We address the problem of accurate, training-free guidance for conditional generation in trained diffusion models. Existing methods typically rely on point-estimates to approximate the posterior score, often resulting in biased approximations that fail to capture multimodality inherent to the reverse process of diffusion models. We propose a sequential Monte Carlo (SMC) framework that integrates over the full denoising distribution to construct an unbiased Monte Carlo estimator of pθ(y∣xt), avoiding the bias of existing training-free guidance methods. To ensure computational tractability, we incorporate variance-reduction schemes based on MultiLevel Monte Carlo (MLMC). We evaluate on three image generation tasks. On CIFAR-10 and ImageNet classifier guidance, our approach improves accuracy by 18 and 38 points respectively over the strongest training-free baseline. On CelebA-HQ joint-attribute guidance, three existing methods (DPS, LGD, TFG) used within our framework all outperform the closest SMC-based comparator (TDS) by 25-37 points, and improve over the strongest baseline by 2-14 points.
Figures & tables
Figure 1: Posterior p(x0∣xt) for an 8-component ring GMM. Panels show high noise ( t=900 , left) and moderate noise ( t=600 , right).
Domain
Task
tstart
β^
ImageNet 256×256
class guidance
68
0.566
Stable Diffusion v1.5
style transfer
68
0.810
Stable Diffusion v1.5
style transfer
76
0.858
CelebA-HQ 256×256
age+gender
20
1.234
CelebA-HQ 256×256
age+gender
52
1.157
Table 1: Empirical values of β^ obtained for MLMC across three conditional-generation problems. β^ is estimated from the slope of logVℓ vs level ℓ over a dyadic stride ladder [8,4,2,1] ( L=3 ).
Dataset
Method
Accuracy
FID
Cost/Success (s)
CIFAR-10
TFG-1
52.0%
91.7
68.5
TFG-4
77.1%
73.9
156.3
SMC-MLMC-DM
95.6%
46.3
23.4
ImageNet
TFG-1
20.2%
231.9
230.0
SMC-MLMC-DM
58.3%
182.3
141.7
Table 2: Classifier guidance results on CIFAR-10 (10 classes) and ImageNet 256×256 (6 classes: 111, 130, 207, 222, 333, 444). Cost/Success: estimated time to generate one valid sample with probability ≥0.95 .
Class
TFG-1
K
TFG best-of- K
SMC-MLMC-DM ( N=4 )
Margin
333 (Hamster)
0.080
8
0.474±0.100
0.487±0.033
+0.013
444 (Tandem Bicycle)
0.102
4
0.351±0.060
0.476±0.033
+0.126
222 (Kuvasz)
0.129
8
0.669±0.083
0.630±0.048
−0.039
130 (Flamingo)
0.200
6
0.737±0.077
0.570±0.033
−0.167
207 (Golden Retriever)
0.280
6
0.875±0.058
0.761±0.028
−0.114
111 (Nematode)
0.406
6
0.957±0.031
0.756±0.048
−0.201
Table 3: TFG-1 best-of-K versus SMC-MLMC-DM (N=4) on ImageNet classifier guidance tasks with matched wall-clock time. Margin is SMC-MLMC-DM success rate minus TFG-1.
Method
Accuracy
Age
Gender
Cost/Success (s)
TFG (Baseline)
0.793
0.784
0.802
6.5
TDS
0.564
0.548
0.581
375.9
SMC-MLMC-DM (TFG-proposal)
0.930
0.921
0.939
88.3
SMC-MLMC-DM (DPS-proposal)
0.857
0.844
0.871
115.1
SMC-MLMC-DM (LGD-proposal)
0.811
0.707
0.914
131.3
Table 4: Performance on CelebA-HQ joint-attribute guidance, averaged across four targets ( {young,old}×{male,female} ). Per-image accuracy is the mean of gender and age accuracies by held-out evaluation classifiers; per-axis columns report marginal accuracy on each classifier. Cost/Success: estimated time to generate one valid sample with probability ≥0.95 .
Method
Twist scale α
Unique ancestors
Distinct images
Distinct- valid
Accuracy
SMC-MLMC-DM
–
6.1
9.0
8.5
0.94
TDS
0.125
5.6
7.2
1.2
0.35
TDS
0.25
5.6
7.6
2.5
0.50
TDS
0.5
4.1
6.0
1.7
0.29
TDS
1
2.3
5.3
1.1
0.26
TDS
2
1.2
3.2
0.4
0.20
Table 5: Diversity study
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
ρˉ
Per-image acc.
