acceptodds
Under review as a conference paper at ICLR 2027

Improving Tail-Class Diversity in Diffusion Models via Training-Free SNR-Adaptive Block Routing

Abstract

Although diffusion models have achieved remarkable performance in image generation, their performance degrades substantially on class-imbalanced datasets that commonly arise in real-world applications, particularly for tail classes, where both generation quality and intra-class diversity deteriorate and the generated distribution tends to collapse toward a limited set of dominant modes. In this work, we study how to improve the coverage of real intra-class modes for tail classes in a class-conditional U-Net diffusion model that has already been trained on long-tailed data, while keeping all model parameters frozen. Through empirical evaluation and internal mechanism analysis, we find that lower training frequency is associated with stronger local class-conditional pressure and narrower representational coverage. Local interventions further show that the effects of class conditioning on local denoising and final generation coverage are not fully aligned, and that appropriately modulating conditioning in the low-resolution encoding region can improve tail-class coverage. Building on this observation, we propose SNR-Adaptive Block Routing (SABR), a training-free sampling strategy that adaptively modulates the strength of local class-conditioning injection according to the model’s internal conditional response and the signal-to-noise ratio (SNR), without requiring any additional training or fine-tuning. Extensive experiments across different long-tailed data distributions and generative baselines demonstrate that SABR generalizes well across settings, improves generation quality in most cases, and effectively mitigates mode concentration for tail classes while increasing coverage of real intra-class modes.

Primary area:
generative models
Keywords:
diffusion models, long-tailed generation, training-free sampling, class-conditional generation, adaptive block routing

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_bkMFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SABR, a training-free sampling method for improving tail-class diversity in diffusion models. The authors identify stronger local conditional responses in tail classes and introduce pressure-guided block selection with SNR-dependent conditioning modulation. Experiments demonstrate improved tail-class generation quality and mode coverage across several benchmarks.

Critical strengths
  1. The analysis of frequency-dependent conditional pressure provides an interesting perspective on the concentration of tail-class patterns.
  2. SABR improves the generation of tail classes without retraining the diffusion model.
  3. The experiments cover multiple datasets and baseline methods.
Critical weaknesses
  1. Novelty in methodology is moderate. SABR mostly combines existing conditional adjustment ideas, while introducing heuristic pressure thresholds and a gate control mechanism based on signal-to-noise ratio.
  2. Component-level verification is insufficient. The independent contributions of signal-to-noise ratio adaptation, pressure budget, and routing decisions have not been clearly separated.
  3. Generalization evidence is limited. The main pattern coverage analysis focuses on CIFAR-100-LT, and its applicability to modern diffusion Transformer architectures remains unclear.
Questions for authors

Please see the weakness part.

2 comments

Thank you for your careful review, your positive comments on our method, and your suggestions on novelty, component analysis, and generalization. We have added experiments and analysis in response. Below, we address the three main weaknesses, which also cover the items referred to in the questions-for-authors section.

Responses to the main weaknesses

Weakness 1 Method novelty and design rationale

SABR builds on the broader idea of adjusting conditioning. Its specific design follows a sequence of measurements and intervention tests aimed at understanding tail-class generation.

We first observed limited mode coverage and stronger concentration in tail-class outputs. To investigate this, we measured the direct class-conditional response at a fixed block input and found higher local pressure in tail classes, especially in Down-Late. Frequency reassignment then linked this higher pressure to training frequency, directing our attention to frequency-related local conditional responses. We next removed direct class conditioning in Down-Late. Tail-class denoising MSE rose by 18.31%, while FID and Recall improved and Precision fell. Reducing local conditioning can therefore improve coverage at a cost to fidelity. This led us to design a method that chooses where to intervene, limits the reduction, and controls its timing.

SABR uses pressure statistics from data-rich classes to build a frozen reference. Normalized block pressure determines intervention priority, a cumulative squared-pressure budget limits the adjustment, and the remaining excess determines the gate through an analytical rule. SNR gating controls when intervention is allowed. The budget comparison also gives lower Few FID and higher Recall with a cumulative budget than with a per-step budget, alongside slightly lower Precision.

The contribution is the finding that local conditional responses vary with training frequency and a method that turns this finding into reference-based control over intervention location, strength, and timing. Pressure measurement, frequency reassignment, and local intervention connect the observed tail-class behavior to the design of SABR.

Weakness 2 Contributions of the individual components

Thank you for suggesting a clearer component analysis. We added a 2 × 2 comparison of cumulative versus per-step budgets and γ\gamma = 0 versus γ\gamma = 1. We also tested block-position permutation and intervention in only the first block. These tests examine the roles of budget history, timing, and spatial routing. All settings use the same frozen CIFAR-100-LT DDPM, 50-step DDIM, CFG scale 2.5, paired initial noise, and 500 images per class.

