5.2× instead of 8× compute. Their remaining differences are local, so pixel-wise voting can replace pixel averaging and keep the heavy echoes that averaging smooths out.

The branching step is read from the generator’s own posterior width. Across six datasets, ForeSay achieves the best heavy-precipitation CSI, 11%–84% above the strongest baseline, and the ablations show that text helps when it predicts the future rather than summarizes the past. These results suggest that a nowcasting ensemble can forecast the storm system once and spend its randomness on where heavy convective cores form.

Limitations and future work. Conditioning all members on a single predicted scenario leaves macroscopic uncertainty unmodeled, so any error in the text propagates across the entire ensemble—especially when radar history cannot resolve whether a storm system will intensify, decay, or split. Sampling several scenarios in such cases, each with its own branches, would capture uncertainty at both scales. Separately, because consensus voting acts pixel-wise, it keeps heavy cores sharp at their forecast locations and thus raises CRPS when those positions drift.

Details of the Branching Operator

Algorithm 1 Branched RF sampling. Lines in teal depart from standard RF sampling.

Require: hh, cc, branching step ss, members NN, strength λ\lambda, RF grid 1=τ0>⋯>τT=01=\tau_0>\cdots>\tau_T=0

Ensure: members x(0),…,x(N−1)x^{(0)},\ldots,x^{(N-1)}, anchor x(0)x^{(0)}

  1. z←ϵ0∼N(0,I)z\leftarrow\epsilon_0\sim\mathcal{N}(0,I)

  2. for i=0…s−1i=0\ldots s-1 do ▷ shared trunk

  3. v←vθ(z,τi;h,c)v\leftarrow v_\theta(z,\tau_i;h,c); z←z+v⋅(τi−τi+1)z\leftarrow z+v\cdot(\tau_i-\tau_{i+1})

  4. end for

  5. x^0←z+τsv\hat{x}_0\leftarrow z+\tau_s v; ϵ^←z+(τs−1)v\hat{\epsilon}\leftarrow z+(\tau_s-1)v ▷ preview

  6. for b=0…N−1b=0\ldots N-1 do ▷ branch, x(0)x^{(0)} is the anchor

  7. ϵb∼N(0,I)\epsilon_b\sim\mathcal{N}(0,I) ▷ re-noise at τs\tau_s

  8. zb←(1−τs)x^0+τs(1−λ2 ϵ^+λϵb)z_b\leftarrow(1-\tau_s)\hat{x}_0+\tau_s\left(\sqrt{1-\lambda^2}\,\hat{\epsilon}+\lambda\epsilon_b\right)

  9. end for

  10. for b=0…N−1b=0\ldots N-1 do

  11. for i=s…T−1i=s\ldots T-1 do

  12. v←vθ(zb,τi;h,c)v\leftarrow v_\theta(z_b,\tau_i;h,c); zb←zb+v⋅(τi−τi+1)z_b\leftarrow z_b+v\cdot(\tau_i-\tau_{i+1})

  13. end for

  14. x(b)←D(zb)x^{(b)}\leftarrow D(z_b)

  15. end for

Algorithm 2 Support-consensus decoding

Require: members x(0…N−1)x^{(0\ldots N-1)}, anchor a=x(0)a=x^{(0)}, threshold thr\mathrm{thr}, vote kk

Ensure: delivered field yy

  1. y←ay\leftarrow a

  2. P^(p)←1N#{b:x(b)(p)≥thr}\hat{P}(p)\leftarrow\frac{1}{N}\#\{b:x^{(b)}(p)\ge\mathrm{thr}\}

  3. ADD←{p:P^(p)≥k/N∧a(p)<thr}\mathrm{ADD}\leftarrow\{p:\hat{P}(p)\ge k/N\land a(p)<\mathrm{thr}\}

  4. REM←{p:P^(p)<k/N∧a(p)≥thr}\mathrm{REM}\leftarrow\{p:\hat{P}(p)<k/N\land a(p)\ge\mathrm{thr}\}

  5. y[ADD]←max⁡bx(b)[ADD]y[\mathrm{ADD}]\leftarrow\max_b x^{(b)}[\mathrm{ADD}]; y[REM]←thr−10−4y[\mathrm{REM}]\leftarrow\mathrm{thr}-10^{-4}

Algorithms 1 and 2 give the full procedures of branched sampling (Eqs. (3)–(4) of the main text) and support-consensus decoding (Eq. (7)).

With τ=0\tau=0 at the data end and τ=1\tau=1 at the noise end, Rectified Flow uses the path zτ=(1−τ)x0+τϵz_\tau=(1-\tau)x_0+\tau\epsilon, and the network predicts the velocity vθ≈x0−ϵv_\theta\approx x_0-\epsilon. At a branching step τs\tau_s the current state can therefore be written as

z≈(1−τs)x^0+τsϵ^z\approx(1-\tau_s)\hat{x}_0+\tau_s\hat{\epsilon}

where x^0=z+τsvθ\hat{x}_0=z+\tau_s v_\theta is the estimated data endpoint and ϵ^=z+(τs−1)vθ\hat{\epsilon}=z+(\tau_s-1)v_\theta the estimated noise endpoint. The branching step in Alg. 1 keeps x^0\hat{x}_0 and re-noises it to the same level τs\tau_s,

zb≈(1−τs)x^0+τs(1−λ2 ϵ^+λϵb).z_b\approx(1-\tau_s)\hat{x}_0+\tau_s\left(\sqrt{1-\lambda^2}\,\hat{\epsilon}+\lambda\epsilon_b\right).

The step τs\tau_s sets the weight between the data and noise ends, and λ\lambda sets how much of the old noise is replaced. At λ=1\lambda=1 the update reduces to zb=z+τs(ϵb−ϵ^)z_b=z+\tau_s(\epsilon_b-\hat{\epsilon}), so that

zb≈((1−τs)x^0+τsϵ^)+τs(ϵb−ϵ^)=(1−τs)x^0+τsϵb.z_b\approx\left((1-\tau_s)\hat{x}_0+\tau_s\hat{\epsilon}\right)+\tau_s(\epsilon_b-\hat{\epsilon})=(1-\tau_s)\hat{x}_0+\tau_s\epsilon_b.

Branching thus removes the old noise component and keeps the data component (1−τs)x^0(1-\tau_s)\hat{x}_0. To match the base sampler, our experiments compute x^0\hat{x}_0 and ϵ^\hat{\epsilon} from the velocity of the last trunk step. An exact variant spends one more forward pass at the branching state, which changes the estimates of x^0\hat{x}_0 and ϵ^\hat{\epsilon} but not the operator. The members share only this endpoint. After branching, each of them evolves independently under the nonlinear velocity field.

Why a useful branching step lies inside the trajectory. The same decomposition predicts where branching can work, independently of the dataset. In the branched state

zb≈(1−τs)x^0+τsϵbz_b\approx(1-\tau_s)\hat{x}_0+\tau_s\epsilon_b

the weight 1−τs1-\tau_s sets how strongly the shared endpoint binds the members to one macroscopic evolution, and τs\tau_s sets how much independent randomness each member keeps.

  • Branching too early (τs→1\tau_s \to1). The shared x^0\hat{x}_0 carries almost no weight and every member restarts from nearly pure noise. The trunk saves little compute, and the ensemble approaches independent sampling.

  • Branching too late (τs→0\tau_s \to0). The fresh noise ϵb\epsilon_b carries almost no weight, all members start close to the same x^0\hat{x}_0, and only a few steps remain for them to diverge. The ensemble collapses toward copies of one forecast and leaves the vote little to work with.

  • In between, x^0\hat{x}_0 is strong enough to hold all members to one evolution, and ϵb\epsilon_b is strong enough to make them disagree on local detail.

A useful branching step therefore lies inside the trajectory. The two failures are asymmetric. Branching too early leaves skill intact and only loses the saving, whereas branching too late hurts both diversity and skill. As the branching step moves later, skill should stay flat and then drop. This argument shows that such a step exists but not where it is. Its location is given by the posterior-width criterion of Eq. (6) in the main text (Appendix § 8) and checked by the sweep (Appendix § 10.§ 1).

Anchor cost. All experiments are self-anchored. The anchor x(0)x^{(0)} is the first branch of the same run. Like every member it receives fresh noise at step ss, and on its own it is a complete TT-step sample (1.0×). An external anchor would cost one more full sampling (+TT steps).

Posterior-Width Probe

Why a probe. Appendix § 7 shows that skill is flat for early branching and drops for late branching, but it does not say where the drop begins. A sweep over ss needs ground truth and a full ensemble at every candidate step, and it locates the drop without explaining it. The posterior width Uf(τ)U_f(\tau) of Eq. (5) measures how much of frame ff the current state still leaves undetermined, and hence how much members branched from that state can still differ. It is computed from the velocity field along one sampling trajectory, without ground truth and without branched members. The criterion of Eq. (6) branches at the last step before the late-frame width starts to contract.

