acceptodds
Under review as a conference paper at ICLR 2027

Traversing the Satisfaction-Diversity Frontier in Text-to-Image Diffusion

Abstract

Text-to-image generation enables users to explore several images generated from the same prompt. For these generated images to be useful, each one must reflect the user's preferences, measured by a learned reward, and differ visually from the others to maintain diversity. Existing methods are limited: they either address reward and diversity separately or combine them in one aggregate score, enabling high diversity to offset low rewards. In this paper, we address these limitations by formulating generation as *satisficing*: every image (candidate) must satisfy a reward floor and the batch of images must satisfy a diversity cutoff. The reward floor controls the balance between worst-candidate reward and batch diversity; we show that varying this floor defines a Pareto frontier. To traverse this frontier, we introduce SatisDive, a training-free inference-time method. SatisDive uses a batch-relative reward cutoff to distinguish lower- from higher-reward candidates, emphasizing reward improvement for candidates below the cutoff and diversity among candidates above it. On Pick-a-Pic, at matched DreamSim, SatisDive improves worst-candidate reward over FK steering by up to with FLUX.1-dev as the base model and HPSv3 as the reward, and by up to with SANA-1.6B as the base model and ImageReward as the reward. More broadly, across their overlapping DreamSim ranges, SatisDive's satisfaction–diversity curve Pareto-dominates FK steering's curve in each setting.

Primary area:
generative models
Keywords:
text-to-image generation, diffusion models, flow matching, inference-time steering, reward-diversity tradeoff

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_wTS5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses an interesting and practical problem at inference time: how to generate a batch of images that maintains high visual diversity while ensuring the worst image in the batch still meets a certain reward threshold. To solve this, they propose SatisDive, a training-free latent optimization method. The technique cleverly uses per-image reward floors and batch diversity constraints, applying batch-relative cutoffs for penalties and replacing consistently poor candidates in the late stages of generation. The results on Pick-a-Pic and HPDv2 are convincing, showing clear improvements over FK steering in worst-case reward at matched diversity levels.

Critical strengths

1.Clear formulation. By decoupling per-image reward requirements from batch-level diversity, the paper makes the intended trade-off highly intuitive. Furthermore, the theoretical analysis offers a solid interpretation of the method's behavior, clearly scoped within its stated assumptions.

2.The proposed method of boosting the reward of weaker candidates while diversifying the stronger ones works well in practice. The empirical Pareto curves convincingly show that it achieves a better worst-candidate reward than FK steering at matched diversity.

Critical weaknesses
  1. While pairwise cosine distances of pooled backbone features serve as a reasonable proxy for diversity, they risk overlooking nuanced semantic variations. More critically, computing pairwise distances across all images in a group incurs a computational overhead that scales quadratically with group size, making it increasingly prohibitive for larger groups.

  2. Higher computational overhead. As shown in Table 14, it takes 60.8 seconds per prompt compared to 37.7 seconds for FK steering, alongside higher GPU memory consumption. These timings are averaged over only three prompts at a small batch size (K=4), which means this method may not generalize well to larger batch sizes under acceptable time and gpu costs.

Questions for authors
  1. How sensitive are the results to diversity and reward cutoffs? Do these hyperparameter settings generalize to other baseline models without extensive retuning? Additionally, what is the recommended strategy for selecting appropriate cutoff values in practice?

  2. What is the rationale behind choosing cosine distance for measuring diversity? Have you evaluated the model's performance using alternative diversity measures?

3 comments

We thank the reviewer for recognizing the clear formulation and the improvements in worst-candidate reward at matched diversity.

Weakness 1: “More critically, computing pairwise distances across all images in a group incurs a computational overhead that scales quadratically with group size, making it increasingly prohibitive for larger groups.”

Weakness 2: “Higher computational overhead. As shown in Table 14, it takes 60.8 seconds per prompt compared to 37.7 seconds for FK steering, alongside higher GPU memory consumption. These timings are averaged over only three prompts at a small batch size (K=4), which means this method may not generalize well to larger batch sizes under acceptable time and gpu costs.”

Response: (a) Cost as batch size grows. Although the number of pairwise distances grows quadratically, our SD1.5 measurements show that doubling the batch size approximately doubles the total runtime of both SatisDive and FK steering over K=4,8,16K=4,8,16. We measured runtime and GPU memory on 50 prompts, using the same RTX A6000 GPU, prompts and starting seeds.

Batch size KKSatisDive time (s/prompt)FK steering time (s/prompt)SatisDive memory (GiB)FK steering memory (GiB)
429.73 ± 0.2416.93 ± 0.3612.337.48
858.69 ± 0.5231.85 ± 0.6516.1110.24
16118.84 ± 1.4764.31 ± 1.8822.5815.74

Time is mean ± standard deviation; memory is the largest sampled GPU memory usage. The full profiling protocol is in Appendix D.15.

(b) Expanded FLUX timing. We have replaced the three-prompt estimate with measurements on 50 prompts, using the same two A100 80 GB GPUs for both methods:

FLUX, K=4K=4Time (seconds/prompt, mean ± standard deviation)
FK steering74.17 ± 0.90
SatisDive138.64 ± 0.29

SatisDive achieves higher worst-candidate reward at matched diversity (Figure 3), with 1.87×1.87\times FK steering's runtime in the FLUX timing study.

