Generative modeling based on stochastic interpolants requires specifying two key choices: an interpolant schedule bridging the initial and target distributions, and an efficient discretization of the resulting probability-flow ODE. In this work, we introduce a new framework for joint schedule/discretization design based on the analyticity properties of the (complexified) ODE trajectories, which in turn ensure good numerical behavior. Our formulation can be solved exactly via conformal mapping, yielding a natural family of variance-preserving elliptic schedules, parametrized by Jacobi elliptic functions. Pairing an elliptic schedule with a tailored discretization, we obtain Zoloflow, a new sampling algorithm closely connected to Zolotarev's optimal rational function approximants. Crucially, both the elliptic schedule and Zoloflow explicitly incorporate the condition number κ of the target distribution and reduce to the classical trig schedule in the limit κ→1. For Gaussian targets, Zoloflow achieves minimax optimality among all oblivious score-span algorithms, attaining total-variation error ε using Θ(logκlog(d/ε)) (exact) score queries. For the log-concave case, the elliptic schedule allows sampling with at most ε 2-Wasserstein-error by requiring merely O(dpolylog(κ,1/ε)) score evaluations. Complementing these theoretical results, we empirically evaluate the elliptic schedule on general (non-log-concave) target distributions, including detailed comparisons with alternative schedule proposals on both synthetic benchmarks and existing image generation datasets.
Figures & tables
Figure 1: Side-by-side comparison of schedules. Images are generated by Stable Diffusion 3.5 Medium (CFG = 4.5) using DPM-Solver++ with only 4 neural network evaluations (4 oracle calls). From left to right, the schedules are the target-rescaled linear, trigonometric, and elliptic (ours). The schedules are rescaled by the hyperparameters shown in Table 2 . The text prompt is shown below the images.
Figure 2: Elliptic schedules for several values of κ : (a) αt , (b) βt and (c) the signal-to-noise ratio βt2/αt2=\jsc2(Kt∣m) on a logarithmic axis. For κ=1 the elliptic schedule is the trigonometric schedule.
Figure 3: Common schedules next to the elliptic schedule ( κ=102 ): (a) βt and (b) the signal-to-noise ratio βt2/αt2 on a logarithmic axis. EDM ( Karras et al., 2022b ) uses the VP-normalized interpolant (αt,βt)=(σρ(t),1)/1+σρ(t)2 , as in Section K.1 , with σmax=80 , σmin=0.002 , and ρ=7 ; its SNR σρ(t)−2 is finite at both ends. The Lipschitz-guided schedule ( Chen et al., 2025 ) and the elliptic schedule use κ=102 .
Figure 4: Endpoint and center–width coordinates. Black dots represent the eigenvalues λi(Λ) of an illustrative precision matrix Λ=Σ−1 . For the matched choice ℓ=λmin and L=λmax , the geometric center λˉ=ℓL lies at the midpoint of the interval on the logarithmic axis, and the interval has logarithmic width logκ=log(L/ℓ) . Here λmin=10−1 and λmax=103 , so λˉ=10 and κ=104 .
Figure 5: Rescaling and the additional benefit of elliptic width. DPM-Solver++(2M) with exact scores and eight score evaluations. The left column shows unscaled linear, rescaled linear, and ground truth; the right shows unscaled trig, rescaled trig, and the best tested elliptic setting. Unscaled baselines use λˉ=1 ; rescaled linear (c), rescaled trig (d), and elliptic (f) share λˉ≈31.6 . Comparing rescaled trig (d) with elliptic (f), at κ≈1.78×104 , isolates the additional width effect at this center; rescaled linear (c) provides another calibrated baseline. Each sample panel contains 105 draws, with common initial noise across sampler runs. Panel (g) shows histogram TV over the tested center–width grid (logarithmic color scale), with markers identifying (b), (d), and (f). Panel (h) minimizes over tested centers at each width; the lighter horizontal lines show unscaled references. The minimizing center remains λˉ≈31.6 from κ=1 through the selected minimum, so this descent also holds at a fixed center. Histogram TV uses 100×100 bins plus an outside bin and a reference of 106 target draws. The grey dotted line shows an independent target draw against that reference, not a lower bound. Rescaled baselines are selected over 30 centers; elliptic is selected over the same centers and 42 widths. These are best observed settings within the tested grids.