Age
Gender
Validity
1.0 (paper default)
0.440
0.022
0.857
1.000
0.5
0.571
0.143
1.000
0.571
0.25
0.509
0.393
0.625
0.571
Appendix
Table 6: TDS guidance-scale sweep on target old+female (1+0). Per-image accuracy is computed over 7 batches × 32 particles. Validity denotes the fraction of batches producing at least one valid sample. Lower validity indicates particle collapse.
Figure 2: Six independent runs of label guidance for CIFAR-10. (a) Without resampling, ESS begins to collapse at t=60. (b) Resampling is performed at {70,60,50,40,30} . The relevant window remains stable across runs, suggesting it can be identified with a single exploratory run.
Allocation
Mean Δ
Mean ∣Δ∣
Outside [0,1]
(5,2,1)
+0.00330
0.02988
30/256 ( 11.7% )
(10,4,2)
+0.00555
0.01460
27/256 ( 10.5% )
(20,8,4)
+0.00481
0.00527
13/256 ( 5.1% )
Appendix
Table 7: Clipping adjustments for three MLMC sample allocations.
Class
TFG-1
K
TFG best-of- K
SMC-MLMC-DM ( N=4 )
Margin
333 (Hamster)
0.080
7
0.434±0.094
0.487±0.033
+0.053
444 (Tandem Bicycle)
0.102
4
0.351±0.060
0.476±0.033
+0.126
222 (Kuvasz)
0.129
7
0.628±0.081
0.630±0.048
+0.002
130 (Flamingo)
0.200
5
0.679±0.074
0.570±0.033
−0.109
207 (Golden Retriever)
0.280
5
0.826±0.060
0.761±0.028
−0.065
111 (Nematode)
0.406
5
0.923±0.037
0.756±0.048
−0.167
Appendix
Table 8: TFG-1 best-of-K versus SMC-MLMC-DM (N=4) on ImageNet classifier guidance tasks with matched VRAM. Margin is SMC-MLMC-DM success rate minus TFG-1.
Figure 3: Computation cost versus accuracy of SMC-MLMC-DM as a function of number of particles N for CIFAR-10.
Target Class
Top-1 Accuracy
FID
Airplane
94.9%
54.8
Automobile
97.6%
35.0
Bird
94.3%
51.0
Cat
92.7%
61.1
Deer
96.1%
39.0
Dog
91.3%
60.2
Appendix
Table 9: Per-class breakdown of SMC-MLMC-DM performance on CIFAR-10 conditional generation.
Figure 4: Qualitative comparison of CIFAR10 automobiles generated by (a) SMC-MLMC-DM and (b) TFG-4 [ Ye et al., 2024 ] .
N
nruns
Accuracy
SEM
4
30
53.3%
9.3
8
30
80.0%
7.4
16
30
96.7%
3.3
32
30
100.0%
0.0
Appendix
Table 10: Particle-count sensitivity on ImageNet (classes 111, 333, 444; 10 seeds per cell, 30 runs per N ). Standard error of the mean aggregates across seed and class.
Figure 5: Qualitative comparison of ImageNet golden retrievers (class 207) by (a) SMC-MLMC-DM and (b) TFG-1 [ Ye et al., 2024 ] .
Target Class
Method
Success Rate
FID ↓
Runtime (s)
Cost/Success (s)
111 (Nematode)
TFG (Baseline)
41.0%
199.5
12
72
SMC-MLMC-DM
71.0%
204.6
34
102
130 (Flamingo)
TFG (Baseline)
20.0%
243.0
12
168
SMC-MLMC-DM
56.5%
167.9
34
136
207 (Golden Retriever)
TFG (Baseline)
28.0%
207.3
12
120
SMC-MLMC-DM
74.5%
143.4
34
102
Appendix
Table 11: Per-class SMC-MLMC-DM performance on ImageNet compared to TFG Ye et al. [2024] .
Method
Validity
Age
Gender
TFG (Baseline)
1.000
1.000
1.000
TDS
0.575
0.563
0.588
SMC-MLMC-DM (TFG-proposal)
1.000
1.000
1.000
SMC-MLMC-DM (DPS-proposal)
1.000
1.000
1.000
SMC-MLMC-DM (LGD-proposal)
0.950
0.900
1.000
Appendix
Table 12: Validity (proportion of batches that have at least one valid sample) across age and gender in CelebA-HQ joint attribute guidance (N=16).