Question 1: “How sensitive are the results to diversity and reward cutoffs? Do these hyperparameter settings generalize to other baseline models without extensive retuning? Additionally, what is the recommended strategy for selecting appropriate cutoff values in practice?”

Response: (a) Cutoff sensitivity. Δ\Delta controls the allowed reward gap from the batch maximum. Figure 3 shows how varying Δ\Delta changes the reward–diversity tradeoff: smaller values favor higher worst-candidate reward, while larger values permit greater diversity. Changing the diversity cutoff δ\delta does not make the diversity step larger or smaller, because we normalize each nonzero diversity gradient as described in Appendix C.

(b) Cutoff selection and transfer. For each prompt, we subtract the lowest reward from the highest reward in the base model's batch. We use about 75% of the median of these differences as a starting value for Δ\Delta. We initialize δ\delta from the median pairwise feature distance at the first update. We use the same selection rules across FLUX, SANA and SD1.5, including FLUX/HPSv3 on both Pick-a-Pic and HPDv2. We compute the cutoffs for each setting rather than reuse the same numerical values.

Weakness 1: “While pairwise cosine distances of pooled backbone features serve as a reasonable proxy for diversity, they risk overlooking nuanced semantic variations.”

Question 2: “What is the rationale behind choosing cosine distance for measuring diversity? Have you evaluated the model's performance using alternative diversity measures?”

Response: Cosine distance compares feature directions without treating larger feature norms as greater diversity, and it is differentiable for latent guidance. Cosine similarity is also used by NegToMe to match diffusion features and by CNO in its contrastive objective on denoised latent predictions.

Although pooled-feature cosine distance may miss some semantic differences, we evaluate the final images separately with the standard perceptual metrics DreamSim and LPIPS. NegToMe also evaluates diversity with DreamSim, while DAS uses LPIPS. Both metrics show greater diversity than FK steering in the FLUX/HPSv3 comparison (Table 1):

MethodDreamSimLPIPS
SatisDive0.5550.531
FK steering0.3060.291

Thus the measured diversity improvement is not specific to the internal cosine distance used for guidance.

Thank you for the response. But my concerns about the computation overload and generalizability of the proposed method remain unresolved.

We thank the reviewer for the follow-up. We further clarify our response.

Computational costs. Although SatisDive is more computationally expensive than FK steering, it achieves higher worst-candidate reward at matched diversity on FLUX/HPSv3 and SANA/ImageReward (Figure 3), as the reviewer notes. The additional computation comes from forward and backward evaluations through the generator to calculate the reward and diversity updates used to guide each candidate. The reported runtime comparison therefore does not establish that FK steering achieves the same reward–diversity outcome at lower cost.

Furthermore, although the number of pairwise cosine comparisons grows quadratically with batch size, we cache and reuse the pooled feature vectors obtained during each candidate’s forward pass. Each additional comparison therefore requires only inexpensive vector calculations, without another forward or backward evaluation through the generator. The costly generator evaluations grow with the number of candidates rather than the number of pairs, which explains the approximately linear total runtime observed on SD1.5: 29.73, 58.69 and 118.84 seconds per prompt for 4, 8 and 16 candidates, respectively. Sampled GPU memory increases from 12.33 to 22.58 GiB between four and sixteen candidates, as shown in Appendix D.15.

Appendix D.1 also reports output quality for SD1.5/ImageReward at K=16K=16: SatisDive achieves DreamSim of 0.77 versus FK steering’s 0.45, with average worst-candidate reward of 0.41 versus 0.69. Thus, SatisDive’s diversity benefit extends to this larger batch size, with a corresponding reduction in worst-candidate reward.

Transfer across backbones without retuning. Regarding generalizability of Δ\Delta across base models, our ImageReward experiments (Tables 3, 12 and 13) use the same Δ=0.5\Delta=0.5 across FLUX, SANA, and SD1.5, without backbone-specific retuning of Δ\Delta. At this setting, SatisDive achieves higher DreamSim than FK steering on all three backbones, with comparable satisfaction rate to FK steering on FLUX and higher satisfaction rate than FK steering on SANA and SD1.5.

Sensitivity to reward tolerance. We additionally illustrate how Δ\Delta affects the results with an example. Consider a candidate whose reward is 0.6 less than the batch's highest. If Δ=0.3\Delta=0.3, this candidate falls below the reward cutoff (since 0.6 > 0.3) and incurs a larger reward penalty. In contrast, if Δ=0.75\Delta=0.75, that same candidate is above the cutoff and is weighted more in the diversity penalty. The tradeoff is visible on SANA in Table 3, in which increasing Δ\Delta from 0.3 to 0.75 improves DreamSim from 0.591 to 0.658, while average worst-candidate reward falls from 1.29 to 1.13.

For practical selection of Δ\Delta, we compute the difference between the highest and lowest rewards in each base-model batch and use about 75% of the median of these differences as a starting value.