Figure 6: Worst-case relative error E against number of score evaluations r on a Gaussian target N(0,Λ−1) with spectrum \spec(Λ)⊂[1,100] . The panels, from left to right, are the Euler, DPM-Solver++(2M), and Elliptic Finite-Difference solvers. Colors distinguish Elliptic, Trig, Linear (VP form), their center-rescaled Trig and Linear variants, and Lipschitz-guided, with λˉ and κ where applicable given in the legend; blue is our elliptic schedule. The black dashed line, the elliptic finite difference discretization with the Elliptic schedule, is repeated in every panel. All six solvers are shown in Figure 13 .
Figure 7: FID-10K over the (λˉ,κ) grid at 4 oracle calls of SiT-XL/2 , using the target-rescaled elliptic schedule with DPM-Solver++(2M). Each cell is annotated with its FID-10K (lower is better), and the colors use a logarithmic scale. The solid box marks the best target-rescaled elliptic configuration, and the dashed box marks the best target-rescaled trigonometric ( κ=1 ) configuration. The white diamond represents the heuristic (λˉ^,κ^) configuration, achieving FID-10K 43.67.
Unscaled
Rescaled
Calls
Schedule
λˉ
κ
FID ↓
IS ↑
λˉ
κ
FID ↓
IS ↑
4
Linear
1
—
40.28
67.9
0.22
—
32.36
92.5
Trig.
1
1
27.93
96.6
0.46
1
27.34
101.8
Elliptic ( ours )
1
1
27.93
96.6
0.46
31.6
25.46
108.5
8
Linear
1
—
14.08
139.5
0.1
—
6.07
198.4
Trig.
1
1
6.88
178.4
0.46
1
4.52
201.9
Table 1: FID-50K and IS by schedule and number of oracle calls of the SiT-XL/2 model. Unscaled schedules fix ℓ=1 , so λˉ=1 for the linear and trigonometric schedules; the unscaled elliptic schedule also fixes ℓ=1 but tunes L , so κ=L and λˉ=κ . Rescaled schedules use the λˉ=ℓL hyperparameter, and the elliptic schedule has its own κ=L/ℓ parameter, which is by default 1 for the trigonometric schedule (so ℓ=L=λˉ for every schedule but the elliptic one). Hyperparameters are chosen from those achieving the best FID-10K in our sweep at each budget. The unscaled elliptic schedule is best at κ=1 (so λˉ=1 ), where it coincides with the trigonometric one. Best per budget in bold.
Unscaled
Rescaled
Calls
Schedule
λˉ
κ
FID ↓
λˉ
κ
FID ↓
4
Linear
1
—
138.13
0.056
—
17.74
Trig.
1
1
186.75
0.056
1
24.95
Elliptic ( ours )
1
1
186.75
0.056
316
17.30
8
Linear
1
—
34.24
0.32
—
12.32
Trig.
1
1
59.74
0.18
1
14.41
Table 2: FID-30K by schedule and number of oracle calls of the SD 3.5M model. Unscaled schedules fix ℓ=1 , so λˉ=1 for the linear and trigonometric schedules; the unscaled elliptic schedule also fixes ℓ=1 but tunes L , so κ=L and λˉ=κ . Rescaled schedules use the λˉ=ℓL hyperparameter, and the elliptic schedule has its own κ=L/ℓ parameter, which is by default 1 for the trigonometric schedule (so ℓ=L=λˉ for every schedule but the elliptic one). Hyperparameters are chosen from those achieving the best FID-5K in our sweep at each budget. The unscaled elliptic schedule is best at κ=1 (so λˉ=1 ), where it coincides with the trigonometric one. Best per budget in bold.
Appendix figures & tables27 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Angular and elliptic clocks (φ,u) for varying parameters m
Figure 9: Angular and elliptic clock speeds (φ,u) for varying parameters m
Function
m=0
m→1
φ(u)=amu
u
arcsin(tanhu)
\snu=sin(amu)
sinu
tanhu
\cnu=cos(amu)
cosu
sechu
\dnu=1−msin2(amu)
1
sechu
Appendix
Table 3: Limiting cases of the Jacobi elliptic functions at m=0 and m=1 .