Table 1. Component comparisons for pressure budgets, SNR gating, and block routing on CIFAR-100-LT. Better results are bold.

SettingOverall FID ↓Few FID ↓Few KID ×1,000 ↓Few Precision ↑Few Recall ↑
Baseline14.366831.71826.426674.06%29.09%
Cumulative, γ\gamma = 013.093127.25804.485274.75%32.75%
Cumulative, γ\gamma = 1 (default)12.951626.62334.397174.98%34.35%
Per-step, γ\gamma = 013.164427.51534.563575.49%33.61%
Per-step, γ\gamma = 113.019326.80254.420774.79%33.56%
Shuffled gate assignment13.282627.67244.357775.96%31.86%
Cumulative, first block only13.213327.87414.922374.81%31.88%

Note: Better results are bold.

The three components play different roles. For the pressure budget, cumulative control gives lower FID than per-step control at both γ\gamma settings, showing the value of retaining information from earlier interventions. For SNR gating, γ\gamma = 1 lowers FID under both budget rules. With the cumulative budget, Few Recall rises from 32.75% to 34.35%, showing that intervention timing affects the result. For block routing, permuting the block positions or using only the first block raises Few FID and lowers Recall. Thus, both the selected locations and the distribution of intervention across blocks matter.

Weakness 3 Generalization across datasets and to DiT

Thank you for asking for more generalization evidence. We added tests of real-mode coverage on another dataset and generation performance with another architecture. The original real-mode analysis focused on CIFAR-100-LT; we now add tail-class mode coverage on CIFAR-10-LT. We also compare DiT with DiT (SABR) on both CIFAR-10-LT and CIFAR-100-LT at 100, 200, 300, 400, and 500 epochs. The results below address these two aspects in turn.

The CIFAR-10-LT analysis compares DDPM and DDPM + SABR using four real-mode metrics. Together with the original CIFAR-100-LT analysis, it tests whether the mode-coverage gains carry over to another dataset.

Table 2. Tail-class real-mode coverage on CIFAR-10-LT. Changes are SABR minus DDPM. Better results are bold.

MetricDDPMDDPM+SABRChange
Effective mode ratio ↑0.71440.7367+0.0223
Mode-mass recall ↑0.68210.7062+0.0241
JS divergence ↓0.11510.0966-0.0185
Maximum mode-share ratio ↓2.34622.2043-0.1419

The effective mode ratio and mode-mass recall rise by 0.0223 and 0.0241, while JS divergence and the maximum mode-share ratio fall by 0.0185 and 0.1419. All four metrics change in the same direction as on CIFAR-100-LT in the paper. The DiT results below examine transfer across architectures.

Table 3. DiT and DiT (SABR) on CIFAR-10-LT at different training epochs. Better results are bold.

EpochMethodOverall FID ↓Overall Recall ↑Few FID ↓Few Recall ↑
100DiT55.25531.55%83.90135.47%
100DiT (SABR)55.11331.79%83.84435.75%
200DiT30.50334.66%37.54136.52%
200DiT (SABR)29.80535.21%36.74037.59%
300DiT20.23735.15%23.84430.05%
300DiT (SABR)19.13136.82%23.63231.75%
400DiT17.33234.82%24.86325.39%
400DiT (SABR)16.59638.04%18.78834.15%
500DiT14.64835.97%23.56724.35%
500DiT (SABR)14.11539.43%17.91833.02%

Table 4. DiT and DiT (SABR) on CIFAR-100-LT at different training epochs. Better results are bold.

EpochMethodOverall FID ↓Overall Recall ↑Few FID ↓Few Recall ↑
100DiT43.15131.49%56.00532.52%
100DiT (SABR)43.27031.48%56.20832.53%
200DiT25.70436.31%39.75837.76%
200DiT (SABR)25.48436.44%38.66038.04%
300DiT19.12437.40%31.31539.58%
300DiT (SABR)18.89937.35%31.10339.54%
400DiT16.31836.73%27.78136.35%
400DiT (SABR)16.10937.78%26.20838.25%
500DiT14.26934.87%24.73731.62%
500DiT (SABR)13.60537.04%22.27936.84%

SABR gives clear tail-class gains on both datasets. At 500 epochs, Few FID falls from 23.567 to 17.918 on CIFAR-10-LT, while Recall rises from 24.35% to 33.02%. On CIFAR-100-LT, Few FID falls from 24.737 to 22.279 and Recall rises from 31.62% to 36.84%. Overall FID and Recall also improve on both datasets. Results at the other epochs show gains at most training stages. These findings extend the results from the original U-Net models to the tested DiT models on two long-tailed datasets. We will add these experiments to the revised paper.