Frame-structured Hutchinson probe with JVPs. The full Jacobian covers all 16×3276816 \times32768 latent coordinates, and even one per-frame block has 32768×3276832768 \times32768 entries. To avoid forming either, we use the Hutchinson estimator E[n⊤Jn]=tr⁡(J)\mathbb{E}[n^\top Jn] = \operatorname{tr}(J) and restrict each probe vector to the coordinates of one frame, so that the estimate counts that frame only:

Frame-structured Hutchinson probe

Figure 1. Listing 1. Frame-structured Hutchinson probe.

signs = randint(0, 2, z.shape) * 2 - 1 # Rademacher +-1, drawn once for z
for fr in range(16): # loop over frames
    n = zeros_like(z); n[:, :, fr] = signs[:, :, fr] # probe on frame fr
    _, jv = jvp(lambda zz: cfg_velocity(zz, t), (z,), (n,)) # jv = J n
    divs[pi, fr, dr] = float((jv * n).sum()) # n^T J n, estimate of tr(J_ff)

The estimator follows three choices. First, each probe vector is nonzero only on the coordinates of one frame, so one Jacobian–vector product (JVP) gives the divergence of one frame. A 16-frame forecast thus needs 16 JVPs per probe sample, each costing about one forward pass. Second, the Rademacher signs signs are drawn once for the whole latent and sliced per frame, so all frames of one sample share the same signs. These common random numbers reduce the noise in the difference between early and late frames. Third, the probe is evaluated only at the grid points of the sampler, and between them zz follows the standard update z←z+vΔτz \leftarrow z + v \Delta\tau. The divergence is therefore measured at states that the sampler actually visits.

Figure 4a of the main text compares the normalized widths of early and late frames across sampling steps. The late-frame width peaks at step 4, where the early frames are also close to their peak, so Eq. (6) gives s∗=4s^* = 4. After this step the late-frame width falls monotonically, to about 0.97, 0.85, 0.67 and 0.43 of its peak at s=5s=5–8. Over the same range the member IoU rises (0.190, 0.220, 0.228, 0.297) and the CSI drops from s=7s=7 on (Table 7).

SettingStepsCostCSI↑ (all 16 frames) (of 8×)CSI↑ (f=8f = 8–16) (of 8×)Member IoU (f=12f = 12)
s=2s = 2, N=8N = 8666.6×0.1743 (97%)0.1366 (91%)0.151
s=3s = 3, N=8N = 8595.9×0.1753 (85%)0.1379 (87%)0.174
s=4s = 4, N=8N = 8 (default)525.2×0.1740 (100%)0.1355 (94%)0.166
s=5s = 5, N=8N = 8454.5×0.1699 (102%)0.1339 (103%)0.190
s=6s = 6, N=8N = 8383.8×0.1688 (93%)0.1275 (82%)0.220
s=7s = 7, N=8N = 8313.1×0.1632 (79%)0.1244 (73%)0.228
s=8s = 8, N=8N = 8242.4×0.1559 (66%)0.1145 (70%)0.297
ens8 (independent sampling)808.0×0.1757 (100%)0.1412 (100%)0.165

Table 7. Sweep of the branching step ss on the Swedish precipitation held-out set (N=8N = 8, T=10T = 10, λ=1\lambda= 1, k=2k = 2). Steps is the number of network evaluations per forecast, s+N(T−s)s + N(T - s), and cost is relative to a single sample (10 steps). CSI is at the highest tier (≥6.3\ge6.3 mm/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 NN members in the 12th forecast frame, and lower means more diverse members.

Accuracy of the Predicted Scenario

We compare the scenario c^\hat{c} predicted from the history alone with labels computed from the ground-truth future radar, without involving the teacher. From c^\hat{c} we extract the motion direction and the intensity and structural changes. The reference motion direction is obtained by optical flow on the true future frames, and the reference intensity and structural changes are computed from the true future radar fields. We report the fraction of the 1000 Swedish precipitation test cases on which the two agree (Table 4 of the main text). The predicted scenario agrees with the ground truth on the main motion direction in 95.1% of the cases and on the 8-way direction in 84.2%. Agreement on intensity and structural change ranges from 73.6% to 90.8%.

Hyperparameter Sweeps

All sweeps use the 500-sample Swedish precipitation held-out set.

Branching Step ss

The posterior width selects s∗s^* without labels. To check the choice on the reported metric, we generate NN branches at every candidate ss and measure two quantities. The heavy-rain CSI tells whether branching still brings a forecast gain, and the IoU of the precipitation supports of all NN members tells whether the members remain diverse (lower is more diverse).

The results (Figure 4b of the main text and Table 7) show a plateau followed by a drop. For s∈[2,4]s \in[2, 4] the all-frame CSI is 0.1740–0.1753, within 0.002 of the independent ensemble (ens8, 0.1757). At s=5s = 5 and 6 it is only 0.0041 and 0.0052 below s=4s = 4, still recovering 102% and 93% of the ens8 gain. From s=7s = 7 on the CSI drops clearly (0.1632 and 0.1559 at s=7s = 7 and 8, recovering 79% and 66%), and the late frames (f=8f = 8–16) drop more. The member IoU rises from 0.166 at s=4s = 4 to 0.297 at s=8s = 8, so later branching gives more similar members and leaves the vote less to work with.

This shape matches Appendix § 7. Early branching leaves skill intact and only costs compute, as s=2s = 2 and 3 reach the CSI of s=4s = 4 at 6.6× and 5.9×. Late branching hurts both diversity and skill.

Why s∗=4s^* = 4. Within the plateau later branching is cheaper, and s=6s = 6 still recovers 93% of the ens8 gain at 3.8×. We keep s=4s = 4 for two reasons.

  1. Margin to the drop. s=6s = 6 lies next to the drop that starts at s=7s = 7, and its late frames (f=8f = 8–16) recover only 82% of the ens8 gain against 94% at s=4s = 4.

  2. Unchanged diversity. The member IoU at s=4s = 4 (0.166) equals that of independent sampling (0.165), while at s=5s = 5 and s=6s = 6 it has already risen to 0.190 and 0.220, as the late-frame width has started to contract (0.97 and 0.85 of its peak).

Deployments with a tighter compute budget may use s=6s = 6 (3.8×) at a slightly lower gain at long lead times.

Number of Votes kk

kk is the number of members that must exceed the threshold for a pixel to count as wet, so a smaller kk gives a larger support. We sweep k=1k = 1–8 on the held-out set (Table 8) and do not simply pick the kk with the highest CSI. k=1k = 1 reaches the highest CSI, but it takes the union of all member supports and forms no consensus. Its FB of 1.92 means that the forecast rain area is about twice the observed one, and its gain comes from this inflated area. k=2k = 2 is the smallest vote in which two members must agree. Its CSI is close to that of k=1k = 1 (0.1740 vs. 0.1760 over all 16 frames), and its area overestimate almost disappears (FB 1.05). From k=3k = 3 on the FB falls below 1 and the area is underestimated, and the CSI falls with kk, to 0.1518 at k=3k = 3 and 0.0420 at k=8k = 8, where all members must agree. A strict vote thus deletes much real precipitation. We therefore use k=2k = 2.

Table 8. Sweep of the number of votes kk (Swedish precipitation held-out set, 500 samples, s=4s = 4, N=8N = 8, λ=1\lambda= 1). CSI is at the highest tier (≥6.3\ge6.3 mm/h). FB is the frequency bias, the number of forecast pixels above the threshold divided by the observed number, so 1 is unbiased, >1> 1 overestimates and <1< 1 underestimates the rain area. k=1k = 1 takes the union of all member supports, and k=2k = 2 is the smallest vote that forms a consensus.

kkCSI (all 16 frames)↑\uparrowCSI (f=8f = 8–16)↑\uparrowFB (closer to 1 is better)
10.17600.14491.92
20.17400.13551.05
30.15180.10990.70
40.12660.08620.50
50.10360.06520.37
60.08210.04550.27
70.06080.02870.19
80.04200.01720.12

Table 8. Sweep of the number of votes kk (Swedish precipitation held-out set, 500 samples, s=4s = 4, N=8N = 8, λ=1\lambda= 1). CSI is at the highest tier (≥6.3\ge6.3 mm/h). FB is the frequency bias, the number of forecast pixels above the threshold divided by the observed number, so 1 is unbiased, >1> 1 overestimates and <1< 1 underestimates the rain area. k=1k = 1 takes the union of all member supports, and k=2k = 2 is the smallest vote that forms a consensus.

Number of Members NN

Table 9 varies the number of members on the held-out set with the vote fixed at k=2k = 2. N=1N = 1 is the anchor alone without voting. At N=2N = 2 the vote requires both members to agree, which removes real precipitation and drops the CSI below the anchor (FB 0.33). From N=4N = 4 on, the CSI rises with NN and the frequency bias moves from an underestimate to an overestimate. N=8N = 8 gives the bias closest to 1 (1.05). N=16N = 16 adds 0.013 CSI at nearly twice the cost, and its bias of 1.53 shows that the extra gain comes largely from an inflated rain area, the same effect as a looser vote (Appendix § 10.§ 2). We therefore use N=8N = 8.

Table 9. Sweep of the number of members NN with the vote fixed at k=2k = 2 (Swedish precipitation held-out set, 500 samples, s=4s = 4, λ=1\lambda= 1). CSI is at the highest tier (≥6.3\ge6.3 mm/h). Cost is s+N(T−s)s + N(T - s) network evaluations relative to a single sample (10 steps). FB is the frequency bias, and 1 is unbiased.

NNCostCSI (f=1f = 1–4)↑\uparrowCSI (f=8f = 8–16)↑\uparrowCSI (all 16 frames)↑\uparrowFB (closer to 1 is better)
1 (anchor, no vote)1.0×0.19860.08920.12390.66
21.6×0.16410.05530.08930.33
42.8×0.21860.10090.13990.65
64.0×0.24290.12450.16490.88
85.2×0.24810.13550.17401.05
1610.0×0.25470.15380.18671.53

Table 9. Sweep of the number of members NN with the vote fixed at k=2k = 2 (Swedish precipitation held-out set, 500 samples, s=4s = 4, λ=1\lambda= 1). CSI is at the highest tier (≥6.3\ge6.3 mm/h). Cost is s+N(T−s)s + N(T - s) network evaluations relative to a single sample (10 steps). FB is the frequency bias, and 1 is unbiased.