Figure 10: Elementary Jacobi Elliptic functions under raw clock u
Figure 11: Elementary Jacobi Elliptic functions under normalized clock t=u/K(m)
Figure 12: Plots of holomorphic extensions to ∣\sn∣ and ∣\cn∣ in normalized time t=K(m)u , with parameter m=0.64 (modulus ℓ=m=0.8 in the plot labels). Zeros are labeled with “ ∘ " and poles are labeled with “ × ." The white lines represent ∣f(z)∣=1 level sets. Lighter indicates ∣f(z)∣ large and darker indicates ∣f(z)∣ small.
Figure 13: Worst-case relative error E against the number of score evaluations r on the spectrum [1,100] . Panels: solvers; colors: the six schedule settings, including Linear in VP form and its center-rescaled variant. Legends give λˉ and κ where applicable; curves are drawn only at the r valid for each pair. The black dashed line, the elliptic FD rule with the Elliptic schedule ( E100,r ), is repeated in every panel.
Figure 14: Worst-case relative error E against κ on the spectrum [1,κ] with r=16 ( r=14 for RK4 with Elliptic, Trig and Linear, including their rescaled variants, where 16 is not valid; RK4 with Lipschitz-guided uses r=16 ). Layout as in Figure 13 .
Figure 15: Gaussian target with spectrum [0.01,1] , r=12 : E against the center λˉ=ℓL , one curve per width κ=L/ℓ ; stars mark each curve’s minimum over λˉ , in the middle of its flat segment where there is one, on the grid log10λˉ∈[−3,1] (step 0.05 ), log10κ∈[0,4] (step 0.125 ). Blue circle: (λˉ,κ)=(λminλmax,λmax/λmin) ; pink triangle: the best κ at λˉ=1 ; orange square: the best rescaled trig, κ=1 .
Figure 16: KL (F#γ∥p1) against the number of score evaluations r . Rows: the two-dimensional targets of Table 4 , grouped by regime (A, B; E1; G); columns: solvers. Compare purple (Elliptic, ours) with orange (rescaled trig) and blue (curvature endpoints); Trig (pink) coincides with rescaled trig wherever c≈1 (A, B, E1). A × on the bottom edge marks a folded map (KL undefined).
Figure 17: KL /d of DPM-Solver++(2M) with r=8 against the center λˉ=ℓL , one curve per width κ=L/ℓ , over the grid log10ℓ∈[−3,2.5] , log10κ∈[0,4] ; top H1, bottom H2. Stars mark each curve’s minimum. The purple cross (Elliptic, ours) sits near the lowest stars; the dashed line is Lipschitz-guided at the same (λˉ,κ) .
target
d
coordinates
A
2
curvature ranges [0.1,1] , [1,10]
B
2
curvature ranges [0.01,1] , [1,100]
E1
2
v=(1,1) , ρ2=(0,0.99)
G
2
Gaussian, v=(1,10−3)
H1
32
vk=10−2k/31 , ρk2=0.8gk
H2
32
vk=10−k/31 , ρk2=0.9+0.09gk
Appendix
Table 4: Log-concave targets. For H1 and H2, k=0,…,31 and gk=(k+1)ω−⌊(k+1)ω⌋ with ω=(5−1)/2 .
Figure 18: Frond covariance-center comparison, DPM-Solver++(2M) with exact scores and r=8 score evaluations. Panel (g) uses the covariance width κ≈36.1 ; panels (b) and (h) use κ=104 , selected by the elliptic NLL sweep and matched for Lipschitz-guided. All rescaled schedules use the same covariance center λˉ≈2.56 . Each panel shows 105 samples, with common initial noise across sampler runs.
Figure 19: Frond width sweep at the fixed covariance center λˉ≈2.56 , with the same solver, score budget and initial noise as Figure 18 . The elliptic curve includes centered trig at κ=1 ; the cross marks the covariance width and the star marks the lowest NLL among the tested widths. The diamond is Lipschitz-guided at the selected elliptic width, not an independently tuned optimum. Horizontal lines show the centered trig and linear baselines.
Figure 20: Helix target, DPM-Solver++(2M) with r=8 score evaluations. Rows, top to bottom: exact samples, Trig, Rescaled trig, Lipschitz-guided, Elliptic.
Figure 21: ECG target, DPM-Solver++(2M) with r=8 score evaluations. Rows, top to bottom: exact samples, Trig, Rescaled trig, Lipschitz-guided, Elliptic.
Figure 22: “ZOLO” text target, DPM-Solver++(2M) with r=8 score evaluations. Rows, top to bottom: exact samples, Trig, Rescaled trig, Lipschitz-guided, Elliptic.
Figure 23: Radial tree target (isotropic covariance, so Elliptic coincides with rescaled trig), DPM-Solver++(2M) with r=8 score evaluations. Panels, left to right: exact samples, Trig, Rescaled trig, Lipschitz-guided, Elliptic.
Model
Use
Samples
Resolution
Implementation
Reference statistics
SiT-XL/2
Figures 7 , 25 and 27
10,000
256×256
pytorch-fid
ADM ImageNet 256
Table 1
50,000
256×256
ADM (TensorFlow)
ADM ImageNet 256
SD 3.5M
Figures 26 and 28
5,000
512×512
pytorch-fid
MS-COCO 2014 val
Table 2
30,000
512×512
pytorch-fid
MS-COCO 2014 val
Appendix
Table 5: FID protocols. ADM is the TensorFlow evaluator of Dhariwal and Nichol (2021) , and pytorch-fid is the implementation of Seitzer (2020) .
Figure 24: Eigenvalue spectrum of the SiT-XL/2 training-data covariance. Histogram of the eigenvalues λ of the empirical covariance Σ^ of the latents on which SiT-XL/2 is trained, using 100 bins of equal width in log10λ . We estimate Σ^ (unbiased) from N=50,000 ImageNet 256×256 training images passed through the exact SiT-XL/2 training pipeline, so each sample is identical in distribution to what the network sees during training. Dashed lines mark 1/L^=0.0566 and 1/ℓ^=10.57 , which are the 0.5th and 99.5th percentiles of the spectrum, respectively.
Figure 25: FID-10K of SiT-XL/2 over the (λˉ,κ) grid of the target-rescaled elliptic schedule. Sampled with DPM-Solver++(2M). Panels show 4 , 8 , and 32 oracle calls of SiT-XL/2, from top to bottom. Each cell is annotated with its FID-10K (lower is better); colors use a single logarithmic scale shared across panels. In each panel, the solid box marks the best target-rescaled elliptic configuration, and the dashed box marks the best target-rescaled trigonometric ( κ=1 ) configuration. The white diamond represents the heuristic (λˉ^,κ^) configuration.
Figure 26: FID-5K of SD 3.5M over the (λˉ,κ) grid of the target-rescaled elliptic schedule. Sampled with DPM-Solver++(2M). Panels show 4 , 8 , and 16 oracle calls of SD 3.5M, from top to bottom. Each cell is annotated with its FID-5K (lower is better); colors use a single logarithmic scale shared across panels. In each panel, the solid box marks the best target-rescaled elliptic configuration, and the dashed box marks the best target-rescaled trigonometric ( κ=1 ) configuration.
Figure 27: FID-10K sweep over the centering parameter λˉ of various schedules at different number of oracle calls of SiT-XL/2. The elliptic schedule uses the best κ configuration from the sweep in Figure 25 .
Figure 28: FID-5K sweep over the centering parameter λˉ of various schedules at different number of oracle calls of SD 3.5M. The elliptic schedule uses the best κ configuration from the sweep in Figure 26 .
Figure 29: Side-by-side comparison of schedules with DPM-Solver++ and 4 oracle calls of SD 3.5M.
Figure 30: Side-by-side comparison of schedules with DPM-Solver++ and 8 oracle calls of SD 3.5M.
Figure 31: Side-by-Side Comparison of Schedules with 8 oracle calls of SiT-XL/2. Images are sampled with DPM-Solver++(2M) from the same initial noise and class label in every panel. The left column has unscaled schedules ( λˉ=1 ), and the right has rescaled schedules. The hyperparameters that have an asterisks are the best from our sweep (Table 1 ).
Diffusion model sampling studies how to generate high-quality samples efficiently by designing reverse processes, solvers, and inference-time controls.