Thank you again for your time and constructive feedback.

Reviewer_jK4jOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses insufficient intra-class mode coverage for tail classes in long-tailed diffusion models and proposes a training-free method, SNR-Adaptive Block Routing (SABR). The authors define conditional pressure and find that tail classes exhibit higher conditional pressure, with controlled frequency-reassignment experiments providing further evidence of its association with training frequency. SABR constructs frozen references from fully conditioned trajectories of data-rich classes and adaptively attenuates local class-conditional residuals based on accumulated squared-pressure excess, normalized pressure-based block ranking, class frequency, and SNR. Experiments show that SABR improves generation quality in most evaluated settings and increases the coverage of real intra-class modes for tail classes on CIFAR-100-LT with the DDPM baseline.

Critical strengths
  1. Improving the intra-class diversity of underrepresented classes without retraining pretrained diffusion models is a meaningful problem.

  2. The paper reports performance improvements across multiple datasets and experimental settings, particularly for tail classes. For example, on CIFAR-100-LT, SABR reduces the DDPM Few-class FID from 31.525 to 27.017.

Critical weaknesses
  1. Prior studies have explored training-free conditioning modulation (CADS [1]), U-Net feature reweighting (FreeU [2]), guidance applied within limited sampling intervals [3], and the stage-wise dynamics of classifier-free guidance (CFG) [4]. Although SABR introduces pressure-guided block routing, its substantive novelty over existing sampling-time control methods remains to be further established.

  2. The paper does not establish a clear causal link between elevated conditional pressure and mode concentration. It remains unclear whether high-pressure trajectories within the same class are more prone to mode concentration or benefit more from pressure-based attenuation.

Questions for authors
  1. The experiments primarily use U-Net architectures, with most conducted at relatively low image resolutions. Is SABR also applicable to current diffusion Transformer (DiT) architectures?

  2. Can the authors further clarify the contributions of individual components? In particular, setting the SNR threshold to zero yields similar overall FID (13.038 vs. 12.977), raising questions about the independent benefit of SNR-dependent gating.

[1] Sadat, S., Buhmann, J., Bradley, D., Hilliges, O., & Weber, R. (2024, May). CADS: Unleashing the diversity of diffusion models through condition-annealed sampling. In International Conference on Learning Representations (Vol. 2024, pp. 23723-23755).

[2] Si, C., Huang, Z., Jiang, Y., & Liu, Z. (2024, June). Freeu: Free lunch in diffusion u-net. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 4733-4743). IEEE.

[3] Kynkäänniemi, T., Aittala, M., Karras, T., Laine, S., Aila, T., & Lehtinen, J. (2024). Applying guidance in a limited interval improves sample and distribution quality in diffusion models. Advances in Neural Information Processing Systems, 37, 122458-122483.

[4] Jin, C., Shi, Q., & Gu, Y. (2026, April). Stage-wise dynamics of classifier-free guidance in diffusion models. In International Conference on Learning Representations (Vol. 2026, pp. 154028-154053).

6 comments

Thank you for your careful review and your comments on novelty, the link between conditional pressure and mode concentration, and the roles of individual components. We have added experiments and analysis to clarify the differences from existing methods, the pressure mechanism, the use of SABR with DiT, and the role of SNR gating. Below, we address the two main weaknesses and the two questions for authors.

Responses to the main weaknesses

Weakness 1 Differences from existing sampling control methods

Thank you for suggesting a clearer comparison with existing methods. These methods adjust generation in different ways. CADS adds annealed noise to the conditioning vector. FreeU reweights U-Net backbone and skip features. Limited-interval guidance and stage-wise CFG mainly study when and how strongly guidance is applied.

SABR turns measured local class-conditional responses into adaptive control at the block level, using a frozen reference. At a fixed block input and time condition, we compare the outputs with and without the current block's class condition to measure its direct conditional response. Head-class response statistics then define the cumulative pressure budget and provide a normalization scale for each block. The budget excess and local pressure determine which blocks to adjust and how much to reduce their conditional residuals, while SNR gating controls when the adjustment is allowed. This design follows our observation that local conditional responses vary with training frequency. The Original, Reverse, and Permuted experiments keep the class set and frequency distribution fixed while changing the mapping between class identity and training frequency. Higher pressure follows the reassigned low-frequency classes, providing a controlled test of this relationship.