Automatic diversity-cutoff selection. The diversity cutoff δ\delta is designed to not require a separate tuning search. As specified in Appendix C, δ\delta is computed automatically from the median pairwise feature distance among candidates at the first gradient update step. For a given batch, changing δ\delta scales the diversity gradient magnitude without changing its direction. We normalize the reward and diversity gradients separately so that differences in their numerical scales do not make one dominate the latent update. Since changing δ\delta only changes the diversity gradient’s magnitude, this normalization also means that changing δ\delta does not alter a given candidate’s diversity update. A separate search over δ\delta would therefore add tuning work without changing these updates. Thus, we can instead adjust Δ\Delta to choose the reward–diversity tradeoff.

Alternative diversity guidance. Our original response explained the choice of cosine distance. We now additionally test LPIPS during guidance, rather than only using perceptual metrics to evaluate the final images. We compared both settings on the same 100 Pick-a-Pic prompts using SD1.5/ImageReward, K=4K=4 and Δ=1\Delta=1.

Diversity guidanceMean worst-candidate ImageReward ↑Mean pairwise DreamSim ↑
Pooled-feature cosine distance0.3350.743
LPIPS0.3300.756

Both guidance choices produce similar observed reward–diversity results in this setting, providing a tested alternative to cosine-based guidance.

We hope these explanations and additional results clarify SatisDive’s computational costs and generalizability. We would be happy to address any further questions.

Reviewer_g3KEOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies batch text-to-image generation, where a user wants K good and mutually different candidates for one prompt. Instead of weighting a diversity penalty, it formalizes the task as satisficing: the target distribution is the base model's batch product restricted by two hard floors — every candidate must clear a reward floor, and the batch must clear a diversity floor.

Critical strengths
  • The satisficing formulation is a real conceptual advance over reward–diversity weighting, and it comes with an optimality story the baselines lack.
  • VASR and DAS get the best min reward (9.99 / 10.06 on FLUX/HPSv3) but at DreamSim 0.053 / 0.059 — effectively K copies of one image, which for batch generation is a failure, not a success.
Critical weaknesses
  • Algorithm 1's Require list is (base model, prompt, reward, batch size K, steps, update steps, tolerance Δ); neither the reward floor nor the diversity floor appears. The threshold is instead recomputed from the current batch maximum at every update, and the diversity floor is set to that run's median pairwise distance.

  • The reward gain comes mainly from a latent-copy step whose replaced pairs are near-duplicates, so the reported diversity and the enforced floor are in tension — and the floor is missed for ~45% of candidates.

Questions for authors

Algorithm 1 recomputes the threshold from the running batch maximum each step, whereas Prop. 2 freezes it. Under what conditions does the running-maximum procedure optimize the fixed-floor tilted distribution? If a floor is meant to be an input, why is it absent from Require, and what changes if it is fixed a priori (e.g. at the base median) instead of derived from the running max?

2 comments

We thank the reviewer for recognizing the conceptual value of the satisficing formulation and the importance of diversity in batch generation.

Weakness 1: “Algorithm 1's Require list is (base model, prompt, reward, batch size KK, steps, update steps, tolerance Δ\Delta); neither the reward floor nor the diversity floor appears. The threshold is instead recomputed from the current batch maximum at every update, and the diversity floor is set to that run's median pairwise distance.”

Question: “Algorithm 1 recomputes the threshold from the running batch maximum each step, whereas Prop. 2 freezes it. Under what conditions does the running-maximum procedure optimize the fixed-floor tilted distribution? If a floor is meant to be an input, why is it absent from Require, and what changes if it is fixed a priori (e.g. at the base median) instead of derived from the running max?”

Response: (a) Why the floors are not inputs. We use a reward tolerance Δ\Delta so users can control the reward gap within a batch without choosing an absolute reward floor beforehand. Algorithm 1 computes τ=max⁡krk−Δ\tau=\max_k r_k-\Delta and recomputes τ\tau as candidate rewards change during denoising. For diversity, we calculate δ\delta from the candidate features rather than require users to specify a feature-distance threshold.

(b) Connection to the fixed-floor target. Proposition 2 applies to the objective with τ\tau and δ\delta held fixed. It assumes that the base model has a nonzero probability of generating a batch satisfying both floors, and zero probability of equality at either floor. Under these conditions, as the temperatures approach zero, the distribution induced by the objective converges to Equation (4) with reward floor ρ=τ\rho=\tau. Algorithm 1 recomputes τ\tau between updates, so the proposition describes the target associated with each cutoff; it does not establish that the adaptive procedure optimizes one fixed-floor distribution throughout generation. We have clarified the cutoff choice after Eq. (5) and the fixed-cutoff condition before Proposition 2.

(c) Results with a fixed reward floor. We compare the adaptive cutoff with a fixed cutoff of 0.1800.180, the median reward of images generated by the base model. Both settings use SD1.5/ImageReward on the same 499 Pick-a-Pic prompts, and we evaluate both against the reward floor ρ=0.180\rho=0.180.

Reward cutoff during generationAverage worst-candidate rewardMean DreamSimBatches where all four candidates satisfy ρ=0.180\rho=0.180
Adaptive, Δ=0.5\Delta=0.50.71130.671074.9%
Fixed, τ=ρ=0.180\tau=\rho=0.1800.57310.677488.0%

The fixed cutoff produces more batches satisfying the specified reward floor. The adaptive cutoff gives higher average worst-candidate reward at similar observed mean DreamSim.

