Noise-unconditioned denoising learns a single score, so generation can be viewed as gradient ascent on a fixed log-density landscape. This optimization interpretation leaves two questions: do trajectories approach meaningful data, and do different initializations retain diverse outcomes? We establish both phenomena for empirical and population multi-scale Gaussian-mixture objectives. Under explicit posterior and landscape conditions, ascent converges almost surely within O(mσmin2/τ) of an m-dimensional reference manifold with reach at least τ. We derive a source--sink identity for the same fixed field and use it to prove finite-resolution regional coverage. In an affine construction, we further identify a positive-bandwidth regime where even empirical-score optimization can avoid training-sample memorization while preserving diversity at the training resolution. The same fixed objective naturally supports penalty-guided controllable generation. Experiments support our predicted mechanisms, show improved robustness under large-step guidance, and demonstrate downstream gains consistent with finite-data generative generalization.
Figures & tables
Figure 1: Numerical verification for E1–E4. (a) Certified first-entry mass per region for two population-sphere grids. (b) Tube entry and twelve-region TV for learned sphere scores, counting failures in a separate bin (mean ± SD across three seeds). The dashed line gives exact-score TV. (c) Fraction of affine outputs at least 1/12 from the training set. (d) Maximum sampled localization ratios and their common uniform bound on the affine grids.
Table 2
Figure 2: Guided ODE sampling at two step budgets. The top and bottom rows show noise-conditioned and noise-unconditioned samples for four-disk and CIFAR-10 horse guidance. The left and right panels use 1,000 and 100 steps, respectively.
Main steps
Selected λ0
Success (%) ↑
Geometric error ↓
Target entropy ↑
50
0.01/0.025
62.2/100.0
1.15/0.00
0.95/0.87
100
0.025/0.025
73.7/96.1
0.57/0.02
0.96/0.96
250
0.025/0.025
90.4/97.9
0.16/0.02
0.95/0.97
500
0.025/0.025
88.0/98.2
0.15/0.01
0.94/0.97
1,000
0.025/0.025
87.5/97.1
0.16/0.02
0.95/0.98
Table 3: Four-disk guidance across step budgets. Entries give C / U. Success is the fraction in the target disk. Geometric error is the mean planar distance to the four-disk union. Target entropy is computed over target hits in 32 equal-area cells and divided by log32 .
Training condition
Cat
Dog
Bal. Acc.
Biased real
11.80 ± 1.25
96.73 ± 0.21
54.27 ± 0.56
Repeated cats (1000)
13.63 ± 0.76
97.13 ± 0.46
55.38 ± 0.41
Diffusion aug.
26.17 ± 1.65
93.33 ± 0.47
59.75 ± 0.95
Unconditioning aug.
38.80 ± 3.45
91.87 ± 0.95
65.33 ± 2.03
Balanced real
79.17 ± 0.76
79.63 ± 1.40
79.40 ± 1.01
Table 4: Downstream classification under minority-class scarcity. Test accuracy (%, mean ± std).
Appendix figures & tables23 assets
Supplementary material from the paper’s appendix.
Appendix
Noise levels
Poisson error (%)
Initialization error (%)
Boundary erosion (%)
Placement loss (%)
Per-region lower (%)
1,024
3.517
0.305
1.051
0.729
2.73
16,384
0.255
0.009
0.017
0.009
8.04
Appendix
Table 5: Continuous first-entry coverage on a uniform sphere at twelve-region resolution. Losses are rounded up and lower bounds down. All probability columns are percentages of total mass. The reference mass is 8.33% per region.
Density
αP
binding bin
split
p99
p99.9
vMF, κ=1
0.789
[0.93,1.00]
—
0.039
0.052
vMF, κ=4
0.799
[0.98,1.00]
—
0.058
0.106
vMF, κ=16
0.852
[0.989,0.995]
—
0.046
0.085
clusters, p=0.5
0.987
binding hemisphere
0.503
0.056
0.109
clusters, p=0.9
0.937
heavier-component hemisphere
0.846
0.059
0.112
Appendix
Table 6: E1: finite-resolution probability and retention across target densities ( M=104 , 105 trajectories; every trajectory reached both stopping surfaces; δstoppop=0.119 ). αP is the minimum, over twelve equal- QM -mass colatitude bins, of the marginal two-sided 95% Wilson lower endpoint for the empirical bin mass divided by its target mass; the binding bin is reported in t=cosθ . For the cluster mixtures, the partition consists of the two component hemispheres, and the measured mass split is also reported. The Wilson intervals are conditional on the realized center cloud and are not a simultaneous confidence band.
Figure 3: E1 ablation ( κ=4 , three center seeds per M ): retention quantile p99 and the measured 99% nearest-center coverage radius εMloc versus M , with an M−1/2 guide and the idealized population benchmark δstoppop . The displayed affine relation p99≈0.037+0.31εMloc is a descriptive fit over the three tested sample sizes.