SABR therefore contributes both the finding of frequency-related differences in local conditional responses and a method that uses this finding for adaptive intervention. Its frozen reference constrains the adjustment, while block selection directs it to the relevant locations.

Weakness 2 Links between pressure and mode concentration

Thank you for this suggestion. We added two analyses to answer whether naturally high-pressure trajectories within the same class show more mode concentration and whether they gain more from pressure-based attenuation of conditioning. The first groups baseline trajectories by pressure; the second compares paired baseline and SABR outputs.

Mode coverage of naturally high-pressure trajectories

We used a frozen DDPM without SABR to generate 500 images for each of the 34 tail classes, with CFG = 2.5 and 50-step DDIM. At each sampling step, we recorded local conditional pressure for the 4 candidate blocks. For each measurement, we fixed the block input and time step, compared its output with and without its class condition, and computed pressure as defined in the paper. These extra computations only record pressure and do not change the baseline generation process. We averaged the pressure values over the 50 steps and 4 blocks to give one mean-pressure score per trajectory.

We formed the groups separately within each class. To keep the group sizes equal, we selected a fixed set of 498 of the 500 trajectories in advance and ranked them by mean pressure. The lowest 166 form the low-pressure group, the next 166 the medium-pressure group, and the highest 166 the high-pressure group. Thus, pressure groups describe relative differences within a class rather than using one threshold across classes. Neither sample selection nor grouping uses the mode-coverage results.

We then evaluated mode coverage for each group. All groups use the same class-specific modes established from real images before evaluation. We assign generated image features to these modes and compute the effective mode ratio, mode-mass recall, JS divergence, and maximum mode-share ratio. Each metric is computed within each class, then averaged equally over the 34 classes.

Table 1. Real-mode metrics for within-class pressure groups on CIFAR-100-LT. Better results are bold.

Pressure groupEffective mode ratio ↑Mode-mass recall ↑JS divergence ↓Maximum mode-share ratio ↓
Low0.46500.46330.29812.8805
Medium0.43970.44240.31763.1590
High0.44170.43830.32443.1982

Compared with the low-pressure group, the high-pressure group has a lower effective mode ratio (0.4417 versus 0.4650) and mode-mass recall (0.4383 versus 0.4633), a higher JS divergence (0.3244 versus 0.2981), and a higher maximum mode-share ratio (3.1982 versus 2.8805). Together, these averages show lower mode coverage, a larger distribution mismatch, and stronger concentration in the largest mode for the naturally high-pressure group in this evaluation.

Pressure groups and gains from SABR

To compare the gains from SABR at different pressure levels, we fixed group membership using within-class pressure measured on the baseline trajectories. We then compared the four mode metrics between the paired baseline and SABR outputs in the low- and high-pressure groups. Group membership stays fixed throughout the paired comparison and is not reassigned from the intervention results.

Table 2. Mode metrics before and after SABR in low- and high-pressure groups. Changes are SABR minus baseline. Better results are bold.

MetricLow pressure baselineLow pressure SABRLow pressure changeHigh pressure baselineHigh pressure SABRHigh pressure change
Effective mode ratio ↑0.39520.4099+0.01470.47090.5022+0.0313
Mode-mass recall ↑0.39590.4087+0.01280.46760.4910+0.0234
JS divergence ↓0.36440.3509-0.01350.29060.2730-0.0176
Maximum mode-share ratio ↓3.35003.2471-0.10292.99532.8053-0.1900

Note: Changes are SABR minus baseline. The SABR values are calculated by adding each change to its baseline value. All values are shown to four decimal places.

SABR improves all four metrics in both groups, with larger average gains in the high-pressure group. For example, the effective mode ratio rises by 0.0313 in the high-pressure group and by 0.0147 in the low-pressure group. The maximum mode-share ratio falls by 0.1900 and 0.1029, respectively. In this paired evaluation, trajectories with higher baseline pressure gain more on average from SABR, linking within-class pressure to the response to intervention.

Responses to the questions for authors

Question 1 Applying SABR to DiT

Thank you for asking about other architectures. The original experiments mainly used U-Nets. We added Diffusion Transformer (DiT) experiments on CIFAR-10-LT and CIFAR-100-LT, comparing DiT with DiT (SABR) at 100, 200, 300, 400, and 500 epochs. This lets us examine the effect across different training stages.

Table 3. DiT and DiT (SABR) on CIFAR-10-LT at different training epochs. Better results are bold.