Weakness 2: "The reward gain comes mainly from a latent-copy step whose replaced pairs are near-duplicates, so the reported diversity and the enforced floor are in tension — and the floor is missed for ~45% of candidates."

Response: We agree that the latent-copy step introduces a reward–diversity tradeoff. The FLUX/HPSv3 ablation in Appendix D.5 quantifies both effects:

FLUX variantWorst-candidate rewardMean DreamSim
Without latent-copy8.580.68
Full SatisDive9.390.56

Although the latent-copy reduces diversity, the full method achieves higher worst-candidate reward than FK steering at matched diversity (main text Figure 3).

Furthermore, the mentioned 45% concerns the fixed evaluation floor ρ≈9.66\rho\approx 9.66 from Table 1 in the main text, calculated as the median reward of images generated by the base model across all prompts. Since SatisDive uses the batch-relative cutoff τ\tau during generation, a candidate can exceed τ\tau while remaining below the evaluated reward floor ρ\rho. The latent-copy step can therefore improve a candidate's reward without necessarily satisfying ρ\rho.

AuthorsOpenReview ↗

Urgent Request to Reviewer

Dear Reviewer,

Thank you again for your thoughtful review and constructive feedback. As the discussion period approaches its end, we wanted to follow up on our rebuttal. We have clarified the relationship between the adaptive cutoff and the fixed-floor theoretical target, added a fixed-floor experiment, and explained the effects of late-stage latent replacement on reward and diversity.

We would be grateful if you could review our responses and let us know whether they address your concerns or whether any further clarification would be helpful.

Thank you for your time and consideration.

Best regards,

The Authors

Reviewer_FU9yOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose SatisDive, an inference-time method for obtaining high-reward samples while preserving the diversity of generated images in text-to-image generation. It combines the target reward and diversity gradients and applies late-stage latent replacement, improving the reward–diversity trade-off over existing inference-time reward alignment methods such as DAS and FK Steering, with diversity measured using DreamSim.

Critical strengths
  • The paper addresses a very practical problem: generating diverse images while ensuring that all images are usable. The result showing higher rewards with almost the same diversity as the base model for sufficiently large values of Δ\Delta (in Figure 3) is especially impressive.
  • The paper is clearly written and easy to follow. It also provides various experiments and ablation studies to enhance the original claim of the paper.
Critical weaknesses
  • The main baseline comparisons (in Tables 1 and 2) are insufficiently aligned with the paper’s motivation of obtaining high-reward samples with sufficient diversity. FK Steering targets a reward-tilted distribution rather than explicitly preserving diversity, while NegToMe does not aim to increase reward. Comparisons with diversity-aware methods such as SGI [1] and CREPE [2] should therefore receive greater emphasis in the main experiments, particularly regarding the Pareto frontier between reward and diversity.
  • The novelty and benefits of late-stage latent replacement are insufficiently addressed. The proposed procedure is equivalent to performing more resampling at later denoising timesteps, which can be achieved by increasing the reward-tilting weight λ\lambda, a technique already mentioned in FK Steering and DAS. The authors should clarify the differences and demonstrate its advantages over a timestep-dependent resampling strategy.

[1] Scaling Group Inference for Diverse and High-Quality Generation (Parmar et al., ICLR 2026)

[2] CREPE: Controlling diffusion with REPlica Exchange (He et al., ICLR 2026)

Questions for authors
  1. Can SatisDive be extended to non-differentiable rewards, such as GenEval [1] or VQAScore [2]?

  2. How robust is performance to the choice of layer feature within each backbone? Could the authors provide an ablation comparing different layers?

  3. Could the authors also report average-candidate and best-candidate reward versus diversity curves, alongside the worst-candidate reward curves in Figure 3, to provide a more complete picture of the quality–diversity trade-off?

[1] GenEval: An Object-Focused Framework for Evaluating Text-to-Image Alignment (Ghosh et al., 2023)

[2] Evaluating Text-to-Visual Generation with Image-to-Text Generation (Lin et al., ECCV 2024)

4 comments
AuthorsOpenReview ↗

Response (1/2)

We thank the reviewer for recognizing the practical value of generating diverse, usable images and the reward–diversity improvements in Figure 3.

Weakness 1: “The main baseline comparisons (in Tables 1 and 2) are insufficiently aligned with the paper’s motivation of obtaining high-reward samples with sufficient diversity. FK Steering targets a reward-tilted distribution rather than explicitly preserving diversity, while NegToMe does not aim to increase reward. Comparisons with diversity-aware methods such as SGI [1] and CREPE [2] should therefore receive greater emphasis in the main experiments, particularly regarding the Pareto frontier between reward and diversity.”

Response: We have moved the CREPE and SGI comparison tables into the main experiments (Section 4.1, Tables 4–5). These methods pursue both reward and diversity, allowing us to compare their tradeoffs with SatisDive's.

With SD1.5/ImageReward, SatisDive achieves higher worst-candidate reward than the evaluated CREPE settings at comparable diversity. The table also reports the percentage of batches in which all four candidates satisfy the reward floor ρ≈0.18\rho\approx0.18, using 499 Pick-a-Pic prompts.