Figure 4: E1: binned density ratio qtr/qM versus cosθ on equal- QM -mass bins, each plotted at the cosθ value splitting its target mass in half, for vMF targets with κ∈{1,4,16} . The ratio is smallest in or near the high-density pole region and exceeds one in the tails, a pattern consistent with mild flattening by the broad scales.
Updates
Relative score RMSE
Tube entry (%)
Regional TV
1k
0.684±0.007
7.9±3.1
0.92±0.03
5k
0.358±0.004
38.3±2.8
0.62±0.03
20k
0.256±0.005
99.9±0.2
0.07±0.03
Appendix
Table 7: Ordinary DSM on the finite sphere mixtures. Values are mean ± sample standard deviation across three cloud and training seeds. Exact-score regional TV is 0.08±0.03 .
Figure 5: Population off-tube radial score. Radial score g(ϱ) on [10−3,R−r0] for m∈{1,2,4} and c∈{1,2,3,4,8,16} . The finite scale sum is combined with Gauss–Legendre quadrature in colatitude. Dashed lines show the post-hoc small- ϱ reference k(m,c)ϱ , where k=(c−1)(c+m−2)/(m+1) . The score is positive at every tested grid node.
Figure 6: Critical structure of empirical sphere mixtures. M=2000 centers lie on S2(1)⊂R12 . (a) Tube distance of all located deduplicated critical points across the σmin sweep; the only located off-tube point is the near-center saddle of the uniform cloud. (b) Median magnitude of the most negative Hessian eigenvalue for in-tube critical points, compared with the 1/σmin2 scale. (c) Hessian spectrum of the near-center uniform-cloud saddle, with three positive and nine negative eigenvalues.
Mean distance
Distance ≥1/12 (%)
σmin
Flow
Cutoff
Flow
Cutoff
0.50
0.1651
0.1321
100.00
64.78
0.55
0.1586
0.1342
91.28
64.88
0.60
0.1663
0.1360
100.00
65.85
Appendix
Table 8: Finite-grid E3. Distances are Euclidean distances to the nearest training point. Cutoff denotes the exact cutoff marginal followed by the Tweedie readout.
s
psep
core mass
ϵcell
npop
q0.05
updates
0.05
0.0000
0.0000
1.0000
0
0.00250
101
0.08
0.0000
0.0000
1.0000
0
0.01096
151
0.10
0.0000
0.0000
1.0000
0
0.02388
151
0.11
0.0000
0.0000
1.0000
0
0.03499
251
0.12
0.0000
0.0000
1.0000
0
0.05004
301
0.13
0.5875
0.5875
0.4779
22
0.06941
751
Appendix
Table 9: E3 bandwidth sweep. Each row uses 8192 numerical outputs with η=0.2 . The separation threshold is 1/12 ; core mass is 1−pout ; npop counts the 41 cores whose estimated mass is at least 1/82 ; q0.05 is the empirical fifth-percentile distance to the training set. All rows meet the recorded residual threshold 10−5 . Statistics are estimates for the simulated outputs, not confidence bounds for the exact flow.
Figure 7: Folded numerical outputs across the bandwidth sweep (E3). The 150 equal-width histogram bins cover [−1/2,1/2) ; training locations, cell centers, and output-core boundaries are marked. All full-grid outputs are pooled. This plot visualizes within-cell displacement only; diversity across cells is quantified by ( F.21 ). The finite-array histograms need not consist of a single exact central atom.
s
η
psep
ϵcell
q0.05
updates
max. residual
0.14
0.2
0.9780
0.1325
0.09861
2451
9.72×10−6
0.14
0.1
0.9764
0.1268
0.08411
4901
9.78×10−6
0.18
0.2
0.9734
0.1060
0.12276
251
3.81×10−6
0.18
0.1
0.9700
0.1039
0.12278
451
9.54×10−6
0.24
0.2
0.9679
0.0901
0.09919
301
3.81×10−6
0.24
0.1
0.9636
0.0844
0.09914
501
9.54×10−6
Appendix
Table 10: E3 paired step-size comparison. The initializations are matched within each bandwidth. The last column is the maximum normalized residual at the final recorded check. The s=0.14 lower-tail margin remains sensitive despite similar residuals, whereas the fifth percentiles at s=0.18 and 0.24 change by less than 4.4×10−5 .
Figure 8: Step-size sensitivity of the E3 numerical outputs. The two step sizes use matched initializations at each bandwidth and the same stopping diagnostic. Separation and core-label TV change modestly at all three settings; the lower-tail distance at s=0.14 is more sensitive than at s=0.18 and 0.24 . This is a finite-step comparison, not an extrapolation to zero step size.
R
fitted slope
B(0.01)
B(0.06)
error at 0.01 (%)
0.5
2.0091
1.2002404
1.2215817
0.04902
1.0
2.0022
1.1997992
1.2049867
0.01225
2.0
2.0006
1.1996890
1.2009770
0.00306
Appendix
Table 11: E4 population bias scaling. Slopes use all six tested lower cutoffs. Here B(s)=Rb⋆/s2 at fixed R and s=σmin . The last column reports the relative discrepancy between B(0.01) and Cfs=1.199652317… .
Figure 9: Population spherical bias with a fixed relative noise schedule (E4). Left: the located near-sphere bias for three sphere radii, with a quadratic guide. Right: the normalized bias approaches Cfs=1.199652317… , computed from the 64-scale schedule rather than fitted to the numerical outputs. The quadratic guide on the left is anchored to one numerical point and illustrates slope only; the horizontal line on the right is the analytic finite-schedule local reference. All reported biases come from population quadrature and radial root finding.
r
Cflat,N(r)
median
99th percentile
sampled maximum
0.039528
1.144095
1.162255
1.5547
1.7547
0.117725
1.289768
1.317887
1.7915
2.1698
0.158114
1.533454
1.558910
2.1233
2.3701
0.219636
1.262671
1.278057
1.6759
1.8251
0.300000
1.250007
1.252572
1.6455
1.9018
Appendix
Table 12: E4 localization at selected radii. All entries use R=1 , one fixed cloud of 104 centers, and 512 queries per radius. The row r=0.158114 corresponds to rmin before rounding. Aggregate discrepancies reported in the text use all 15 radii, not only these five rows.
Figure 10: Sampled posterior-localization diagnostic (E4). At each of 15 radii, 512 tube queries are evaluated against one fixed cloud of 104 uniform spherical centers. The median closely follows the radius-dependent flat reference for the same 24-scale schedule, while the higher quantiles and sampled maximum show the variation across query locations.
Seed
Continuous lower (%)
Euler lower (%)
Observed minimum (%)
0
7.21
7.17
7.98
1
7.28
7.24
7.92
2
7.18
7.14
7.94
Appendix
Table 13: Per-region mass for the same three learned circle fields and first-entry rule. Continuous and Euler lower bounds include prior placement and neural drift. The Euler column also includes the interpolation defect. The observed column gives the smallest empirical region mass in the separate RK4 runs.
Figure 11: Unguided 100-step conditioning baseline. This ablation isolates the failure in Figure 2 to the interaction between coarse steps and guidance rather than the reduced step count alone.
Main steps
λ0=0.1
λ0=0.5
λ0=1
50
1.32
15.71
32.59
1,000
0.01
0.10
0.52
Appendix
Table 14: Percentage increase in C’s mean noise-scaled normal-score RMSE at the guided proposal, relative to its paired unguided proposal.
Training condition
Cat
Dog
Bal. Acc.
Biased real
11.80 ± 1.25
96.73 ± 0.21
54.27 ± 0.56
Balanced real reference
79.17 ± 0.76
79.63 ± 1.40
79.40 ± 1.01
Repeated cats (1000)
13.63 ± 0.76
97.13 ± 0.46
55.38 ± 0.41
Pretrained diffusion aug.
77.53 ± 1.34
82.07 ± 1.36
79.80 ± 1.35
Pretrained unconditioning aug.
77.00 ± 2.62
82.33 ± 2.01
79.67 ± 2.28
100-cat diffusion aug.
26.17 ± 1.65
93.33 ± 0.47
59.75 ± 0.95
Appendix
Table 15: Full downstream augmentation results. Test accuracy (%, mean ± std). The two 100-cat generator rows are the data-scarce comparison emphasized in Section 4.4 .
Training condition
Overall
Cat
Other-9
Biased real
81.49 ± 1.02
22.40 ± 2.25
88.06 ± 0.95
Balanced real reference
86.00 ± 1.01
71.30 ± 2.31
87.63 ± 0.90
Repeated cats (1000)
79.90 ± 1.43
14.73 ± 1.04
87.14 ± 1.53
Pretrained diffusion aug.
86.17 ± 0.93
69.40 ± 2.26
88.03 ± 0.83
Pretrained unconditioning aug.
84.77 ± 1.05
66.87 ± 1.72
86.76 ± 1.27
100-cat diffusion aug.
81.59 ± 0.29
29.50 ± 0.53
87.37 ± 0.31
Appendix
Table 16: Ten-class CIFAR-10 classification with a scarce cat class. Accuracy (%, mean ± std over three classifier seeds). “Other-9” is the mean per-class accuracy over the nine non-cat classes.
Training condition
Subset 1
Subset 2
Subset 3
Biased real
54.48±0.28
54.69±0.03
53.28±0.66
Repeated cats
55.29±0.36
55.20±0.41
55.13±0.62
Diffusion augmentation
59.21±1.07
59.98±0.71
60.31±0.45
Unconditioning augmentation
64.34±0.68
65.05±0.14
65.95±0.51
Appendix
Table 17: Balanced accuracy (mean ± sample SD, in percent) across three independently drawn 100-cat subsets.
Method
NN-ratio proxy (%) ↓
PRDC coverage (%) ↑
Diffusion
47.1
72.2
Unconditioning
23.6
89.1
Appendix
Table 18: Training-image proximity and reference coverage on the same 1,000 generated cats per method.