EpochMethodOverall FID ↓Overall Recall ↑Few FID ↓Few Recall ↑
100DiT55.25531.55%83.90135.47%
100DiT (SABR)55.11331.79%83.84435.75%
200DiT30.50334.66%37.54136.52%
200DiT (SABR)29.80535.21%36.74037.59%
300DiT20.23735.15%23.84430.05%
300DiT (SABR)19.13136.82%23.63231.75%
400DiT17.33234.82%24.86325.39%
400DiT (SABR)16.59638.04%18.78834.15%
500DiT14.64835.97%23.56724.35%
500DiT (SABR)14.11539.43%17.91833.02%

Table 4. DiT and DiT (SABR) on CIFAR-100-LT at different training epochs. Better results are bold.

EpochMethodOverall FID ↓Overall Recall ↑Few FID ↓Few Recall ↑
100DiT43.15131.49%56.00532.52%
100DiT (SABR)43.27031.48%56.20832.53%
200DiT25.70436.31%39.75837.76%
200DiT (SABR)25.48436.44%38.66038.04%
300DiT19.12437.40%31.31539.58%
300DiT (SABR)18.89937.35%31.10339.54%
400DiT16.31836.73%27.78136.35%
400DiT (SABR)16.10937.78%26.20838.25%
500DiT14.26934.87%24.73731.62%
500DiT (SABR)13.60537.04%22.27936.84%

SABR gives clear tail-class gains on both datasets. At 500 epochs, Few FID falls from 23.567 to 17.918 on CIFAR-10-LT, while Recall rises from 24.35% to 33.02%. On CIFAR-100-LT, Few FID falls from 24.737 to 22.279 and Recall rises from 31.62% to 36.84%. Overall FID and Recall also improve on both datasets. Results at the other epochs show gains at most training stages. These findings extend the results from the original U-Net models to the tested DiT models on two long-tailed datasets. We will add these experiments to the revised paper.

Question 2 The separate contribution of SNR gating

Thank you for asking about the separate role of SNR gating. We agree that the original Overall FID values of 13.038 for γ\gamma = 0 and 12.977 for the default setting are close, and that this metric alone gives an incomplete picture. We added a paired 2 × 2 comparison of cumulative versus per-step budgets and γ\gamma = 0 versus γ\gamma = 1. All settings use the same frozen CIFAR-100-LT DDPM model, 50-step DDIM, CFG = 2.5, paired initial noise, and 500 images per class. This tests the effects of SNR gating on generation quality and tail-class recall under both budget rules.

Table 5. Comparison of pressure-budget rules and SNR gating on CIFAR-100-LT. Better results are bold.

SettingOverall FID ↓Few FID ↓Few KID ×1,000 ↓Few Precision ↑Few Recall ↑
Baseline14.366831.71826.426674.06%29.09%
Cumulative, γ\gamma = 013.093127.25804.485274.75%32.75%
Cumulative, γ\gamma = 1 (default)12.951626.62334.397174.98%34.35%
Per-step, γ\gamma = 013.164427.51534.563575.49%33.61%
Per-step, γ\gamma = 113.019326.80254.420774.79%33.56%

With the other components fixed, SNR gating adds gains in generation quality and coverage. Under the default cumulative budget, changing γ\gamma from 0 to 1 lowers Overall FID from 13.0931 to 12.9516, Few FID from 27.2580 to 26.6233, and Few KID × 1,000 from 4.4852 to 4.3971. Few Precision rises from 74.75% to 74.98%, and Few Recall rises from 32.75% to 34.35%. These results show improvements in both overall quality and tail-class quality and recall. We will add this comparison to explain how SNR gating works with the cumulative budget.

Reviewer_e1z1OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies limited intra-class diversity in diffusion models trained on long-tailed data. It measures conditional pressure by comparing a block’s outputs with and without class conditioning while holding its input fixed. Frequency-reassignment experiments link lower training frequency to higher pressure and narrower representation spectra. The proposed method, SABR, uses reference trajectories from data-rich classes to decide when and where to reduce conditioning within a predefined low-resolution encoder region. Its gates depend on cumulative pressure, class-frequency groups, and SNR, without updating model parameters.

The evaluation includes several long-tailed datasets, DDPM/OC/CBDM checkpoints, different U-Net configurations and samplers, and a text-conditioned ArtBench experiment. The paper reports improved FID and Recall across all six CIFAR dataset–model pairs and improved tail-class mode coverage on CIFAR-100-LT. The gains are less consistent with CBDM, whose FID worsens on ImageNet-LT and ArtBench. Fixed gates, random block assignment, frozen mean gates, and CADS provide additional comparisons.