Comparison on Radar Extrapolation Benchmarks

Beyond the Swedish and MRMS datasets, we compare with existing methods on four public radar extrapolation benchmarks, SEVIR, MeteoNet, Shanghai and CIKM. CSI-M averages over thresholds and hides where a gain occurs. The main text (Table 1) therefore reports CSI at the three highest thresholds together with HSS, SSIM and LPIPS, and Table 11 lists CSI-M and CSI at every threshold of each benchmark. As in § 4.1, we follow the protocol of DiffCast [49], AlphaPre [20] and PercpCast [8] and evaluate all metrics at 128 × 128 (Appendix § 12). Best results are in bold and second best are underlined. Table 10 gives the test split, sample count and forecast setting of each benchmark.

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.

DatasetEvaluation splitSamplesCondition frames → forecast framesIntensity thresholds
SEVIR (VIL)official test80965 → 2016/74/133/160/181/219
MeteoNettest13105 → 2012/18/24/32
Shanghaiofficial test5265 → 2020/30/35/40
CIKMtest40005 → 1020/30/35/40

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.

MethodSEVIR (8096) CSI-M↑CSI-16↑CSI-74↑CSI-133↑CSI-160↑CSI-181↑CSI-219↑MeteoNet (1310) CSI-M↑CSI-12↑CSI-18↑CSI-24↑CSI-32↑Shanghai (526) CSI-M↑CSI-20↑CSI-30↑CSI-35↑CSI-40↑CIKM (4000) CSI-M↑CSI-20↑CSI-30↑CSI-35↑CSI-40↑
MIMO-VP [29]0.29310.66430.55620.25090.14920.11020.02800.28780.47010.38380.24690.05050.31080.49540.37220.25860.11710.24920.59320.19330.13510.0752
SwinLSTM [41]0.31100.67610.60150.27380.15850.11880.03730.37780.51610.45580.36310.17620.38390.53470.44090.34600.21410.30020.62170.29900.20020.0799
DiffCast [49]0.30830.65860.54100.25780.17400.14350.07500.36930.50840.44270.34960.17670.36550.50100.40720.32540.22860.30670.60530.27920.21610.1261
AlphaPre [20]0.32490.69840.60070.28830.17540.13320.05350.36270.49260.43150.34610.18070.39720.54970.45140.35910.22870.28530.59280.26040.18100.1070
PercpCast [8]0.34920.68560.58440.31010.22030.19290.10220.40390.52790.47370.39330.22070.39950.53160.44590.36280.25770.31630.61080.27660.22230.1553
exPreCast [39]0.34010.66360.60040.32660.21330.16490.07160.39730.52650.47040.38330.20900.39010.54120.43990.34710.23230.29270.62210.27290.18730.0886
FlowCast [34]0.35060.70830.62300.32760.21490.16610.06380.38980.53670.47050.37150.18030.42720.57250.48150.39270.26220.31620.61770.29540.21930.1324
SDIR [52]0.36060.71080.62000.33630.22010.18280.09340.42700.54720.50370.42200.23490.43930.58170.49190.40400.27970.32880.64320.29880.22100.1524
PDRF [26]0.36230.68180.59120.32890.24080.21100.11990.43850.55250.50530.42850.26790.44090.56810.48560.40660.30320.33290.62770.29380.23810.1720
Ours†0.37060.68150.59740.33410.24770.21730.14580.44700.54820.50580.43690.29710.47140.59330.51660.43690.33870.35070.63170.31720.25980.1941

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.

†Ours is the decoded ensemble with thr at the highest tier of each benchmark (SEVIR: VIL ≥219\ge219, MeteoNet: ≥32\ge32 dBZ, Shanghai and CIKM: ≥40\ge40 dBZ).

On the Evaluation Resolution

We evaluate at two resolutions, the native 256×256256 \times256 output and the 128×128128 \times128 protocol of prior radar benchmarks.

Use of the two resolutions. Our model outputs 256×256256 \times256 fields on the four radar benchmarks. All metrics in Table 1 of the main text and Table 11 (CSI, HSS, SSIM, LPIPS) are evaluated at 128×128128 \times128, the resolution at which DiffCast [49], AlphaPre [20], PercpCast [8] and others report results on these benchmarks. Our forecasts and the ground truth are mapped to [0,1][0,1], bilinearly downsampled to 128×128128 \times128 and then evaluated. The baselines output 128×128128 \times128 fields and are not rescaled. On CIKM, whose native grid is 101×101101 \times101, all methods including ours are evaluated on 101×101101 \times101 fields zero-padded to 128×128128 \times128. The CSI and HSS in Table 14 follow the same 128×128128 \times128 protocol. Its MAE, RMSE, CRPS, SSIM and LPIPS are computed on our native 256×256256 \times256 output, with MAE and RMSE in the original units of each dataset. That table contains no baseline, so these metrics need not follow the 128×128128 \times128 protocol.

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———
6th(16) 0.6774 → 0.6815———
CSI-M (multi-threshold avg.)0.3604 → 0.37060.4407 → 0.44700.4668 → 0.47140.3451 → 0.3507
Skill score (128 × 128 protocol)
HSS↑0.4664 → 0.48000.5779 → 0.58600.6041 → 0.60980.4468 → 0.4546
Pixel-level and perceptual metrics (256 × 256)
MAE↓9.158 → 8.9761.183 → 1.1771.6220 → 1.62304.7930 → 4.8038
RMSE↓18.374 → 18.0143.460 → 3.4665.1052 → 5.11317.3979 → 7.4176
CRPS↓0.0359 → 0.03520.01314 → 0.013070.01802 → 0.018030.05990 → 0.06005
SSIM↑0.7019 → 0.70610.8365 → 0.83850.8346 → 0.83500.7066 → 0.7057
LPIPS↓0.2361 → 0.23460.1528 → 0.15270.1723 → 0.17220.3548 → 0.3544

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 § 12 compares the two resolutions.

Numerical comparison under the two resolutions. Table 12 gives our results under both protocols. The conclusions do not depend on the resolution. On all four datasets decoding improves CSI-M and HSS, and the gain concentrates in the highest tier. The absolute values shift systematically. Compared with 256×256256 \times256, the 128×128128 \times128 protocol gives a higher HSS on all four datasets, a higher CSI-M on all but CIKM and a lower LPIPS, while SSIM rises on MeteoNet and falls on the other three, most notably on CIKM (0.706 →\to 0.675). Values from the two protocols are therefore not interchangeable.

MetricProtocolSEVIRMeteoNetShanghaiCIKM
CSI-M↑256×256256 \times2560.3484 →\to 0.35790.4081 →\to 0.41530.4456 →\to 0.45090.3482 →\to 0.3525
128×128128 \times1280.3604 →\to 0.37060.4407 →\to 0.44700.4668 →\to 0.47140.3451 →\to 0.3507
HSS↑256×256256 \times2560.4531 →\to 0.46590.5454 →\to 0.55500.5835 →\to 0.59010.4268 →\to 0.4329
128×128128 \times1280.4664 →\to 0.48000.5779 →\to 0.58600.6041 →\to 0.60980.4468 →\to 0.4546
SSIM↑256×256256 \times2560.7019 →\to 0.70610.8365 →\to 0.83850.8346 →\to 0.83500.7066 →\to 0.7057
128×128128 \times1280.6926 →\to 0.69780.8439 →\to 0.84510.8218 →\to 0.82220.6757 →\to 0.6754
LPIPS↓256×256256 \times2560.2361 →\to 0.23460.1528 →\to 0.15270.1723 →\to 0.17220.3548 →\to 0.3544
128×128128 \times1280.1671 →\to 0.16700.0993 →\to 0.10000.1300 →\to 0.13020.2703 →\to 0.2691

Table 12. Our results under the two evaluation resolutions (single model →\to ensemble). The ensemble setting is that of Table 14, and the samples and forecast setting are those of Table 10. 256×256256 \times256 is the native output and 128×128128 \times128 is the protocol of prior work.

Anchor versus Decoded Ensemble on All Six Datasets

Table 13 compares the anchor alone with the decoded ensemble on the Swedish precipitation dataset and MRMS, and Table 14 does so on the official test sets of SEVIR, MeteoNet, Shanghai and CIKM. The ensemble uses s=4s = 4, N=8N = 8, λ=1\lambda= 1, 52 network evaluations in total (5.2×), with thr at the highest tier of each dataset. The single-model baseline is the anchor of that ensemble, the field that support-consensus decoding starts from. It costs the 4 trunk steps plus 6 branch steps, 10 steps in total (1.0×). Both share the same anchor field and differ only in whether the kk-of-NN vote corrects it, so the comparison is paired and the anchor’s own sampling noise cancels. Table 14 gives the full results. Since the thresholds differ across datasets, its CSI rows are ordered by intensity tier from highest to lowest.

SwedishMRMS
CSI↑, highest(6.3) 0.0984 →\to 0.1357(8) 0.0278 →\to 0.0360
CSI↑, weak(0.06) 0.5926 →\to 0.5927(1) 0.4346 →\to 0.4347
FSS↑0.9149 →\to 0.91480.7052 →\to 0.7034
SSIM↑0.7467 →\to 0.74660.8985 →\to 0.8983
CRPS↓0.1159 →\to 0.12380.1360 →\to 0.1364
LPIPS↓0.1897 →\to 0.18960.1306 →\to 0.1308

Table 13. Anchor (single model, 1.0×) →\to 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 Table 5 of the main text.