Method / settingWorst-candidate rewardMean DreamSimBatches where all four candidates satisfy ρ\rho
CREPE, scale 10−0.330.76133.5%
CREPE, scale 1000.070.75347.3%
SatisDive, Δ=1.0\Delta=1.00.440.76263.5%

SatisDive also returns more batches in which all four candidates satisfy the reward floor. Appendix D.8, Table 14 reports the batch-level results.

For FLUX/HPSv3, we compare SatisDive with SGI across five diversity weights. Both methods return four images; SGI selects them from 128 initial candidates.

Method / settingInitial candidatesWorst-candidate rewardMean DreamSim
SatisDive, Δ=3\Delta=348.920.646
SGI, diversity weight 01289.320.647
SGI, diversity weight 1.51289.170.693
SGI, diversity weight 61288.890.744
SGI, diversity weight 241288.410.803
SGI, diversity weight 961288.120.823

SGI achieves higher worst-candidate reward at comparable diversity, but starts with 128 candidates rather than four. It uses 340 diffusion-model passes per prompt, compared with at most 200 for SatisDive (Figure 6).

Weakness 2: “The novelty and benefits of late-stage latent replacement are insufficiently addressed. The proposed procedure is equivalent to performing more resampling at later denoising timesteps, which can be achieved by increasing the reward-tilting weight λ\lambda, a technique already mentioned in FK Steering and DAS. The authors should clarify the differences and demonstrate its advantages over a timestep-dependent resampling strategy.”

Response: (a) Difference from increasing λ\lambda. Increasing FK’s λ\lambda makes resampling favor higher-reward candidates; FK does not use a reward cutoff to decide which candidates to keep. In contrast, SatisDive keeps every candidate that satisfies the reward cutoff τ\tau, even if another candidate has higher reward. It replaces a candidate only when both its current and extrapolated rewards fall below τ\tau, copying the latent of a uniformly chosen candidate that meets τ\tau.

(b) Comparison with late-weighted resampling. To test the reviewer's suggestion of increasing resampling strength later in denoising, we compared constant and late-weighted FK on 499 Pick-a-Pic prompts with SD1.5/ImageReward. We kept FK's resampling rule unchanged. For each strength λ=5,10,20\lambda=5,10,20, the late-weighted schedule uses 0.25×λ0.25\times\lambda during the first half of denoising, then increases linearly to the full value. SatisDive at Δ=0.75\Delta=0.75 achieves higher worst-candidate reward and DreamSim than all six tested settings:

MethodStrength / endpointWorst-candidate rewardDreamSim
SatisDiveΔ=0.75\Delta=0.750.53810.7229
FK, constant50.19430.6139
FK, late-weighted5−0.00040.6737
FK, constant100.35490.4875
FK, late-weighted100.27900.5431
FK, constant200.44250.4060
FK, late-weighted200.43020.4472

Late weighting increases FK's diversity at each endpoint. The paired 95% intervals for SatisDive's gains are above zero for both metrics in all six comparisons. Appendix D.16, Figure 8 reports the full curves and protocol.

AuthorsOpenReview ↗

Response (2/2)

Question 1: “Can SatisDive be extended to non-differentiable rewards, such as GenEval [1] or VQAScore [2]?”

Response: Yes. We have now trained a differentiable model to predict VQAScore from an image and its prompt, then used its scores and gradients to guide SatisDive with SD1.5. We evaluated the final images with VQAScore on 499 Pick-a-Pic prompts, with four candidates per prompt.

Method / settingWorst-candidate VQAScoreMean DreamSim
Base model0.61720.8608
SatisDive, Δ=0.25\Delta=0.250.67150.7461
SatisDive, Δ=0.5\Delta=0.50.64210.8325

SatisDive improves worst-candidate VQAScore at both tolerances. Increasing Δ\Delta retains more diversity, with a smaller reward improvement.

Question 2: “How robust is performance to the choice of layer feature within each backbone? Could the authors provide an ablation comparing different layers?”

Response: Worst-candidate reward and DreamSim change little across the three tested layers in FLUX, SANA and SD1.5. For each backbone, we use the same 100 Pick-a-Pic prompts and starting seeds, changing only the feature layer.

FLUX blockWorst-candidate HPSv3DreamSim
89.77790.4788
12 (default)9.75410.4903
169.77420.4920
SANA blockWorst-candidate ImageRewardDreamSim
31.15360.6247
7 (default)1.14690.6312
151.14720.6182
SD1.5 layerWorst-candidate ImageRewardDreamSim
Last down-block residual0.42170.6963
Mid-block (default)0.44280.6927
First up-block residual0.42980.6784

For each alternative layer, the paired 95% intervals for changes in reward and DreamSim relative to the default include zero. We have added these results to Appendix D.14, Tables 17–19.

Question 3: “Could the authors also report average-candidate and best-candidate reward versus diversity curves, alongside the worst-candidate reward curves in Figure 3, to provide a more complete picture of the quality–diversity trade-off?”

Response: We have added average- and best-candidate reward curves for FLUX and SANA to Appendix D.13, Figure 7, using the same 499 prompt batches and settings as Figure 3. For both backbones, SatisDive achieves higher average- and best-candidate reward than FK steering at matched diversity along the plotted curves.

Thank you for the response and for adding the SGI comparison. However, my concern regarding this comparison remains unresolved. SGI achieves higher worst-candidate reward at comparable diversity, whereas SatisDive uses fewer diffusion-model forward passes. These different compute budgets make it difficult to assess the relative performance of the two methods or substantiate SatisDive’s claimed efficiency advantage.