Critical strengths

1. The paper provides a useful investigation of the effect of training frequency. The Original/Reverse/Permuted experiments show that high conditional pressure follows the reassigned low-frequency classes rather than remaining attached to particular class identities. This is more informative than comparing fixed head and tail groups. The finding that removing Down-Late conditioning worsens denoising MSE but improves tail FID and Recall is also interesting. It motivates selective attenuation while showing why simply removing conditioning is not enough as precision decreases.

2. The method improves frozen models without additional training. The baseline and SABR evaluations use the same checkpoints. On CIFAR-10-LT and CIFAR-100-LT, overall FID and Recall improve for all three base models, and Few-class FID also improves (Tables 1–2). The real-mode analysis provides further evidence beyond FID, while the random-assignment control shows that intervention placement matters. These results support the usefulness of the method within the evaluated setting.

Critical weaknesses

1. Comparison with other diversity methods is limited.

SABR outperforms the tested CADS configuration for Overall, Medium, and Few classes in Table E7, including on Precision. However, CADS uses settings transferred from ImageNet-256 without tuning for CIFAR-100-LT (Appendix E.5). It is unclear whether this advantage would persist after tuning CADS on the target dataset with a comparable validation budget.

I also could not determine whether CFG[1] is used in the reported experiments or, if so, which guidance scales are applied. For checkpoints that support CFG, comparisons across guidance scales and guidance intervals would help determine whether local gating offers a better quality–diversity trade-off than adjusting guidance alone.

2. The effects of reference selection and class-group rules are not sufficiently tested.

Both the decision to intervene and the ranking of blocks depend on the head-class reference. The paper does not show how the gates or final results change when this reference is rebuilt using different calibration seeds, fewer trajectories, or different head-class subsets. The cross-architecture experiments construct a separate reference for each model, and the mismatch analysis in Appendix D.3 is theoretical. Neither establishes how stable the gains are under these calibration choices.

The class-group rules introduce another dependency. Many classes are not modified under the default setting, so their FID is identical to the baseline by construction (Table 2). This does not show that pressure alone can identify when intervention is unnecessary. Appendix B.1 defines settings that allow intervention on all groups, but their results are not reported. Known class frequencies are valid inputs here; what remains unclear is how much the reported results depend on the particular group-specific rules.


References

[1] Classifier-Free Diffusion Guidance, NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications 2021

Questions for authors

1. What exactly is gated in the text-conditioned experiment? For ArtBench/miniSD, which module output is interpolated, and how is hl0h_l^0 computed at the same block input? Is the text representation zeroed, replaced with an empty-prompt representation, or is a text-conditioning contribution removed? This distinction matters for understanding how the intervention relates to the one defined for class-conditioned residual blocks.

2. How are the gate values distributed in practice? Appendix E.6 reports E[1−g]=0.2683\mathbb{E}[1-g]=0.2683 and E[1−g2]=0.2745\mathbb{E}[1-g^2]=0.2745 for SABR. Do these statistics use the same averaging, and do they include non-intervention gates equal to one? Under identical averaging, the reported values imply E[g(1−g)]≈0.0062\mathbb{E}[g(1-g)]\approx 0.0062. Among modified blocks (g<1g<1), how often are gates near zero versus at intermediate values? This would clarify whether most interventions partially attenuate conditioning or nearly remove the selected conditional residual.

8 comments

Thank you for your careful review and for recognizing the frequency-reassignment experiments, the improvements to frozen models, and the real-mode analysis. We have added experiments and implementation details on fair comparisons, reference calibration, and class-group rules. Below, we address the two main weaknesses and the two questions for authors.

Responses to the main weaknesses

Weakness 1 Comparison with other diversity methods

CADS tuning on the target dataset

Thank you for pointing out how the choice of CADS settings affects a fair comparison. We agree that Table E7, which used settings transferred from ImageNet-256, did not establish how SABR compares with CADS tuned on the target dataset. We therefore retuned CADS on CIFAR-100-LT with a validation budget comparable to that used for SABR.

Table 1. Comparison of CADS and SABR on CIFAR-100-LT. Better results are bold.

Class groupMethodFID ↓KID×1000 ↓IS ↑Precision% ↑Recall% ↑
OverallCADS13.2583.66713.47480.1543.12
OverallSABR12.9773.37013.63180.5543.66
ManyCADS17.4435.46514.50588.5345.25
ManySABR18.9506.00314.69188.9944.07
MediumCADS20.4085.06110.78079.2543.25
MediumSABR20.4844.68810.70579.8843.08
FewCADS29.4405.9809.88175.5529.21
FewSABR27.0174.35410.05975.3932.96

SABR performs better on all five overall metrics. For Few classes, it improves FID, KID, IS, and Recall, with slightly lower Precision. The two methods each have advantages for Medium and Many classes. SABR thus gives better overall generation results and better tail-class quality and recall in this comparison. We will add the search ranges, the rule for choosing the configuration, and the validation budget to the revised paper.