Breadth of the gain. The improvements are spread across intensity tiers. Of the 22 CSI entries over the four datasets (18 tiers and 4 CSI-M values), 19 improve, 1 ties and 2 regress, each by at most 0.0005 (Shanghai at 30 dBZ and CIKM at 20 dBZ). HSS improves on all four datasets. Decoding thus causes no material loss at any tier.

Concentration of the gain. The size of the gain concentrates in the highest tier. Gains in the lower tiers are mostly 0.0003 to 0.0086, while the highest tier gains 0.0170 (Shanghai) to 0.0256 (SEVIR). Since CSI-M is the mean of the per-threshold CSIs, the share of each tier can be read off directly. On MeteoNet, Shanghai and CIKM the highest tier gains 0.017 to 0.022 while every other tier changes by at most 0.0024, so the highest tier accounts for 88% to 92% of the CSI-M increment. SEVIR also improves by 0.004 to 0.009 at all five lower tiers, and its highest tier accounts for 42%.

Pixel-level cost. At the pixel level the anchor and the ensemble differ little and not in a consistent direction. Both deliver one field, so CRPS equals the MAE on fields normalized to [0, 1] and the two always move together. On SEVIR MAE and RMSE improve by about 2%, on MeteoNet MAE improves and RMSE is flat, and on Shanghai and CIKM both worsen slightly. No regression exceeds 0.27% (the largest is RMSE on CIKM, 7.3979 → 7.4176). Over the same experiments the highest-tier CSI improves by 5.3% (Shanghai) to 21.3% (SEVIR), 20 to 80 times more. The correction is thus local. It raises the heavy-echo hit rate and leaves the rest of the forecast field nearly unchanged.

Perceptual quality. At the native resolution, LPIPS improves on all four datasets, and SSIM improves everywhere except a 0.0009 drop on CIKM. This follows from the construction of the decoder. Added pixels take values from an actual member and removed pixels are only pushed below the threshold, so raising the heavy-echo hit rate introduces no blur. On the four radar benchmarks decoding therefore raises the heavy-echo hit rate at a small pixel-level cost and without loss of perceptual quality. On the Swedish precipitation dataset the pixel-level cost is larger (CRPS +6.8%), as analyzed in § 4.2 of the main text.

Per-Dataset Results

This section reports, for each of the six datasets, the skill versus lead time and two forecast examples, one subsection per dataset. The two precipitation datasets (Swedish and MRMS) also get the full comparison table. Both forecast 16 frames (5 to 80 min) from 4 input frames and are evaluated at the native 256 × 256 resolution. Ours is the decoded ensemble (5.2×).

On the four radar benchmarks, Figure 13–Figure 19 plot each metric against the forecast frame, under the same 128 × 128 protocol as Table 11. CSI is shown at every intensity threshold of each dataset. HSS, POD and FAR are shown averaged over the thresholds and at the highest threshold. For MeteoNet, Shanghai and CIKM, which have fewer thresholds, MAE and RMSE are added. Ours is the decoded ensemble. CRPS, MAE and RMSE are outside the benchmark protocol, and for the baselines they appear only in these figures. For the other metrics, the frame means equal the corresponding entries of Table 1 and Table 11.

Skill metric curves versus forecast frame on SEVIR

Figure 13. Skill versus forecast frame on SEVIR (5 →\to 20 frames, 128 ×\times 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.

Sixteen skill-versus-forecast-frame panels with legend entries Ours, MIMO-VP, SwinLSTM, DiffCast, AlphaPre, PercpCast, exPreCast, FlowCast, SDIR, and PDRF; panels (a) CSI-M, (b) CSI ($\ge 20$ dBZ), (c) CSI ($\ge 30$ dBZ), (d) CSI ($\ge 35$ dBZ), (e) CSI ($\ge 40$ dBZ), (f) HSS ($\ge 40$ dBZ), (g) POD ($\ge 40$ dBZ), (h) FAR ($\ge 40$ dBZ), (i) HSS (thr. avg.), (j) POD (thr. avg.), (k) FAR (thr. avg.), (l) MAE, (m) RMSE, (n) SSIM, (o) LPIPS, and (p) CRPS

Figure 19. Skill versus forecast frame on CIKM (5 → 10 frames, 128 × 128 protocol). The horizontal axis of every panel is the forecast frame.

Across the four benchmarks two patterns hold. At the highest threshold our CSI is above all baselines at every frame, with a FAR that never rises above the range spanned by the baselines, and on SEVIR and CIKM the decoder keeps forecasting heavy echoes after most baselines have stopped forecasting them. At the lowest thresholds the strong methods stay within a few hundredths of each other over the whole horizon, so the differences between methods sit in the heavy tail, which is where the decoder acts. Each benchmark is discussed in its own subsection below, with its curves and two forecast examples.

Swedish Precipitation

Table 15 completes Table 2 with FSS and CRPS. Ours is best in both CSI columns and in SSIM and second in LPIPS. FlowCast is best in FSS and CRPS, where ours ranks fifth. § 4.2 of the main text analyzes this gap.

MethodCSI0.06_{0.06} ↑CSI6.3_{6.3} ↑FSS ↑CRPS ↓SSIM ↑LPIPS ↓
PySteps [31]0.4730.0490.8720.1440.5890.284
DGMR [33]0.4330.0390.8820.1430.7020.204
DiffCast [49]0.4390.0290.9030.1490.6600.292
SSLDM-ISI [21]0.5580.0400.9060.1220.7180.191
AlphaPre [20]0.5410.0250.8940.1490.6470.333
DTCA [16]0.5720.0460.9110.1150.7320.180
RainHCNet [45]0.4990.0250.3860.1530.3320.320
SSRF-Net [25]0.4850.0450.4020.2220.5720.310
exPreCast [39]0.4890.0280.9170.1320.5200.348
FlowCast [34]0.4820.0090.9270.1090.6800.324
SDIR [52]0.4240.0190.9220.1120.7180.307
PDRF [26]0.4740.0370.9050.1490.5710.314
LangPrecip [22]0.5860.0740.9230.1320.7320.193
Ours0.5930.1360.9150.1240.7470.190

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.

Temporal Analysis

At 6.3 mm/h our CSI is above every baseline at every lead time. It falls from 0.266 at 5 min to 0.080 at 80 min, against 0.132 and 0.040 for LangPrecip, the strongest baseline at every lead time. The ratio between the two stays between 1.7 and 2.0 over the whole horizon, so the margin does not narrow with lead time. Seven of the thirteen baselines fall below CSI 0.01 at this threshold, first FlowCast at 25 min and SDIR at 30 min, last DiffCast at 75 min. At 0.06 mm/h our CSI leads LangPrecip by 0.034 at 5 min. The two are level at 45 min, ours trails by at most 0.007 afterwards, and both stay above all other baselines at every lead time.

FSS is the metric where ours trails most. It starts within 0.003 of the best value (SDIR), and the gap to the best baseline grows to about 0.02 at 80 min, where FlowCast and LangPrecip lead. Our SSIM is the highest of all methods up to 35 min and stays within 0.005 of the highest afterwards. Our LPIPS is second to DTCA up to 45 min and stays within 0.011 of it over the whole horizon. Our CRPS is the lowest at 5 and 10 min and then grows by 0.084 over the horizon, against 0.043 for FlowCast and 0.056 for SDIR. FlowCast has the lowest CRPS from 20 min on (0.125 against our 0.153 at 80 min). As on CIKM (§H.6.1), this is the double penalty that pixel errors assign to sharp heavy cores once their position drifts.

Visualization

Predicted scenario c^\hat{c}: The radar echoes drift steadily rightward with a subtle downward tilt, tracing a smooth trajectory across the frame as fragmented, irregular blobs—concentrated in the upper right and lower left—slide cohesively without breaking apart or merging. Their diffuse, filamentous edges remain unchanged as the entire pattern glides uniformly, maintaining internal voids and structural integrity while advancing at a constant pace. No deformation or reorientation emerges; the motion is purely translational, with all elements locked in synchronized drift, preserving their relative positions and shapes throughout the sequence. The system evolves not through transformation but through persistent, linear displacement—right and slightly down—its form frozen in motion, its dynamics reduced to unwavering advection across the screen.

Predicted scenario c^\hat{c}: The radar echoes drift steadily rightward across the frame in a near-straight trajectory, maintaining their fragmented, blob-like forms with soft, undefined edges and scattered internal voids. As they slide uniformly to the right, the entire field undergoes subtle deformation, with no curvature or reorientation emerging. The motion remains consistent in speed and direction, with no acceleration or deviation, while the shapes neither merge nor break apart, preserving their irregular, disorganized layout throughout. No new structures form, and no existing ones dissolve; the system simply translates intact, its evolution defined purely by lateral displacement and gentle contour change, devoid of dynamic reorganization or structural transformation.

Case 1 (Figure 9a) contains two systems. A group of fragmented cells in the north-east, mostly below 30 dBZ, drifts north-east, and a band along the southern edge, whose western end holds cores above 35 dBZ, drifts east while its heavy cores spread along it. Ours keeps a heavy core at the western end of the southern band through +80 min, together with fewer cores above 35 dBZ further east than observed. The northern group keeps its position and orientation and from +50 min merges into one patch with fewer internal gaps than observed. DTCA and LangPrecip keep the fine texture of both systems, and the cores of their southern band fall mostly below 35 dBZ after about +50 min. PySteps and AlphaPre forecast smooth blobs, AlphaPre with cores above 45 dBZ in the first frames that decay below 30 dBZ by +50 min. exPreCast and SDIR shrink the echo area over the horizon, and SDIR keeps only a few weak fragments at +80 min. RainHCNet and SSRF-Net over-smooth the echoes into broad, weak areas, while PDRF forecasts a speckled field with isolated cores above 45 dBZ.