Could you provide either a compute-matched comparison or a compute v.s. worst-candidate reward trade-off curve for both methods, with diversity controlled at comparable levels?

Thank you for the follow-up. We have now run an approximately FLOP-matched comparison, setting SGI's initial pool to M=72M=72 candidates. Both methods return four images and are evaluated on the same 50 fixed Pick-a-Pic prompts with FLUX.1-dev/HPSv3.

Matched computation. On a fixed profiling prompt, our reported FLUX setting uses 18.061 PFLOPs, and SGI with M=72M=72 is comparable at 17.950 PFLOPs. These measured FLOP budgets include generation, scoring and backward computation. SatisDive uses a six-update schedule at denoising steps 4, 8, 12, 16, 20 and 24. SGI uses its reported 50% pruning rule, reducing 72 initial candidates to four.

SGI: fixed initial pool M=72M=72, four returned candidates.

Diversity weight λ\lambdaMean worst-candidate HPSv3 ↑Mean pairwise DreamSim ↑
09.3020.634
1.59.1780.668
68.8360.716
248.4940.756
968.1420.776

SatisDive: four candidates throughout, normal six-update schedule.

Reward tolerance Δ\DeltaMean worst-candidate HPSv3 ↑Mean pairwise DreamSim ↑
0.00019.6570.282
0.59.6310.389
19.4340.504
1.59.3270.560
38.9910.638

These approximately FLOP-matched results show SatisDive and SGI cover different reward–diversity ranges; SatisDive reaches higher reward at lower diversity, while SGI extends to higher diversity. In particular, SGI achieves a worst-candidate HPSv3 of 9.302 and DreamSim of 0.634. Comparatively, SatisDive achieves a worst-candidate HPSv3 of 8.991 at very similar 0.638 DreamSim.

We have revised the manuscript's computational-efficiency comparison. The main text now reports the matched results (Table 5), while the original M=128M=128 comparison remains in Appendix D.9. We also clarify that Figure 6 (Appendix D.10) reports model passes, while Table 5 reports the profiled neural FLOPs for this two-method comparison. Both methods return four images but allocate computation differently: SGI selects from 72 initial candidates, whereas SatisDive guides four candidates throughout denoising.

Reviewer_46njOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the tradeoff between worst-candidate reward and within-batch diversity in text-to-image generation. It formulates an ideal satisficing target requiring every candidate to exceed a reward floor and the batch to exceed a diversity cutoff. The proposed training-free method, SatisDive, combines batch-relative reward gates, a reward-weighted diversity penalty, and late-stage replacement of low-reward latents. Experiments across FLUX, SANA, and Stable Diffusion show favorable reward–diversity tradeoffs against the evaluated baselines. The theoretical analysis characterizes reward-conditioned distributions under a support-based diversity assumption and establishes a zero-temperature limit for a fixed-threshold surrogate distribution.

Critical strengths
  1. The empirical reward–diversity improvements are meaningful. Figure 3 shows favorable tradeoffs against the evaluated FK configurations, particularly for SANA with ImageReward. The qualitative examples also demonstrate visibly different compositions while avoiding some failures of diversity-only guidance. Evaluation across multiple generators, rewards, datasets, and batch sizes provides useful evidence that the approach is not restricted to a single configuration. The appendices further report held-out rewards, component ablations, uncertainty estimates, and computation costs.

  2. The focus on the weakest candidate addresses a relevant limitation of aggregate objectives. When users need several usable alternatives, a strong batch average can conceal an unacceptable candidate. The paper makes this distinction explicit and proposes a practical combination of reward-dependent guidance and selective replacement. This perspective is consequential even if the theoretical claims ultimately need to be narrowed.

Critical weaknesses
  1. The theoretical results do not justify the claimed frontier for the experimental metric or Algorithm 1. Theorem 3 assumes diversity is monotone under support restriction. Appendix B.2 explicitly states that distributions with identical supports have identical diversity, irrespective of their probability masses. This assumption excludes expected pairwise cosine distance and analogous perceptual-distance statistics. For example, consider two equally rewarded images at cosine distance 1. A base distribution assigning probabilities (0.9,0.1)(0.9,0.1) has expected pairwise distance 0.180.18, while a distribution assigning (0.5,0.5)(0.5,0.5) has expected distance 0.50.5. Both have identical support, identical guaranteed reward, and finite KL divergence to the base distribution. Thus, reward conditioning need not maximize pairwise diversity at a fixed guaranteed reward. Proposition 2 introduces a separate gap. Appendix B.3 fixes the thresholds and defines a density proportional to $
∏kpθ(xk∣c)exp⁡[−J(x1:K)]\prod_k p_\theta(x_k\mid c)\exp[-J(x_{1:K})]

$. Its convergence result concerns this constructed density. Algorithm 1 instead recomputes the relative cutoff during denoising, performs replacement, and uses detached, separately normalized, clipped, and scheduled gradients as specified in Appendix C. No argument connects its output distribution to the constructed density.

  1. The conditional mathematical arguments appear internally consistent, but their scope falls short of the central “provably traverses” claim. This substantially affects the theoretical contribution. The empirical evaluation does not directly establish the stated satisfaction requirement or isolate the proposed objective’s contribution. Algorithm 1 receives Δ\Delta and uses τ=max⁡krk−Δ\tau=\max_k r_k-\Delta, rather than receiving an absolute reward floor ρ\rho. Meeting this relative criterion does not establish satisfaction of an external quality requirement. Moreover, Equation 11 reports candidate-level satisfaction, whereas the proposed target requires every candidate and the batch diversity constraint to be satisfied jointly.For example, Table 1 reports candidate-level satisfaction 0.5520.552 and mean minimum reward 9.399.39, below the evaluation floor 9.669.66. The all-candidates-success rate cannot exceed 0.5520.552. These results demonstrate improvement, but leave the practical batch-feasibility claim insufficiently evaluated. The component attribution also remains incomplete. In Table 8, removing replacement reduces minimum reward from 9.399.39 to 8.588.58, approximately base performance. Removing the diversity penalty changes DreamSim from 0.560.56 to 0.540.54, while increasing minimum reward to 9.439.43. The diversity difference is reported as statistically significant, but this single operating point does not establish whether the full objective improves the attainable frontier beyond reward guidance and selective replacement. Finally, at the last scheduled update, m=0m=0, so every below-threshold candidate satisfies the replacement condition. For FLUX, no subsequent diversity-gradient update occurs after these replacements. The explanation that both JdJ_d and stochastic continuation diversify copied trajectories therefore needs qualification for these events.
Questions for authors
  1. Scope of the guarantee: What precise guarantee is intended for Algorithm 1 and the experimentally measured diversity? Can the authors establish the missing connection, or explicitly restrict the theoretical claims to the ideal target family? Please also clarify how separate gradient normalization preserves the intended diversity-cutoff behavior.

  2. Batch-level satisfaction: How can a user specify an absolute reward floor? Using the existing generated samples, can the authors report the fractions of batches satisfying $

min⁡krk≥ρ\min_k r_k\ge\rho

$ and satisfying both the reward and diversity requirements, with the final diversity statistic and cutoff clearly defined?

  1. Contribution of the diversity objective: In one principal setting, can the authors provide a Δ\Delta sweep for the variant without JdJ_d, comparing its reward–diversity curve with the full method? This would help determine whether the proposed diversity objective materially improves the frontier. Please also clarify the mechanism responsible for diversification after the final replacement, when no further JdJ_d update occurs.
3 comments
AuthorsOpenReview ↗

Response (1/2)

Thank you for the detailed questions on the theoretical guarantees and batch-level evaluation. We clarify their scope and report the requested batch and ablation analyses below.

Question 1: “What precise guarantee is intended for Algorithm 1 and the experimentally measured diversity? Can the authors establish the missing connection, or explicitly restrict the theoretical claims to the ideal target family? Please also clarify how separate gradient normalization preserves the intended diversity-cutoff behavior.”

Response: (a) Scope of the guarantee. We restrict the theoretical guarantees to the ideal target family and its smooth relaxation, rather than Algorithm 1's output distribution or mean pairwise DreamSim. Proposition 2 (p. 5) connects our smooth objective to the hard reward and diversity constraints in Equation (4). With the cutoffs fixed and under the stated feasibility and boundary conditions, the distribution defined by the objective converges to the constrained target as the temperatures approach zero. Algorithm 1 uses the gradients of these smooth penalties to update candidate latents during denoising.

Theorem 3 (p. 6) establishes a frontier for the ideal reward-conditioned distributions under its support-restriction assumption. The two-image example shows why this result does not extend to expected pairwise distance. Our empirical result holds across the overlapping DreamSim ranges in Figure 3: SatisDive achieves higher worst-candidate reward than the evaluated FK settings at matched DreamSim on FLUX/HPSv3 and SANA/ImageReward.

Related work likewise supports comparative improvements through experiments while stating its theoretical guarantees separately. DAS establishes asymptotic sampling guarantees (§3.3.4) and evaluates Pareto tradeoffs between reward metrics empirically (§4.2). SGI evaluates quality–diversity tradeoffs empirically (§4.2), including with metrics not used during selection (Appendix A.1).

(b) Separate gradient normalization. The normalized diversity update does not become smaller solely because weighted feature diversity DD exceeds its cutoff δ\delta. As described in Appendix C, dividing a nonzero diversity gradient by its norm cancels the positive, cutoff-dependent factor from Eq. (8) but preserves the direction toward increasing DD.

AuthorsOpenReview ↗

Response (2/2)

Question 2: “How can a user specify an absolute reward floor? Using the existing generated samples, can the authors report the fractions of batches satisfying [min⁡krk≥ρ][\min_k r_k\geq\rho] and satisfying both the reward and diversity requirements, with the final diversity statistic and cutoff clearly defined?”

Response: (a) Specifying an absolute reward floor. Users can choose ρ\rho by generating images with the base model on a separate set of prompts and selecting a percentile of their reward scores, such as the median. Our experiments use the median reward of the base-model images as the evaluation floor. We also evaluated a version that keeps the reward cutoff fixed at the chosen floor, τ=ρ\tau=\rho, throughout generation. With SD1.5/ImageReward, all four candidates satisfy ρ=0.180\rho=0.180 in 88.0% of the 499 batches, as reported in Appendix D.19.

(b) Batch-level satisfaction. We now report how often all four candidates satisfy the reward floor, both with and without a minimum batch-diversity requirement. In the FLUX and SANA comparisons below, SatisDive uses Algorithm 1's batch-relative cutoff τ=max⁡krk−Δ\tau=\max_k r_k-\Delta. We have added these results to Appendix D.18.

Each percentage is out of all 499 batches of four images. We measure batch diversity by averaging the Euclidean distances between the DreamSim embeddings of all pairs of final images.

FLUX / HPSv3: reward floor ρ≈9.66\rho\approx9.66.

MethodAll four satisfy reward floorAll four satisfy reward floor and batch diversity ≥0.50\geq0.50All four satisfy reward floor and batch diversity ≥0.75\geq0.75
SatisDive, Δ=0.0001\Delta=0.000147.1%6.6%0.0%
SatisDive, Δ=0.5\Delta=0.546.5%12.0%0.6%
SatisDive, Δ=1\Delta=146.7%22.2%2.0%
SatisDive, Δ=1.5\Delta=1.543.7%26.7%2.8%
SatisDive, Δ=3\Delta=336.1%28.5%3.4%
FK steering48.5%8.0%0.4%
DAS57.9%0.0%0.0%

SANA / ImageReward: reward floor ρ≈1.19\rho\approx1.19.

MethodAll four satisfy reward floorAll four satisfy reward floor and batch diversity ≥0.50\geq0.50All four satisfy reward floor and batch diversity ≥0.75\geq0.75
SatisDive, Δ=0.1\Delta=0.174.7%38.7%4.2%
SatisDive, Δ=0.2\Delta=0.271.7%48.1%7.0%
SatisDive, Δ=0.3\Delta=0.368.9%54.1%10.6%
SatisDive, Δ=0.5\Delta=0.563.1%53.5%11.6%
SatisDive, Δ=1.5\Delta=1.546.7%41.7%8.8%
FK steering58.9%5.6%0.6%
DAS59.3%0.0%0.0%

On SANA, with Δ=0.1,0.2,0.3,\Delta=0.1,0.2,0.3, or 0.50.5, SatisDive produces more batches where all four images meet the reward floor than FK steering or DAS. At all five tested tolerances, SatisDive produces more batches that meet both the reward and diversity requirements.

Question 3: “In one principal setting, can the authors provide a Δ\Delta sweep for the variant without JdJ_d, comparing its reward–diversity curve with the full method? This would help determine whether the proposed diversity objective materially improves the frontier. Please also clarify the mechanism responsible for diversification after the final replacement, when no further JdJ_d update occurs.”

Response: (a) Contribution of the diversity objective. We swept Δ\Delta with and without JdJ_d on the same 499 Pick-a-Pic prompts using SD1.5/ImageReward (Appendix D.17, Figure 9). Both versions use the same reward guidance and late-stage latent replacement. Only the diversity update changes.

Δ\DeltaFull: worst rewardNo JdJ_d: worst rewardFull: DreamSimNo JdJ_d: DreamSim
0.50.69200.70460.65700.6564
0.750.53810.55550.72290.7150
1.00.42580.44360.77230.7556
1.250.29500.34150.80040.7891

At Δ=1.0\Delta=1.0 and Δ=1.25\Delta=1.25, JdJ_d increases DreamSim but lowers worst-candidate reward. The curves remain close over their shared diversity range, indicating a modest contribution beyond reward guidance and replacement.

(b) Diversification after final replacement. FLUX's final scheduled update occurs at step 24 of 28. Replacement makes the two candidate latents identical, but each candidate receives different random noise during the remaining denoising steps. This noise can separate their trajectories without another JdJ_d update. We have clarified in the Method that diversity updates can follow earlier replacements, but not the final scheduled replacement.

AuthorsOpenReview ↗

Urgent Request to Reviewer

Dear Reviewer,

Thank you again for your thoughtful review and constructive feedback. As the discussion period approaches its end, we wanted to follow up on our rebuttal. We have clarified the scope of the theoretical guarantees and added batch-level satisfaction results, a fixed-floor comparison, and an ablation isolating the contribution of the diversity objective.

We would be grateful if you could review our responses and let us know whether they address your concerns or whether any further clarification would be helpful.

Thank you for your time and consideration.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Manuscript Revisions

We thank all reviewers for their careful reading and constructive feedback. We have revised the manuscript to address their comments, with additions and revised text highlighted in blue. The main updates are:

  • Theory and implementation: Clarified the theoretical guarantees and the roles of cutoff selection and gradient normalization (Section 3 and Appendix C).
  • Diversity-aware comparisons: Brought the CREPE and SGI comparisons into the main experiments to emphasize comparisons with diversity-aware methods (Section 4.1, Tables 4–5).
  • Expanded evaluation: Added average- and best-candidate reward curves, feature-layer and runtime studies, resampling and diversity-update ablations, whole-batch satisfaction measurements, and a fixed-floor comparison (Appendix D.13–D.19).
  • Non-differentiable rewards: Demonstrated an extension using a learned VQAScore predictor, with final images evaluated by VQAScore itself (Appendix D.20).

Our individual responses provide the corresponding explanations and results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.