CFG scales and guidance intervals

Thank you for noting that the original paper did not explain the CFG settings and their relationship with SABR clearly enough. We agree that results at the default CFG setting alone do not show whether the gains depend on a particular scale or interval. We have added comparisons across CFG scales and guidance intervals to examine this relationship.

CFG scale comparison

We added CFG scales of 1.0 and 1.5 alongside the original scale of 2.5. SABR improves all five tail-class metrics—FID, KID, IS, Precision, and Recall—at all three scales. Few FID falls by 10.80%, 12.67%, and 14.30%, while Few Recall rises by 4.59, 4.76, and 5.22 percentage points, respectively. The tail-class gains remain consistent across these scales, showing that they extend beyond the default CFG setting.

Table 2. Tail-class results at different CFG scales on CIFAR-100-LT. Better results are bold.

CFGMethodFewFID↓KID×1000↓IS↑Precision%↑Recall%↑
1.0DDPM19.9883.84710.1873.3440.40
1.0SABR17.8293.30410.3074.8544.99
1.5DDPM23.6023.92310.0775.0235.46
1.5SABR20.6123.45210.3375.5040.22
2.5DDPM31.5256.2889.5975.1027.74
2.5SABR27.0174.35410.0675.3932.96

Guidance interval comparison

We also tested whether changing only the time interval of global guidance could produce similar gains. To study early and late guidance separately, we set the boundaries at 0.2 and 0.8 in normalized diffusion time in advance. This separates roughly the first and last 20% of the time axis from its middle and defines three candidate intervals.

Sampling proceeds from high-noise states at larger t/Tt/T to low-noise states at smaller t/Tt/T. The interval [0, 0.8] reduces early guidance, [0.2, 1] reduces late guidance, and [0.2, 0.8] reduces guidance at both ends while keeping stronger guidance in the middle. These settings test the separate and combined effects of reducing guidance at the two ends. We use CIFAR-100-LT with IR100 and 50-step DDIM and generate 50,000 images for each setting. CFG is 2.5 inside the interval and 1 outside it. The DDPM and SABR controls use guidance throughout sampling, and SABR also retains its local gates.

Overall results

Table 3. Overall results for guidance-interval controls on CIFAR-100-LT. Better results are bold.

Method and stronger guidance interval t/TFID ↓KID×1000 ↓IS ↑Precision% ↑Recall% ↑
DDPM [0, 1]14.2273.70213.38080.2041.67
SABR [0, 1]12.9773.37013.63180.5543.66
Interval CFG [0.2, 0.8]13.5083.59813.31679.8745.54
Interval CFG [0, 0.8]13.6793.51613.53580.1742.38
Interval CFG [0.2, 1]14.1313.90013.27479.8442.02

Few-class results

Table 4. Tail-class results for guidance-interval controls on CIFAR-100-LT. Better results are bold.

Method and stronger guidance interval t/TFID ↓KID×1000 ↓IS ↑Precision% ↑Recall% ↑
DDPM [0, 1]31.5256.2889.58975.1027.74
SABR [0, 1]27.0174.35410.05975.3932.96
Interval CFG [0.2, 0.8]28.7535.2189.93675.5330.41
Interval CFG [0, 0.8]29.5226.1589.80475.1428.89
Interval CFG [0.2, 1]30.4275.4409.76975.2230.01

Among the three preset intervals, [0.2, 0.8] gives the lowest Overall and Few FID and the highest Overall and Few Recall. SABR further improves four tail-class metrics over this interval: Few FID falls from 28.753 to 27.017, Recall rises from 30.41% to 32.96%, a gain of 2.55 percentage points, KID × 1000 falls from 5.218 to 4.354, and IS rises from 9.936 to 10.059. Overall FID also falls from 13.508 to 12.977, showing better overall generation quality.

These comparisons show the added value of adjusting local conditioning based on the current generation state. SABR gives better tail-class quality and coverage than changing only the global guidance interval in the tested settings. There are also metric trade-offs: Few Precision is slightly lower with SABR (75.39% versus 75.53%), as is Overall Recall (43.66% versus 45.54%). Its main advantages here are lower Overall FID and further gains in tail-class quality and coverage.

Weakness 2 Reference selection and class-group rules

Head-class reference selection and calibration stability

Thank you for raising the question of reference dependence. We examined calibration randomness, the number of calibration trajectories, and the choice of head-class subsets. We describe the setup and findings of each test below.