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^\hat{c} that the fine-tuned VLM predicts from the input frames alone.

Case 2 consists of scattered showers over the whole domain, with stratiform areas in the north-west and north-east and a cluster of convective cells in the south whose cores stay above 35 dBZ through +80 min. Over the horizon the small cells in the middle of the domain multiply. Ours keeps the southern cluster at the observed position with cores above 35 dBZ through +80 min, and it keeps the north-western area. It under-predicts the convective cells in the middle after about +40 min, missing the initiation of new cells, and its north-eastern area contracts faster than observed. LangPrecip and DTCA keep the texture of the field, with southern cores mostly below 35 dBZ after about +40 min. DiffCast intensifies the southern cluster above 45 dBZ from +50 min on. PySteps and AlphaPre forecast smooth blobs, PySteps with a core that grows in the north-eastern area. exPreCast and SDIR lose most echoes in the middle of the domain, and RainHCNet and SSRF-Net over-smooth the cells into broad, contiguous areas.

Figure 10 shows all eight members behind the forecast of a third Swedish case, the case of Figure 3 in the main text.

Figure 10. All eight members of the branched ensemble in Figure 3 of the main text at five lead times (s∗=4s^{*} = 4, N=8N = 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\approx6.3 mm/h) at each pixel. Ours (ens.): the field delivered by support-consensus decoding.

MRMS

Table 16 completes Table 2 with FSS and CRPS. Ours ties LangPrecip for the best CSI at 1 mm/h and has the best CSI at 8 mm/h. In FSS, CRPS, SSIM and LPIPS it ranks fourth, sixth, fifth and third.

MethodCSI1_1 ↑CSI8_8 ↑FSS↑CRPS↓SSIM↑LPIPS↓
PySteps [31]0.3380.0240.5790.1670.8710.187
DGMR [33]0.3670.0220.6330.1540.8830.137
DiffCast [49]0.3280.0180.6130.1680.8690.183
SSLDM-ISI [21]0.4160.0240.7040.1350.8970.120
AlphaPre [20]0.3600.0240.6990.1440.8910.197
DTCA [16]0.4230.0260.7110.1310.8990.118
RainHCNet [45]0.3200.0020.6060.1530.8770.207
SSRF-Net [25]0.3360.0090.5730.2190.8440.213
exPreCast [39]0.3940.0150.6930.1900.8280.203
FlowCast [34]0.3930.0130.6990.1330.9000.192
SDIR [52]0.3490.0130.6440.1260.9070.189
PDRF [26]0.3830.0200.6700.1660.8750.190
LangPrecip [22]0.435<u>0.031</u>0.7250.125<u>0.905</u>0.142
Ours0.4350.0360.7030.1360.8980.131

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.

Temporal Analysis

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×).

MRMS is the dataset where our absolute margin is smallest. At 8 mm/h our CSI is the highest at every lead time, by 0.001 to 0.007 over the strongest baseline. It falls from 0.139 at 5 min to 0.010 at 80 min, against 0.136 (LangPrecip) and 0.007 (AlphaPre) for the strongest baseline at the two ends. The relative gain thus grows from 2% at the first frame to between 27% and 44% from 55 min on, at very small absolute values. Every baseline falls below CSI 0.01 at this threshold, between 15 min (RainHCNet) and 65 min (PySteps), while ours stays at or above 0.010. Heavy echoes are very sparse on MRMS, so the vote corrects few pixels. The absolute gain is small (+0.008) although the relative gain is large (+29.5%, Table 5). At 1 mm/h ours leads at 5 min (0.654 against 0.639 for LangPrecip), trails LangPrecip by up to 0.009 between 10 and 30 min, and stays within 0.004 for its 35 min on.

LangPrecip has a higher FSS than ours at every lead time. The gap reaches 0.033 at 55 min and narrows to 0.008 at 80 min, as the FSS of LangPrecip drops faster in the last frames. In SSIM and CRPS ours is level with LangPrecip at 5 min and falls behind the best baseline afterwards, by at most 0.015 and 0.017. SDIR is best in both from about 60 min on. Our LPIPS is third at every lead time, behind DTCA and SSLDM-ISI and ahead of LangPrecip. On MRMS the decoder thus leaves the forecast close to LangPrecip in the area-based and continuous metrics and adds a small but consistent gain at the heavy threshold.

Visualization

Case 1 (Figure 12a) is a rain shield in the northern half of the domain with an embedded core above 8 mm/h near its centre, which persists over the whole horizon with a slight eastward drift, and scattered showers in the south. Ours keeps the outline of the shield and the textured showers in the south. Its core stays near the observed position, with a smaller area above 8 mm/h than observed over the whole horizon. LangPrecip also keeps a textured shield, with a core around 8 mm/h. DiffCast overestimates the core, with a broad area above 8 mm/h and a spurious isolated cell in the south. DGMR, SSLDM-ISI and DTCA keep the texture and weaken the core below 8 mm/h after the first frames. PySteps, AlphaPre, RainHCNet, SSRF-Net, FlowCast and SDIR forecast smooth fields, and the core of AlphaPre, FlowCast and SDIR decays below 4 mm/h. exPreCast shows block artifacts and a spreading weak fringe, and PDRF forecasts a spurious intensification in the south-east.

Figure 12. (a) Case 1. Predicted scenario c^\hat{c}: The radar echo structure drifts steadily rightward with a subtle upward tilt, tracing a near-straight trajectory across the frame as it maintains remarkable directional consistency. As it moves, the entire mass undergoes subtle curvature, its internal bright cores and dark voids reorienting in unison, while the whole system deforms—edges drawing closer together without fragmentation. The leading edge, once sharply defined, softens slightly as it advances, yet the core remains cohesive, resisting dispersion despite rapid translation. Brightness gradients shift subtly within, hinting at internal reorganization, but no new features emerge or break away; instead, the structure evolves through synchronized motion—sliding right, deforming, and compressing—all while preserving its elongated, irregular silhouette. This is a unified, dynamic entity: moving swiftly to the right, turning gently, and holding its form against the flow.

(b) Case 2. Predicted scenario c^\hat{c}: The radar echoes drift steadily rightward across the frame in a near-straight trajectory, maintaining their fragmented, blob-like form without coalescing or breaking apart. Scattered patches, concentrated in the upper-left and trailing toward the lower-right, slide uniformly as a loose cluster, their fuzzy edges never sharpening or organizing into lines or cores. Over time, the entire field glides smoothly to the right with no curvature or deformation—just pure translation, frame after frame, at a constant pace. Internal structure remains inert: no pulses, no splits, no surges—only passive advection carries the disorganized mass horizontally, preserving its shapeless, low-contrast character from start to finish. The motion is linear, unidirectional, and mechanically consistent, as if the whole scene were locked to an invisible conveyor belt moving rightward without turbulence or transformation.

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^\hat{c} that the fine-tuned VLM predicts from the input frames alone.

Case 2 is widespread moderate rain of 1 to 4 mm/h over the western half of the domain with an eastern lobe. Rain above 8 mm/h appears only in a narrow strip along the western edge, which intensifies after +50 min. Ours keeps the layout and texture of the field and forecasts a heavy patch near the western edge from about +20 min on, slightly displaced from the observed strip. The light echoes that fill the gap between the western area and the eastern lobe in the observation are partly missed. PySteps keeps a heavy strip at the western edge and leaves an empty region at the south-western boundary as the field is advected. LangPrecip, DGMR, SSLDM-ISI and DTCA keep the texture with few pixels above 8 mm/h. DiffCast and exPreCast cover the south with a veil of light rain below 0.5 mm/h. AlphaPre, RainHCNet and SSRF-Net blur the field, and SSRF-Net adds a spurious core of 4 to 8 mm/h. PDRF over-intensifies a broad western area to 4 to 8 mm/h after +40 min, and FlowCast and SDIR preserve the layout as an overly smooth field.

SEVIR

Temporal Analysis

At VIL ≥\ge 219 our CSI falls from 0.549 at the first frame to 0.035 at frame 20, while PDRF, the strongest baseline at both ends, goes from 0.489 to 0.023. The relative gain thus rises from 12% to 51%. Seven of the nine baselines fall below CSI 0.01 at this threshold, between frame 7 (SwinLSTM) and frame 19 (DiffCast), whereas ours stays above 0.03. Our POD at this threshold decays from 0.709 to 0.051 and is the highest of all methods from frame 14 on (PercpCast and PDRF lead before, at a higher mean FAR), while our FAR rises from 0.29 to 0.90. At VIL ≥\ge 16 the best baseline stays 0.02 to 0.04 above ours at every frame. Our LPIPS is the lowest at every frame and grows by 0.19 over the horizon, against 0.33 for SDIR. SDIR and FlowCast have a higher mean SSIM and a lower mean CRPS.

At VIL 16 and 74 SDIR and FlowCast reach a higher CSI than ours, and at VIL 133 SDIR does (0.3363 vs. 0.3341). The gap is already present in the anchor, since the decoder acts only at the highest tier (Table 14). Both methods forecast smoother fields, as their higher SSIM and higher (worse) LPIPS show (Table 1) and as the forecast examples in Figure 14 illustrate. A smooth field covers the broad weak-echo area well and loses the heavy cores, where their CSI falls far below ours.