Target
Baseline
TDS
TFG-prop
DPS-prop
LGD-prop
young+female (0+0)
0.939
0.641
0.936
0.928
0.920
young+male (0+1)
0.892
0.722
0.998
0.977
0.956
old+female (1+0)
0.753
0.470
0.834
0.603
0.530
old+male (1+1)
0.589
0.425
0.952
0.922
0.836
Mean
0.793
0.564
0.930
0.857
0.811
Appendix
Table 13: Per-image accuracy across four CelebA-HQ joint-attribute targets ( N=16 ).
Target
Baseline
TDS
TFG-prop
DPS-prop
LGD-prop
young+female (0+0)
0.981
0.856
0.981
0.963
0.991
young+male (0+1)
0.925
0.844
0.997
0.959
0.972
old+female (1+0)
0.738
0.141
0.803
0.609
0.188
old+male (1+1)
0.494
0.350
0.903
0.844
0.678
Mean
0.784
0.548
0.921
0.844
0.707
Appendix
Table 14: Age-axis accuracy across four CelebA-HQ joint-attribute targets ( N=16 ).
Target
Baseline
TDS
TFG-prop
DPS-prop
LGD-prop
young+female (0+0)
0.897
0.425
0.891
0.894
0.850
young+male (0+1)
0.859
0.600
1.000
0.994
0.941
old+female (1+0)
0.769
0.800
0.866
0.597
0.872
old+male (1+1)
0.684
0.500
1.000
1.000
0.994
Mean
0.802
0.581
0.939
0.871
0.914
Appendix
Table 15: Gender-axis accuracy across four CelebA-HQ joint-attribute targets ( N=16 ).
Method
Accuracy
Age
Gender
Cost/Success (s)
TFG (Baseline)
0.791
0.780
0.801
6.5
TDS
0.643
0.607
0.678
319.2
SMC-MLMC-DM (TFG-proposal)
0.902
0.874
0.931
161.4
SMC-MLMC-DM (DPS-proposal)
0.875
0.886
0.863
212.0
SMC-MLMC-DM (LGD-proposal)
0.819
0.821
0.816
215.7
Appendix
Table 16: CelebA-HQ joint-attribute guidance at N=32 , averaged across four targets. Per-axis columns report marginal accuracy on each classifier. Cost/Success: estimated time to generate one valid sample with probability ≥0.95 .
Target
Baseline
TDS
TFG-prop
DPS-prop
LGD-prop
young+female (0+0)
0.942
0.853
0.938
0.915
0.833
young+male (0+1)
0.895
0.708
0.987
0.989
0.973
old+female (1+0)
0.754
0.440
0.797
0.663
0.580
old+male (1+1)
0.571
0.571
0.888
0.931
0.888
Mean
0.791
0.643
0.902
0.875
0.819
Appendix
Table 17: Per-image accuracy across four CelebA-HQ joint-attribute targets.
Target
Baseline
TDS
TFG-prop
DPS-prop
LGD-prop
young+female (0+0)
0.969
0.991
0.973
0.973
0.987
young+male (0+1)
0.946
0.844
0.973
0.982
0.987
old+female (1+0)
0.737
0.022
0.772
0.705
0.536
old+male (1+1)
0.469
0.571
0.777
0.884
0.777
Mean
0.780
0.607
0.874
0.886
0.821
Appendix
Table 18: Age-axis accuracy across four CelebA-HQ joint-attribute targets, evaluated using the held-out age classifier.
Target
Baseline
TDS
TFG-prop
DPS-prop
LGD-prop
young+female (0+0)
0.915
0.714
0.902
0.857
0.679
young+male (0+1)
0.844
0.571
1.000
0.996
0.960
old+female (1+0)
0.772
0.857
0.821
0.621
0.625
old+male (1+1)
0.674
0.571
1.000
0.978
1.000
Mean
0.801
0.678
0.931
0.863
0.816
Appendix
Table 19: Gender-axis accuracy across four CelebA-HQ joint-attribute targets, evaluated using the held-out gender classifier.
tstart
remaining
base
β^
V0
V1
V2
V3
12
88
11
−0.331
0.0321
0.0577
0.0701
0.0914
20
80
10
−0.810
0.0122
0.0200
0.0377
0.0615
36
64
8
+0.337
0.0086
0.0173
0.0144
0.0108
52
48
6
+0.505
0.0137
0.0065
0.0045
0.0032
68
32
4
+0.810
0.0066
0.0029
0.0018
0.0009
76
24
3
+0.858
0.0038
0.0023
0.0019
0.0007
Appendix
Table 20: Full β^(t) calibration on Stable Diffusion v1.5 style transfer. Six intervals across the denoising schedule, 50 trials per level.
tstart
rem.
base
β^
V0
V1
V2
V3
ρ1
ρ2
ρ3
20
80
10
+1.234
12.205
5.685
1.952
1.027
0.481
0.148
0.085
36
64
8
+0.925
3.204
4.529
2.867
1.256
0.559
0.311
0.113
52
48
6
+1.157
0.0124
0.0076
0.0196
0.00154
0.470
0.176
0.039
68
32
4
+0.883
0.0246
0.0362
0.0236
0.01064
0.444
0.145
0.061
84
16
2
+0.701
0.0600
0.0741
0.0355
0.02803
0.493
0.199
0.157
Appendix
Table 21: Full β^(t) calibration on CelebA-HQ DDPM with joint age+gender guidance.
Stochastic differential equation (SDE)-based generative models have achieved substantial progress in conditional generation via training-free differentiable loss-guided approaches. However, existing methodologies utilizing posterior sam- pling typically confront a substantial estimation error, which results in inaccu- rate gradients for guidance and leading to inconsistent generation results. To mitigate this issue, we propose that performing an additional backward denois- ing step and Monte-Carlo sampling (ABMS) can achieve better guided diffu- sion, which is a plug-and-play adjustment strategy. To verify the effectiveness of our method, we provide theoretical analysis and propose the adoption of a dual-focus evaluation framework, which further serves to highlight the critical problem of cross-condition interference prevalent in existing approaches. We conduct experiments across various task settings and data types, mainly includ- ing conditional online handwritten trajectory generation, image inverse problems (inpainting, super resolution and gaussian deblurring) molecular inverse design and so on. Experimental results demonstrate that our approach can be effec- tively used with higher order samplers and consistently improves the quality of generation samples across all the different scenarios.
Minsi Ren, Wenhao Deng, Ruiqi Feng +1
AI for Scientific Simulation and Discovery Lab, Westlake University
Training-free conditional diffusion provides a flexible alternative to task-specific conditional model training, but existing samplers often allocate computation inefficiently: independent guided trajectories can vary widely in quality, and additional function evaluations along a single trajectory may not recover from poor early decisions. We propose Tempered Guided Diffusion (TGD), an annealed sequential Monte Carlo framework for training-free conditional sampling with diffusion priors. TGD targets tempered posterior distributions over the clean signal, using noisy diffusion states only as auxiliary variables for proposing reconstructions and propagating particles. Particles are reweighted by incremental likelihood ratios, resampled, and propagated across noise levels, concentrating computation on trajectories plausible under both the prior and observation. Under idealized exact-reconstruction assumptions, full TGD yields a consistent particle approximation to the posterior as the number of particles grows. For expensive reconstruction tasks, Accelerated TGD (A-TGD) retains early particle exploration but prunes to a single high-likelihood trajectory partway through sampling. Experiments on a controlled two-dimensional inverse problem and image inverse problems show improved posterior approximation and favorable wall-clock speed-quality tradeoffs over independent multi-trajectory baselines.
Andreas Makris, Paul Fearnhead, Chris Nemeth
Department of Mathematics and Statistics Lancaster University, UK · Department of Mathematics and Statistics2026 Lancaster University, UK
Diffusion models achieve strong generative performance but often rely on classifier-free guidance (CFG), an inference-time intervention that modifies the sampling trajectory. Since standard denoising score matching (DSM) should recover the target distribution when the score is learned accurately, the practical necessity of CFG suggests deficiencies in learned conditional scores. We identify weak inter-class separation as one such limitation: samples generated from the same noise under different labels can retain similar class-agnostic structure. To address this deficiency, we introduce MCLR, a principled fine-tuning objective that explicitly maximizes inter-class likelihood-ratios, thereby strengthening class-specific structures under standard sampling. Across diverse models and datasets, MCLR substantially improves guidance-free conditional generation and narrows the gap to CFG. Theoretically, we prove that the CFG-guided score is the unique minimizer of a sample-adaptively weighted, ELBO-approximated MCLR objective, interpreting CFG as an implicit inference-time contrastive alignment procedure. We further compare MCLR with related contrastive objectives, including DPO, CCA, and CFM, clarifying their connections and key differences in an otherwise fragmented literature.
Xiang Li, Yixuan Jia, Xiao Li +3
University of Michigan · Michigan State University