Calibration randomness

With the generation and evaluation settings unchanged, we rebuilt the head-class reference using calibration seeds 91001 and 91002. Each reference was evaluated on 17,000 tail-class images. Few FID was 26.957 and 26.941, compared with 26.989 for the original reference. The similar results show stable tail-class performance when the calibration seed changes.

Number of calibration trajectories

In an earlier single-block configuration on CIFAR-10-LT, we used 4, 8, 16, and 32 nested calibration trajectories per head class, with paired initial noise and all other settings fixed. This test examines how calibration sample count affects the reference and the final generation results.

Table 5. Effect of calibration trajectory count in the single-block CIFAR-10-LT configuration. Better results are bold.

Calibration trajectories per head classFew FID ↓Precision % ↑Recall % ↑
428.308884.2040.52
828.096084.3041.39
1628.175184.2541.65
3228.156984.2341.63

Increasing the count from 4 to 32 changes Few FID from 28.3088 to 28.1569 and Recall from 40.52% to 41.63%. Results are close once the count reaches 8 trajectories per class.

Head-class reference subsets

We used the original SABR on CIFAR-100-LT at CFG = 2.5 and fixed the total calibration budget at 132 trajectories. The full reference uses all 33 head classes with 4 trajectories per class. Subsets A and B each use 11 head classes with 12 trajectories per class. A = {0, 3, 6, 9, 12, 15, 18, 21, 24, 27, 30} and B = {1, 4, 7, 10, 13, 16, 19, 22, 25, 28, 31}. The subsets were chosen in advance by class ID. All three settings are evaluated on the same tail classes, 66–99.

Table 6. Head-class reference subsets on CIFAR-100-LT with a fixed total calibration budget. Better results are bold.

ReferenceFew FID ↓KID×1000 ↓IS ↑Precision% ↑Recall% ↑
All 33 head classes27.0174.35410.05975.3932.96
Subset A27.1254.42010.04575.333.24
Subset B27.1924.47410.15675.3433.61

Relative to the full 33-class reference, Few FID increases by 0.108 for subset A and 0.175 for subset B, corresponding to relative increases of less than 0.7% in both cases. Precision changes by less than 0.1 percentage points, and Recall increases for both subsets. With the same total calibration budget, both 11-class subsets give tail-class results close to those of the full reference, showing that SABR can build a useful reference from different head-class selections.

Tail-class performance remains close when we change the calibration seed, the number of calibration trajectories, or the head-class subset. Rebuilding the reference causes only small changes in Few FID. With a fixed total budget, reducing the reference from 33 to 11 classes changes Few FID by less than 0.7% for both subsets. The trajectory-count test also shows that generation results stabilize with relatively few calibration samples. Together, these results show stable reference construction across the tested choices.

Class-group rules and head-class protection

Thank you for asking about the role of class-group rules. We added a comparison that allows intervention in every class. The default setting preserves the original behavior for head classes and applies the intervention rules to Medium and Few classes. The all-class setting makes every class eligible, with actual changes still determined by pressure feedback and the budget. Comparing the results by class group shows how this broader intervention range affects quality and coverage.

Table 7. Default class-group rules versus intervention eligibility for all classes on CIFAR-100-LT. Better results are bold.

Class groupSettingFID↓KID×1000↓IS↑Precision%↑Recall%↑
OverallDefault12.9773.37013.63180.5543.66
OverallAll classes eligible13.1423.58113.64680.8743.23
ManyDefault18.9506.00314.69188.9944.07
ManyAll classes eligible19.4956.62714.68689.7442.65
MediumDefault20.4844.68810.70579.8843.08
MediumAll classes eligible20.3444.70610.91179.5743.19
FewDefault27.0174.35410.05975.3932.96
FewAll classes eligible27.0174.35410.05975.3932.96

When all classes are eligible, FID for the Many group rises from 18.950 to 19.495 and Recall falls from 44.07% to 42.65%. Overall FID rises from 12.977 to 13.142 and Recall falls from 43.66% to 43.23%. Precision increases for both the Many group and the overall evaluation, and some Medium metrics also improve. This shows the quality–coverage trade-off from expanding the intervention range. The tail-class rules stay the same, and all five metrics for the Few group remain unchanged.

The default protection preserves head-class quality and coverage while retaining SABR's tail-class gains. Targeting Medium and Few classes gives better Overall FID and Recall than making every class eligible. Head-class protection therefore serves a clear purpose in SABR: it allows different treatment of class groups while preserving the original head-class behavior.

Responses to the questions for authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.