Forecast examples on SEVIR for Case 1 and Case 2
Case 1 forecast examples

(a) Predicted scenario c^\hat{c}: The radar echo complex, anchored in the lower-left quadrant as a bright, irregular core with a sharp leading edge and diffuse trailing tail, surges steadily rightward across the frame with unwavering directionality, propelled by rapid, coherent motion. As it glides horizontally, the entire structure undergoes subtle curvature, its internal gradients and edges reorienting in response to a dominant deformation that twists its form without breaking cohesion. Simultaneously, the system deforms, its boundary sharpness increasing as if drawn by an invisible force, compressing its mass while preserving its elongated silhouette. Though minor ripples flicker along its periphery, the echo resists fragmentation, maintaining structural integrity as it drifts rightward, deforms, and reorients in unison—a dynamically locked entity evolving through synchronized translation, deformation, and contour change, all unfolding smoothly over time without disruption or decay.

Case 2 forecast examples

(b) Predicted scenario c^\hat{c}: The radar echo drifts steadily rightward and slightly upward across the frame, propelled by a consistent motion vector, while its irregular, blob-like form—soft-edged and internally mottled—slowly deforms, with faint voids and brighter cores shifting within as the entire structure reorients. As it glides, the echo’s edges subtly stretch and fray, with internal-core shift and boundary sharpness changes, yet it holds its coherence, never fragmenting or collapsing, maintaining a stable, organized evolution. The deformation gently warps its shape over time, while the reorientation nudges its mass, creating a dynamic but controlled transformation—no sharp breaks, no sudden bursts—just a smooth, persistent drift to the right, with subtle curvature and contour change, all unfolding in harmony as the system evolves without losing its identity.

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^\hat{c} that the fine-tuned VLM predicts from the input frames alone.

Visualization

Case 1 (Figure 14a) is an isolated cell that keeps a core above VIL 219 over the whole horizon while its envelope above VIL 16 widens after t+12t + 12. Our forecast keeps the core above 219 until t+14t + 14 and patches above 181 until t+20t + 20, so the heavy core survives in the frames where the top-tier CSI of most baselines approaches zero (Figure 13g). Two errors are visible. From t+12t + 12 on ours develops a spurious heavy appendage south-west of the core, and its interior at VIL 74 to 100 is larger than observed. SwinLSTM, AlphaPre, FlowCast and SDIR lose the core above 181 by about t+8t + 8, leaving a diffuse blob. exPreCast keeps an intense but oversized patch, and DiffCast breaks the cell into fragments between t+4t + 4 and t+10t + 10. MIMO-VP, PDRF and PercpCast also keep a core above 219, with an envelope that is too small for MIMO-VP and irregular for PDRF.

Case 2 is a decaying system that drifts north-east and reorganizes into an east–west band with thin streaks above VIL 181. Ours keeps a core above 219 until t+10t + 10 and pockets of 160 to 181 afterwards, inside a compact envelope that keeps the original north-west to south-east orientation. It thus captures the intensity of the streaks better than their arrangement. SwinLSTM, exPreCast and MIMO-VP widen the envelope along the old orientation, AlphaPre, FlowCast and SDIR keep its size with a weaker core, and DiffCast and PDRF reproduce the streaky interior in a fragmented form.

MeteoNet

Temporal Analysis

On MeteoNet our margin at the highest threshold is small in the first frames and grows with lead time. At 32 dBZ our CSI is 0.717 against 0.708 (PDRF) at the first frame and 0.132 against 0.111 (SDIR) at frame 20, so the relative gain grows from 1% to 19%. SDIR has a higher POD at this threshold up to frame 15, together with a clearly higher FAR. At 12 dBZ our CSI stays within 0.021 of the best baseline over the whole horizon. Our LPIPS is the lowest from frame 2 on and shows the smallest increase of all methods (0.14). Ours also has the best mean SSIM, while SDIR and FlowCast reach a slightly lower CRPS.

Visualization

Case 1 (Figure 16a) is a convective band whose heavy cells above 32 dBZ are concentrated in its southern part and drift slowly east. Ours keeps a heavy region at the observed location through t+20t + 20 and preserves fine texture in the weaker part of the band. Its heavy region is a contiguous patch that grows from t+10t + 10 on, larger and smoother than the observed group of scattered cells, which is consistent with the rise of our FAR at 32 dBZ to 0.74 at frame 20 (Figure 15h). SwinLSTM, MIMO-VP and FlowCast keep the band outline with few cells above 32 dBZ at the end. exPreCast, PDRF and SDIR keep heavy cores, and SDIR forecasts a large area above 40 dBZ that is almost absent from the observation.

Sixteen-panel line charts showing skill versus forecast frame on MeteoNet

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. (a) Case 1. Predicted scenario c^\hat{c}: The radar echo structure drifts steadily rightward and slightly upward across the frame, maintaining a coherent, elongated form with a bright, dense core that softens into fragmented, filamentous edges as it extends toward the upper-right. As it moves, the entire system undergoes subtle curvature, its internal features reorienting in unison, while simultaneously deforming—its boundaries becoming less sharp, its mass diffusing without breaking apart. The motion is smooth and persistent, with no abrupt shifts or fragmentation; instead, the echo evolves through gradual deformation, its shape subtly warping as it glides diagonally across the screen. This is a unified, dynamic entity—translating, deforming, and reorienting in harmony—as it sweeps from lower-left to upper-right, leaving no voids or sharp breaks in its wake.

(b) Case 2. Predicted scenario c^\hat{c}: The radar echo structure drifts steadily upward and slightly rightward across the frame, maintaining a coherent, elongated form with a dense core near center-left that gradually tapers toward the upper-right. As it moves, the entire system undergoes subtle curvature, its filamentous arms—especially along the right flank—twisting into curved contours that evolve smoothly without breaking apart. These arms show sharp-edged gradients, particularly on the left where abrupt transitions suggest strong shear, yet the whole structure holds its shape, resisting fragmentation as it glides. Internal voids flicker faintly within the core but do not disrupt the flow, while the motion remains unwavering in direction and speed, propelled by a stable, large-scale force. The deformation is subtle but persistent, curling the edges without distorting the whole, as the system advances diagonally upward-right, its form locked in dynamic equilibrium—translating, reorienting, but never unraveling.

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^\hat{c} that the fine-tuned VLM predicts from the input frames alone.

Case 2 has two heavy clusters made of fine-grained cells, one along the north-eastern band and one in the south-west. Ours keeps both clusters through t+20t + 20 with a textured interior and merges the north-eastern cells into a single patch larger than observed from about t+12t + 12. MIMO-VP loses all echoes above 32 dBZ after t+4t + 4, and AlphaPre breaks both clusters into scattered, noisy fragments. exPreCast and SDIR keep the clusters with cores above 40 dBZ that are larger than observed. SwinLSTM, DiffCast and FlowCast keep the north-eastern cluster and weaken the south-western one.

Shanghai

Temporal Analysis

On Shanghai the margin at the highest threshold is small at both ends of the horizon. At 40 dBZ our CSI goes from 0.756 to 0.167, against 0.728 (PDRF) and 0.158 (PercpCast) for the strongest baseline at the two ends, a relative gain of 4% at the first frame and 6% at the last. PDRF and then PercpCast have a higher POD at this threshold, at a higher mean FAR, and our FAR is below the baseline range in the first two frames and within it afterwards. At 20 dBZ ours leads the best baseline at every frame, by at most 0.018. Shanghai is the only benchmark where ours has the best mean value of every continuous metric shown, namely SSIM, LPIPS, CRPS, MAE and RMSE.

Visualization

Case 1 (Figure 18a) is an elongated line whose interior stays between 35 and 45 dBZ, with scattered cores above 45 dBZ and a narrow rim. Ours keeps the elongated outline, the narrow rim and the textured interior with scattered cores above 45 dBZ through t+20t+20, in line with its low LPIPS (Figure 17). It misses the few pixels above 55 dBZ that appear between t+10t+10 and t+18t+18. SwinLSTM, exPreCast, AlphaPre, FlowCast and SDIR forecast a smooth band with a wide weak rim, and MIMO-VP weakens the interior below 45 dBZ. PDRF and PercpCast fill most of the interior above 45 dBZ and partly above 55 dBZ, which overestimates the heavy area. DiffCast keeps the fine texture but breaks the line into fragments.

Sixteen-panel line charts showing skill versus forecast frame for Shanghai across CSI-M, CSI, HSS, POD, FAR, MAE, RMSE, SSIM, LPIPS, and CRPS metrics

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^\hat{c} that the fine-tuned VLM predicts from the input frames alone.

Case 2 is a bow-shaped echo whose core above 45 dBZ persists and widens to the north over the horizon. Ours keeps the curvature over all frames and the core above 45 dBZ until about t+10t+10, after which the core falls to 40 to 45 dBZ and the outline frays. SDIR loses most of the intensity by t+10t+10 and AlphaPre by t+18t+18. SwinLSTM and FlowCast keep a heavy core inside a shrunken or bloated bow, and PDRF turns the bow into a compact blob. PercpCast keeps both the curvature and a core above 45 dBZ until t+20t+20, with an interior above 55 dBZ that is more intense than observed.

CIKM

Temporal Analysis

CIKM forecasts ten frames. At 40 dBZ our CSI goes from 0.5331 to 0.0337, against 0.5050 (SDIR) and 0.0294 (PercpCast) for the strongest baseline at the two ends, and six of the nine baselines fall below 0.01 by frame 9. The relative gain grows from 6% to 15%, and our POD at this threshold is the highest from frame 5 on. At 20 dBZ ours trails the best baseline at every frame, by 0.002 at the first frame and 0.017 at the end. CIKM is also the benchmark where the cost of sharp forecasts is most visible. PercpCast has the lowest LPIPS at every frame, and our MAE, RMSE and CRPS grow faster than those of FlowCast and SDIR, with RMSE rising by 5.14 over the ten frames against 4.47 and 4.58. This matches the double penalty that pixel errors assign to sharp echoes once their position drifts.

Visualization

Case 1 (Figure 20a) is a band of echoes above 40 dBZ that drifts toward the bottom edge and broadens eastward, under a widening area of 30 to 35 dBZ in the upper right. Ours keeps a core above 40 dBZ near the observed position through t+10t+10. The core stays more compact than the observed band, and the upper-right area is forecast at 25 to 30 dBZ, weaker than observed. SwinLSTM and exPreCast over-smooth the echoes into a broad 30 to 40 dBZ region, and exPreCast also shows block artifacts. AlphaPre and SDIR lose the core by t+7t+7. DiffCast and PDRF keep a heavy core of similar quality to ours, FlowCast keeps an elongated and displaced core, and PercpCast keeps the core with a noisy texture.

Figure 20. (a) Case 1. Input t−4t-4 t−3t-3 t−2t-2 t−1t-1 tt; Ground truth t+1t+1 t+2t+2 t+3t+3 t+4t+4 t+5t+5 t+6t+6 t+7t+7 t+8t+8 t+9t+9 t+10t+10; Ours; SwinLSTM; MIMO; exPreCast; AlphaPre; DiffCast; PDRF; PercpCast; FlowCast; SDIR; Reflectivity (dBZ). Predicted scenario c^\hat{c}: The echo complex initiates motion in the upper-right quadrant, rapidly translating rightward with a slight downward drift along a consistent trajectory, resulting in a noticeable displacement over time. As it advances, the elongated, irregular band undergoes deformation, with its leading edge sharpening while the trailing side diffuses, and internal bright cores shift amid dark voids, indicating dynamic reorganization. A distinct curvature emerges, particularly in the central region, where contour changes develop, reinforcing a coherent mesoscale structure. Throughout the sequence, the system maintains structural integrity, with no fragmentation or dissipation, as deformation and reorientation interact to sustain a stable, organized evolution. The overall motion combines rapid translation, persistent curvature, and gradual contour change, tracing a continuous path of coordinated movement and transformation across the frame. (b) Case 2. Input t−4t-4 t−3t-3 t−2t-2 t−1t-1 tt; Ground truth t+1t+1 t+2t+2 t+3t+3 t+4t+4 t+5t+5 t+6t+6 t+7t+7 t+8t+8 t+9t+9 t+10t+10; Ours; SwinLSTM; MIMO; exPreCast; AlphaPre; DiffCast; PDRF; PercpCast; FlowCast; SDIR; Reflectivity (dBZ). Predicted scenario c^\hat{c}: The radar echo structure drifts steadily rightward with a subtle downward tilt, tracing a smooth trajectory across the frame at a steady pace, while its dense core—anchored in the lower-left—pulls fragmented, wispy tendrils behind it as they stretch and dissolve toward the upper-right. As it glides, the entire mass undergoes subtle curvature, its leading edge curling like a slow contour change, while internal voids shift and reconfigure without breaking cohesion. The outer fringes continuously fray and re-knit, yet the system holds its form, resisting fragmentation as it slides unswervingly right-downward, its motion fluid and persistent, propelled by a balance of deformation and contour change that keeps its shape intact even as its edges ripple and reform in response to unseen forces. 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^\hat{c} that the fine-tuned VLM predicts from the input frames alone.

Case 2 starts with a heavy band on the left and a core above 45 dBZ at the bottom. After t+4t+4 the observation breaks up into small scattered cells of 35 to 45 dBZ over the whole domain while the bottom core decays. No method reproduces this rapid break-up. Ours keeps the left band until t+5t+5 and then forecasts a broad field of 25 to 35 dBZ with few heavy cells, which overestimates the area of moderate echoes. This case illustrates the growth of our pixel errors on CIKM (Figure 19, m). AlphaPre and PercpCast produce a speckled field closer in texture to the observation, with cells at the wrong positions, and exPreCast keeps the bottom core long after it has decayed.

Implementation Details

Table 17 lists the settings of the generator, the autoencoder, the generator optimization and the student VLM. The 5.2× cost in the main text counts network evaluations (52 against 10). Measured on one RTX 5090, the wall-clock ratio is 4.4×, as the eight branches after the shared trunk are evaluated together in a single batch at each step.

Generator
BackboneDiT, 28 layers, width 1152, 16 heads, patch (1, 2, 2) over (time, height, width)
Parameters583.93M (435.29M visual backbone + 148.64M text cross-attention)
Frames (condition + forecast)Swedish precipitation / MRMS 4 + 16, SEVIR / MeteoNet / Shanghai 5 + 20, CIKM 5 + 10
Tokens per sampleSwedish precipitation / MRMS 5120, SEVIR / MeteoNet / Shanghai 6400, CIKM 3840
ObjectiveRectified Flow
SamplerEuler, T=10T = 10 steps
Average inference time2.0 s single forecast, 8.8 s 8-member ensemble (one RTX 5090, batch size 1)
Autoencoder
ArchitectureFLUX.2 autoencoder [2], base channels 128, multipliers [1, 2, 4, 4], 2 residual blocks per level
Parameters84.05M (encoder 34.43M + decoder 49.62M)
Latent per frame32 × 32 × 32 for a 256 × 256 frame (8× spatial downsampling, 2× fewer values)
Input / outputrange [−1, 1], first of the 3 output channels kept
Generator optimization
OptimizerAdamW, β=(0.9,0.999)\beta= (0.9, 0.999), weight decay 0.01, ϵ=10−8\epsilon= 10^{-8}
Learning rate1×10−51 \times10^{-5}
Scheduleconstant learning rate after 200 warm-up steps
Regularization and precisionEMA 0.9999, gradient clipping 1.0, bf16, gradient checkpointing
Batch size32–36
Hardware6× NVIDIA RTX 4090
Student VLM
ModelQwen3.5-9B [32], base weights frozen, condition frames hh as input, maximum length 1024 tokens
LoRArank 8, α=32\alpha= 32, dropout 0.05, all linear layers of the language model, vision encoder and merger
OptimizerAdamW, β=(0.9,0.95)\beta= (0.9, 0.95), weight decay 0.1, ϵ=10−8\epsilon= 10^{-8}
Learning rate2×10−52 \times10^{-5}, cosine schedule after 100 warm-up steps
Training2 epochs, batch size 12, bf16

Table 17. Implementation details.

References

  1. [1]Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Teo Karras, and Ming-Yu Liu. eDiff-I: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022.arxiv.org/abs/2211.01324
  2. [2]Black Forest Labs. FLUX.2: Frontier Visual Intelligence. https://bfl.ai/blog/flux-2, 2025.
  3. [3]Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22563–22575, 2023.arxiv.org/abs/2304.08818
  4. [4]Adam J. Clark. Generation of ensemble mean precipitation forecasts from convection-allowing ensembles. Weather and Forecasting, 32(4):1569–1583, 2017.DOI
  5. [5]Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In Proceedings of the 35th International Conference on Machine Learning (ICML), 2018.arxiv.org/abs/1802.07687
  6. [6]Yilun Du, Mengjiao Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Brian Ichter, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B. Tenenbaum, Leslie Kaelbling, Andy Zeng, and Jonathan Tompson. Video language planning. In International Conference on Learning Representations (ICLR), 2024.arxiv.org/abs/2310.10625
  7. [7]Elizabeth E. Ebert. Ability of a poor man’s ensemble to predict the probability and distribution of precipitation. Monthly Weather Review, 129(10):2461–2480, 2001.doi.org/10.1175/1520-0493(2001)129<2461:aoapms>2.0.co;2
  8. [8]Wenzhi Feng, Xutao Li, Zhe Wu, Kenghong Lin, Demin Yu, Yunming Ye, and Yaowei Wang. Perceptually constrained precipitation nowcasting model. In Proceedings of the 42nd International Conference on Machine Learning (ICML), pages 16777–16795, 2025.
  9. [9]Gabriele Franch, Elena Tomasi, Rishabh Wanjari, Virginia Poli, Chiara Cardinali, Pier Paolo Alberoni, and Marco Cristoforetti. GPTCast: a weather language model for precipitation nowcasting. Geoscientific Model Development, 18(16):5351–5371, 2025.DOI
  10. [10]Zhihan Gao, Xingjian Shi, Hao Wang, Yi Zhu, Yuyang Bernie Wang, Mu Li, and Dit-Yan Yeung. Earthformer: Exploring space-time transformers for earth system forecasting. Advances in Neural Information Processing Systems, 35:25390–25403, 2022.arxiv.org/abs/2207.05833
  11. [11]Zhangyang Gao, Cheng Tan, Lirong Wu, and Stan Z. Li. SimVP: Simpler yet better video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3170–3180, 2022.arxiv.org/abs/2206.05099
  12. [12]Zhihan Gao, Xingjian Shi, Boran Han, Hao Wang, Xiaoyong Jin, Danielle Maddix, Yi Zhu, Mu Li, and Yuyang Bernie Wang. PreDiff: Precipitation nowcasting with latent diffusion models. Advances in Neural Information Processing Systems, 36:78621–78656, 2023.arxiv.org/abs/2307.10422
  13. [13]Junchao Gong, Lei Bai, Peng Ye, Wanghan Xu, Na Liu, Jianhua Dai, Xiaokang Yang, and Wanli Ouyang. CasCast: Skillful high-resolution precipitation nowcasting via cascaded modelling. In Proceedings of the 41st International Conference on Machine Learning, pages 15809–15822. PMLR, 2024.arxiv.org/abs/2402.04290
  14. [14]Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In Advances in Neural Information Processing Systems (NeurIPS), 2022.
  15. [15]Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022.arxiv.org/abs/2106.09685
  16. [16]ChaoRong Li, XuDong Ling, YiLan Xue, WenJie Luo, LiHong Zhu, FengQing Qin, Yaodong Zhou, and Yuanyuan Huang. Precipitation nowcasting using diffusion transformer with causal attention. IEEE Transactions on Geoscience and Remote Sensing, 62:1–16, 2024.arxiv.org/abs/2410.13314
  17. [17]Haobo Li, Zhaowei Wang, Jiachen Wang, Yue Ya Wang, Alexis Kai Hon Lau, and Huamin Qu. Cllimate: A multimodal benchmark for weather and climate events forecasting. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 17547–17573, 2025.arxiv.org/abs/2409.19058
  18. [18]Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. LLM-grounded video diffusion models. In International Conference on Learning Representations (ICLR), 2024.arxiv.org/abs/2309.17444
  19. [19]Han Lin, Ahyah Zala, Jiamin Chao, and Mohit Bansal. VideoDirectorGPT: Consistent multi-scene video generation via LLM-guided planning. In Conference on Language Modeling (COLM), 2024.arxiv.org/abs/2309.15091
  20. [20]Kenghong Lin, Baoquan Zhang, Demin Yu, Wenzhi Feng, Shidong Chen, Feifan Gao, Xutao Li, and Yunming Ye. AlphaPre: Amplitude-phase disentanglement model for precipitation nowcasting. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 17841–17850, 2025.DOI
  21. [21]XuDong Ling, ChaoRong Li, LiHong Zhu, FengQing Qin, Ping Zhu, and Yuanyuan Huang. Spacetime separable latent diffusion model with intensity structure information for precipitation nowcasting. IEEE Transactions on Geoscience and Remote Sensing, 62:1–15, 2024.DOI
  22. [22]XuDong Ling, ChaoRong Li, Tianxi Huang, Qian Dong, and Guiduo Duan. LangPrecip: Language-aware multimodal precipitation nowcasting. In Proceedings of the 43rd International Conference on Machine Learning (ICML), 2026.arxiv.org/abs/2512.22317
  23. [23]Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In International Conference on Learning Representations (ICLR), 2023.arxiv.org/abs/2210.02747
  24. [24]Xingchao Liu, Chaoqiye Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In International Conference on Learning Representations (ICLR), 2023.arxiv.org/abs/2209.03003
  25. [25]Wenj ie Luo, Chaorong Li, Xudong Ling, Chuanhu Deng, and Zhuo Wang. SSRF-Net: A stagewise scheduled rainfall forecasting network with an asymmetric architecture. IEEE Transactions on Geoscience and Remote Sensing, 63:1–18, 2025.DOI
  26. [26]Wenjie Luo, Chaorong Li, Chuanhu Deng, and Zhuo Wang. Physically-guided data-space rectified flow for precipitation nowcasting. In Proceedings of the 43rd International Conference on Machine Learning (ICML), 2026.
  27. [27]Xinyin Ma, Gongfan Fang, and Xinchao Wang. DeepCache: Accelerating diffusion models for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024.arxiv.org/abs/2312.00858
  28. [28]Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations (ICLR), 2022.arxiv.org/abs/2108.01073
  29. [29]Shuliang Ning, Mengcheng Lan, Yanran Li, Chaofeng Chen, Qian Chen, Xunlai Chen, Xiaoguang Han, and Shuguang Cui. MIMO is all you need: A strong multi-in-multi-out baseline for video prediction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1975–1983, 2023.
  30. [30]William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195–4205, 2023.arxiv.org/abs/2212.09748
  31. [31]Seppo Pulkkinen, Daniele Nerini, Andrés A Pérez Hortal, Carlos Velasco-Forero, Alan Seed, Urs Germann, and Loris Foresti. PySteps: An open-source Python library for probabilistic precipitation nowcasting (v1.0). Geoscientific Model Development, 12(10):4185–4219, 2019.
  32. [32]Qwen Team. Qwen3.5: Towards native multimodal agents. https://qwen.ai/blog/?id=qwen3.5, 2026.
  33. [33]Suman Ravuri, Karel Lenc, Matthew Willson, Dmitry Kangin, Remi Lam, Piotr Mirowski, Megan Fitzsimons, Maria Athanassiadou, Sheleem Kashem, Sam Madge, Rachel Prudden, Amol Mandhane, Aidan Clark, Andrew Brock, Karen Simonyan, Raia Hadsell, Niall Robinson, Ellen Clancy, Alberto Arribas, and Shakir Mohamed. Skilful precipitation nowcasting using deep generative models of radar. Nature, 597(7878):672–677, 2021.DOI
  34. [34]Bernardo Perrone Ribeiro and Jana Faganeli Pucer. FlowCast: Advancing precipitation nowcasting with conditional flow matching. In International Conference on Learning Representations (ICLR), 2026.
  35. [35]Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022.arxiv.org/abs/2112.10752
  36. [36]Lei She, Chenghong Zhang, Xin Man, and Jie Shao. LLMDiff: Diffusion model using frozen LLM transformers for precipitation nowcasting. Sensors, 24(18):6049, 2024.DOI
  37. [37]Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems, 28, 2015.arxiv.org/abs/1506.04214
  38. [38]Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-video generation without text-video data. In International Conference on Learning Representations (ICLR), 2023.arxiv.org/abs/2209.14792
  39. [39]Changhoon Song, Teng-Yuan Chang, and Youngjoon Hong. Extreme weather nowcasting via local precipitation pattern prediction. In International Conference on Learning Representations, pages 109146–109177, 2026.arxiv.org/abs/2602.05204
  40. [40]Cheng Tan, Siyuan Li, Zhangyang Gao, Wenfei Guan, Zedong Wang, Zicheng Liu, Lirong Wu, and Stan Z. Li. OpenSTL: A comprehensive benchmark of spatio-temporal predictive learning. In Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2023.arxiv.org/abs/2306.11249
  41. [41]Song Tang, Chuang Li, Pu Zhang, and RongNian Tang. Swinlstm: Improving spatiotemporal prediction accuracy using swin transformer and lstm. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 13424–13433. IEEE, 2023.arxiv.org/abs/2308.09891
  42. [42]Shuo Tang, Jian Xu, Jiadong Zhang, Yi Chen, Qizhao Jin, Lingdong Shen, Chenglin Liu, and Shiming Xiang. MeteorPred: A meteorological multimodal large model and dataset for severe weather event prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22910–22919, 2026.arxiv.org/abs/2508.06859
  43. [43]Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. MCVD: Masked conditional video diffusion for prediction, generation, and interpolation. In Advances in Neural Information Processing Systems (NeurIPS), 2022.
  44. [44]Wan Team. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314; Wan2.2 release, 2025.arxiv.org/abs/2503.20314
  45. [45]Lei Wang, Zheng Wang, Wenjun Hu, and Cong Bai. RainHCNet: Hybrid high-low frequency and cross-scale network for precipitation nowcasting. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 18:8923–8937, 2025.
  46. [46]Yunbo Wang, Mingsheng Long, Jianmin Wang, Zhifeng Gao, and Philip S. Yu. PredRNN: Recurrent neural networks for predictive learning using spatiotemporal LSTMs. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 1, 2
  47. [47]Jiarui Xing, Song Wang, and Jian Wang. Divergence is uncertainty: A closed-form posterior covariance for flow matching. arXiv preprint arXiv:2605.00941, 2026. 5
  48. [48]Yilun Xu, Mingyang Deng, Xiang Cheng, Yonglong Tian, Ziming Liu, and Tommi Jaakkola. Restart sampling for improving generative processes. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 3arxiv.org/abs/2306.14878
  49. [49]Demin Yu, Zuxi Li, Yunming Ye, Baoquan Zhang, Chuyao Luo, Kuai Dai, Rui Wang, and Yunlai Chen. DiffCast: A unified framework via residual diffusion for precipitation nowcasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27758–27767, 2024. 1, 2, 6, 7, 8, 17, 18, 20, 24arxiv.org/abs/2312.06734
  50. [50]Haoran Zhao, Tong Bai, Lei Huang, and Xiaoyu Liang. SAGE: Semantic-aware shared sampling for efficient diffusion. arXiv preprint arXiv:2509.15865, 2025. 3arxiv.org/abs/2509.15865
  51. [51]Xiangyu Zhao, Zhiwang Zhou, Wenlong Zhang, Yihao Liu, Xiangyu Chen, Junchao Gong, Hao Chen, Ben Fei, Shiqi Chen, Wanli Ouyang, Xiao-Ming Wu, and Lei Bai. WeatherGFM: Learning a weather generalist foundation model via in-context learning. In International Conference on Learning Representations (ICLR), 2025. 3arxiv.org/abs/2411.05420
  52. [52]Yunlong Zhou, Chen Zhao, Danyang Peng, Fanji Li, and Xiao-Tong Yuan. Learning to refine: Spectral-decoupled iterative refinement framework for precipitation nowcasting. In Proceedings of the 43rd International Conference on Machine Learning (ICML), 2026. 2, 6, 8, 18, 20, 24arxiv.org/abs/2606.02661
  53. [53]Zhiwang Zhou, Yuandong Pu, Xuming He, Yidi Liu, Yixin Chen, Junchao Gong, Xiang Zhuang, Wanghan Xu, Qinglong Cao, Shixiang Tang, Yihao Liu, Wenlong Zhang, and Lei Bai. Omni-Weather: A unified multimodal model for weather radar understanding and generation. In International Conference on Learning Representations (ICLR), 2026. 3arxiv.org/abs/2512.21643

Paper details

vlnc:W_vzc4famx
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents