Pith. sign in

REVIEW 4 major objections 3 minor 58 references

Progressive Seed Pruning scores half-denoised images and prunes low-ranked seeds early, outperforming best-of-N and search baselines at matched compute.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 06:59 UTC pith:BNKR3ALQ

load-bearing objection A simple, clearly-tested idea—prune seeds early using intermediate rewards—that mostly delivers, but the headline claims outrun the error bars. the 4 major comments →

arxiv 2607.21591 v1 pith:BNKR3ALQ submitted 2026-07-23 cs.CV

Inference-Time Scaling of Diffusion Models via Progressive Seed Pruning

Rogerio Guimaraes , Pietro Perona This is my paper
classification cs.CV
keywords diffusion modelsflow matchinginference-time scalingreward-guided generationprompt alignmentseed pruningprogressive seed pruningbest-of-N
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that the constant-memory constraint in inference-time scaling of diffusion models is unnecessary: if you are allowed to hold a variable number of candidates at different denoising stages, the best use of a fixed compute budget is to explore many noise seeds early and prune aggressively. The proposed method, Progressive Seed Pruning (PSP), starts with twice the seeds of best-of-N, scores cheap one-step clean-image estimates at 25% and 50% progress, and keeps only the top half each time. Across Stable Diffusion v1.5, SDXL, and Stable Diffusion 3.5, PSP reports higher GenEval prompt-alignment and better human evaluations than best-of-N, FK-Steering (importance sampling), and DSearch (tree search) at equal total denoising steps, and the advantage grows with compute. A sympathetic reader would care because it offers a training-free, gradient-free way to turn extra compute into better prompt alignment, and it works with deterministic samplers, which resampling-based methods cannot.

Core claim

PSP's claim is that intermediate rewards computed on predicted clean images x̂0 are informative early enough that a fixed geometric halving schedule—start with 2N seeds, prune to N at 25% progress, prune to N/2 at 50%—lets a fixed budget of N·T denoising steps behave like a best-of-2N search. The paper shows survival probabilities of the final best seed near 80–90% at the two prune points and regret near 0.01 relative to the 2N-seed upper bound, and that this regret decreases as the compute multiplier grows. On all three backbones PSP reports the best GenEval and human alignment scores among matched-compute baselines, and the authors argue PSP is less prone to reward hacking because it only

What carries the argument

The central object is the one-step denoised estimate x̂0(x_t, t, c), a predicted clean image that the generator's noise/velocity prediction already implies and that costs no extra model evaluations to compute. PSP scores this estimate with a black-box reward (ImageReward), prunes trajectories to the top-k by that score at predetermined progress points, and advances only survivors. The identity that carries the argument is the compute bookkeeping C = Σ k_t: front-loading the pool multiplies the number of evaluated seeds while keeping the total step count fixed, and the geometric schedule k_t = 2N → N → N/2 turns an N-fold budget into a 2N-fold seed pool.

Load-bearing premise

The method's gain rests on the assumption that a reward score computed on a half-denoised image ranks candidates almost as well as the reward on the final image, so that halving the pool at 25% and 50% progress rarely discards the eventual best seed.

What would settle it

Run the matched-compute comparison on a task whose decisive features are generated late, such as fine-grained texture or aesthetic detail: if the survival rate of the final-best seed under top-half pruning at 25%/50% progress falls toward chance, PSP's advantage over best-of-N should shrink or reverse. This can be measured directly from cached trajectories by computing the survival curve of Fig. S1 on such a reward.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Relaxing constant memory during diffusion inference is a legitimate scaling axis: a fixed halving schedule outperforms constant-memory baselines at matched compute on prompt alignment.
  • Deterministic samplers suffice; resampling-based methods need stochasticity that can degrade flow-matching generation, while PSP does not.
  • PSP's quality continues to increase as the compute multiplier grows from 2 to 16, with pruning regret shrinking relative to the upper bound set by all initial seeds.
  • Because sampling is deterministic, pruning schedules can be tuned offline by replaying cached intermediate rewards, without rerunning the generator.
  • The method generalizes to selecting among prompt rewrites for a fixed noise seed, and it remains beneficial on top of DPO-finetuned backbones.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the principle generalizes to any iterative generative process with a cheap intermediate proxy for the final sample; the same front-loading-then-prune allocation could be tested on autoregressive token generation or video generation.
  • Editorial inference: the measured correlation curves suggest an adaptive schedule guided by the reward's information profile could outperform a fixed geometric schedule when the informativeness of intermediate scores varies across tasks; this is a direct testable extension.
  • Editorial inference: the bandit-style connection to successive halving implies PSP's advantage is largest when the reward surface is seed-dominated and the best seed is rare; in domains where many seeds are nearly equivalent, the extra exploration buys little and BoN may match it.
  • Editorial inference: the reward-hacking argument—PSP never modifies trajectories—implies a testable prediction: on adversarial rewards that reward early-image artifacts, PSP should show lower reward-vs-alignment divergence than resampling methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 3 minor

Summary. The paper proposes Progressive Seed Pruning (PSP), an inference-time scaling method for diffusion and flow-matching text-to-image models. Under a fixed compute budget measured in denoiser forward calls, PSP initializes 2N noise seeds, denoises all seeds to a first pruning point, scores one-step denoised estimates with a black-box reward (ImageReward), prunes to N, later prunes to N/2, and fully denoises the survivors. Compute is matched to Best-of-N through C = Σ k_t (Eq. 4). Experiments on Stable Diffusion v1.5, SDXL, and Stable Diffusion 3.5 compare PSP with Best-of-N, FK-Steering, DSearch, NTS, RBF, BFS, and SVDD, reporting ImageReward, HPS v2, GenEval, and human prompt-alignment scores. The central claim is that PSP consistently improves reward-guided selection and achieves higher GenEval and human-evaluation scores than these baselines at matched compute, and that performance continues to scale with compute.

Significance. If the empirical claims hold, the paper identifies a simple and practical design principle—variable particle counts with early reward-based pruning—that can improve over constant-memory seed-search methods at a fixed denoiser budget. The strengths include transparent compute accounting via Eq. (4), the use of deterministic solvers with cached intermediate rewards, which enables cheap offline schedule search, comparisons against six baselines, and a human evaluation whose protocol is validated against prior work. The paper also states its limitations honestly in Section 5. However, the decisive evidence needs strengthening: Table 1 has no variance or significance information, several headline numbers are ties or not best, and the core pruning assumption is supported only by per-prune statistics rather than the full sequential schedule. These gaps are fixable and the method is likely useful, but the current support is weaker than the abstract implies.

major comments (4)
  1. [§4.2, Table 1] The abstract and the text in §4.2/§5 state that PSP "consistently improves reward-guided selection" and "achieves higher GenEval scores" than baselines at matched compute. Table 1 does not support these statements as written: on SD 3.5, GenEval for PSP is 0.747, identical to BoN's 0.747 and BFS's 0.747; on SDXL, BFS obtains a higher ImageReward (1.247 vs 1.224) and NTS/RBF/DSearch obtain higher HPS. Results are averaged over 3 random seeds but no error bars, confidence intervals, or significance tests are reported. Please either correct the claims or provide paired significance tests and variances for the near-tie SD 3.5 GenEval comparisons and the SDXL IR/HPS comparisons.
  2. [§4.2, Fig. S1] The core mechanism is that pruning at 25% and 50% progress discards the eventual best seed rarely enough. Fig. S1 reports only per-prune statistics: survival of the final best sample for a single K→K/2 prune (~80% at 25% for K=8, ~90% at 50% for K=4) and per-prune regret. The actual default schedule is sequential 8→4→2, so the quantity that determines the output is the joint probability that the best of the initial 8 seeds survives both prunes, and the cumulative regret after both prunes. If the two pruning events were independent, joint survival would be at most 0.72; correlation could make it lower. Report end-to-end survival and cumulative regret for the full default schedule, and for the scaled schedules used in §4.4.
  3. [§4.2, Algorithm 1] There is no ablation that replaces reward-based pruning with random pruning at the same 8→4→2 schedule and the same initial seed pool. Such a control is necessary to isolate the contribution of the intermediate reward signal from the effect of starting with a larger pool and using a variable-particle-count schedule. Without it, the comparison to BoN, FK-Steering, and DSearch does not establish that the informativeness of early x̂0 rewards drives the gains, which is the stated mechanism of Algorithm 1. Given the cached trajectories used in §4.6, this ablation is cheap and directly tests the paper's central mechanism.
  4. [§4.4, Fig. 3] The scaling claim ('performance scales with compute') is demonstrated only on ImageReward. Fig. S2 (right) shows that IR saturates with respect to GenEval, and Table 1 already shows no GenEval improvement for SD 3.5 at N=4. To substantiate that scaling transfers to prompt alignment, report GenEval or another reward-agnostic metric for the ¯N=2,4,8,16 sweeps, not only IR. Otherwise the scaling conclusion applies only to the guidance reward, which is weaker than the claim in the abstract.
minor comments (3)
  1. [§4.3] Human evaluation uses a single seed per prompt/method/backbone triplet while automated metrics are averaged over 3 seeds. Please report whether the same seed is used across methods for each prompt and provide per-method confidence intervals or a significance test for the human alignment differences; the differences between PSP and BoN/FK-Steering on SD 3.5 are small (0.841 vs 0.831/0.837).
  2. [§4.1] For SD 3.5, FK-Steering uses a tuned stochasticity γ=0.005, selected as the largest value that does not degrade baseline quality. Since PSP is deterministic, this hyperparameter choice affects the fairness of the comparison; a brief sensitivity analysis of γ (e.g., 0 vs 0.005 vs 0.01) would help.
  3. [§4.6] The claim that tuned schedules show 'little to no improvement' is based on Table S2, but the default schedule is itself a tuned choice (halving with two prune points). Consider rephrasing to 'the simple default is competitive with the best schedule found by grid search on this benchmark' to avoid implying the default was not selected with knowledge of its performance.

Circularity Check

0 steps flagged

No significant circularity; only minor self-referential reporting, no load-bearing reduction.

full rationale

PSP's central claim is empirical and is evaluated against metrics external to the selection signal: GenEval detector-based checks, HPSv2, and human prompt-alignment ratings, in addition to the ImageReward (IR) used for pruning. Using IR both as the selection reward and as one reported metric is self-referential as a sanity check, but it is not the basis of the claim; notably, BFS exceeds PSP on IR for SDXL while PSP wins on GenEval, so the IR column is not forced by construction. The default schedule (8 -> 4 at 25% progress, 4 -> 2 at 50%) is fixed a priori by geometric halving and compute matching (Sec. 4.6), not fitted to the test set; schedule tuning is performed on a separate Benchmark IR split and transferred to GenEval via offline replay (Sec. S5), which is standard held-out validation. The core assumption that intermediate xhat0 rewards rank candidates well is supported directly by correlation, survival, and regret measurements (Fig. S1), not assumed by definition. The only self-citation ([15], Guimaraes et al.) appears in related work to note that coarse early features are exploited in perception tasks; the property itself is attributed to external [31] and is not load-bearing for the main result. No equation in the paper reduces a prediction to a fitted constant or to a self-citation chain.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim rests on empirical assumptions about reward informativeness, determinism, compute accounting, and baseline fairness, all stated in the paper. No new physical entities are introduced; the only free choices are the pruning schedule, tuned schedule variants, and baseline hyperparameters.

free parameters (4)
  • PSP default pruning schedule = SD1.5/SDXL: k_init=8, t=(16,32), k=(4,2), T=64; SD3.5: k_init=8, t=(8,16), k=(4,2), T=32
    Hand-chosen halving schedule that matches compute multiplier N=4. Not fitted to data for the main Table 1, but the method's reported performance depends on it (Sec. 4.6, S7).
  • PSP tuned schedules per model/reward = e.g. SDXL(IR): k_init=16,t=(4,8,24,40),k=(8,4,3,2); SD3.5: k_init=8,t=(8,12,22,30),k=(4,3,2,1)
    Selected by grid search on the held-out BenchmarkIR set (S5). Used for tuned results in Table S2 and Fig. 4, not for the main Table 1. Legitimate model selection, but still a fitted hyperparameter.
  • SD3.5 baseline stochasticity gamma = 0.005 (clamped to [0, sqrt(2)-1])
    Tuned on the held-out IR benchmark as the largest noise level that does not degrade baseline generation quality; affects fairness of FK-Steering/DSearch/BFS/SVDD on SD3.5 (S4).
  • Baseline hyperparameter choices = FK-Steering lambda=10,K=4; DSearch w=2, oversample_rate=2; NTS eps=0.4,lambda=0.15; RBF max_nfe=256; SVDD num_particles=
    Chosen to match compute or follow author recommendations (S7). The central claim 'PSP beats baselines' is contingent on these choices, especially for stochasticity-sensitive flow matching.
axioms (5)
  • domain assumption Intermediate denoised estimates x̂0, scored by the reward r, rank-order final outputs well enough that early pruning to top half at 25%/50% progress discards the eventual best seed only rarely.
    Algorithm 1 prunes purely on s_t = r(x̂0(x_t,t,c)); Fig. S1 shows survival ~80% at 25% and ~90% at 50% prune, regret ~0.01. This is empirical support, not a proof, and it is load-bearing.
  • domain assumption All randomness for PSP arises from the initial noise seed because deterministic solvers (DDIM eta=0, Euler discrete) are used.
    Sec. 4.1 states this; PSP's seed-search framing and offline schedule replay depend on deterministic trajectories.
  • domain assumption ImageReward is a valid black-box reward for prompt alignment, and its scores on intermediate x̂0 estimates are meaningful.
    ImageReward is the guidance signal throughout (Tab. 1, S1); its correlation with final reward is documented in Fig. S1, but this is an empirical property of the chosen reward model, not a given.
  • domain assumption Total inference compute is adequately measured by the number of denoising steps C = Σ k_t.
    Sec. 3.3 defines the budget this way; Table 2 shows PSP incurs extra VAE-decode and reward-scoring overhead, so 'matched compute' does not include all real costs.
  • domain assumption The public implementations and chosen hyperparameters of FK-Steering, DSearch, NTS, RBF, BFS, SVDD are faithful and fairly configured.
    S7 lists many baseline hyperparameters; the matched-compute comparison depends on them, especially since stochasticity can degrade flow-matching quality.

pith-pipeline@v1.3.0-alltime-deepseek · 24362 in / 16955 out tokens · 155518 ms · 2026-08-01T06:59:22.097693+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Inference-Time Scaling of Diffusion Models via Progressive Seed Pruning." pith.science (2026). https://pith.science/paper/BNKR3ALQ

@misc{pith2026260721591,
  author       = {Pith},
  title        = {Pith review of: Inference-Time Scaling of Diffusion Models via Progressive Seed Pruning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BNKR3ALQ}},
  note         = {Machine review of arXiv:2607.21591}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Diffusion and flow-matching models dominate conditional image generation, yet inference-time scaling for these models is far less developed than for autoregressive language models. Because final quality is highly sensitive to the initial noise seed, many approaches spend extra compute on seed search or resampling under a black-box reward, but typically maintaining a constant memory footprint throughout inference. We show that relaxing this constraint enables an underexplored inference-time scaling axis: by front-loading exploration, evaluating many seeds early, and pruning aggressively, we can use a fixed compute budget more effectively. \emph{Progressive Seed Pruning} (\PSP) scores intermediate denoised estimates and progressively narrows the candidate set so that only promising trajectories are fully denoised, while keeping the total number of model evaluations fixed. Across diffusion and flow-matching backbones, \PSP \ consistently improves reward-guided selection and achieves higher GenEval scores (automated) and better human evaluation on prompt-alignment than best-of-$N$, importance-sampling, and tree-search baselines at matched compute. Project page: https://www.vision.caltech.edu/psp. Code: https://github.com/rogerioagjr/psp.

Figures

Figures reproduced from arXiv: 2607.21591 by Pietro Perona, Rogerio Guimaraes.

Figure 1
Figure 1. Figure 1: Examples of image generation improvement with [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Compute allocation profiles for inference-time scaling. Top [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Scaling behavior of PSP. Left: final guidance reward versus compute mul￾tiplier N¯, comparing PSP to BoNat matched compute. Middle: regret of PSP rel￾ative to BoN with double compute that would select the best seed among all 2N¯ candidates after full denoising; regret decreases with N¯, indicating pruning becomes increasingly safe at larger budgets. Right: reward versus approximate FLOPs, showing that infe… view at source ↗
Figure 4
Figure 4. Figure 4: Tuning the pruning strategy improves PSP. Left. We compare our de￾fault, off-the-shelf schedule (solid) to the best schedule found by grid search under the same compute budget (hatched). We search on prompts from the IR Benchmark and report results on prompts from GenEval, using IR as guidance signal. Right. Visual￾ization of all schedules in the grid search colored by their performance on Benchmark IR. Be… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

58 extracted references · 2 canonical work pages

  1. [1]

    In: International Conference on Learning Representations

    Bansal,A.,Chu,H.M.,Schwarzschild,A.,Sengupta,R.,Goldblum,M.,Geiping,J., Goldstein, T.: Universal guidance for diffusion models. In: International Conference on Learning Representations. vol. 2024, pp. 51304–51323 (2024)

    2024

  2. [2]

    In: Proceed- ings of the AAAI conference on artificial intelligence

    Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Podstawski, M., Giani- nazzi, L., Gajda, J., Lehmann, T., Niewiadomski, H., Nyczyk, P., et al.: Graph of thoughts: Solving elaborate problems with large language models. In: Proceed- ings of the AAAI conference on artificial intelligence. vol. 38, pp. 17682–17690 (2024)

    2024

  3. [3]

    com/papers/dall-e-3.pdf, accessed: July 22, 2026

    Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., Manassra, W., Dhariwal, P., Chu, C., Jiao, Y., Ramesh, A.: Improving image generation with better captions (2023),https://cdn.openai. com/papers/dall-e-3.pdf, accessed: July 22, 2026

    2023

  4. [4]

    In: International Conference on Learning Represen- tations

    Black, K., Janner, M., Du, Y., Kostrikov, I., Levine, S.: Training diffusion models with reinforcement learning. In: International Conference on Learning Represen- tations. vol. 2024, pp. 4965–4987 (2024)

    2024

  5. [5]

    Finance and Stochastics13(4), 613–633 (Sep 2009).https://doi.org/10.1007/s00780-009-0098-8,http://link.springer

    Carmona, R., Fouque, J.P., Vestal, D.: Interacting particle systems for the compu- tation of rare credit portfolio losses. Finance and Stochastics13(4), 613–633 (Sep 2009).https://doi.org/10.1007/s00780-009-0098-8,http://link.springer. com/10.1007/s00780-009-0098-8

    work page doi:10.1007/s00780-009-0098-8 2009

  6. [6]

    In: The Eleventh International Con- ference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023

    Chung, H., Kim, J., McCann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sampling for general noisy inverse problems. In: The Eleventh International Con- ference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net (2023),https://openreview.net/forum?id=OnD9zGAGT0k

    2023

  7. [7]

    CoRRabs/2110.14168(2021),https://arxiv

    Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., Schulman, J.: Training verifiers to solve math word problems. CoRRabs/2110.14168(2021),https://arxiv. org/abs/2110.14168

    Pith/arXiv arXiv 2021

  8. [8]

    Probability and its Applications, Springer New York, New York, NY (2004).https://doi.org/10.1007/978-1-4684-9393- 1,http://link.springer.com/10.1007/978-1-4684-9393-1

    Del Moral, P.: Feynman-Kac Formulae. Probability and its Applications, Springer New York, New York, NY (2004).https://doi.org/10.1007/978-1-4684-9393- 1,http://link.springer.com/10.1007/978-1-4684-9393-1

    work page doi:10.1007/978-1-4684-9393- 2004

  9. [9]

    CoRRabs/2505.14683(2025).https://doi.org/10.48550/ARXIV.2505

    Deng, C., Zhu, D., Li, K., Gou, C., Li, F., Wang, Z., Zhong, S., Yu, W., Nie, X., Song, Z., Guang, S., Fan, H.: Emerging properties in unified multimodal pretrain- ing. CoRRabs/2505.14683(2025).https://doi.org/10.48550/ARXIV.2505. 14683,https://doi.org/10.48550/arXiv.2505.14683

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.14683 2025

  10. [10]

    Statistics and Computing31(6), 81 (2021)

    Elvira, V., Miguez, J., Djurić, P.M.: On the performance of particle filters with adaptive number of particles. Statistics and Computing31(6), 81 (2021)

    2021

  11. [11]

    In: Forty-first international conference on machine learning (2024) Inference-Time Scaling of Diffusion Models via Progressive Seed Pruning 17

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., Müller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al.: Scaling rectified flow transformers for high-resolution image synthesis. In: Forty-first international conference on machine learning (2024) Inference-Time Scaling of Diffusion Models via Progressive Seed Pruning 17

    2024

  12. [12]

    Euler, L.: Institutiones calculi integralis, vol. 1. impensis Academiae imperialis scientiarum (1792)

  13. [13]

    Advances in neural information processing systems14(2001)

    Fox, D.: KLD-sampling: Adaptive particle filters. Advances in neural information processing systems14(2001)

    2001

  14. [14]

    Advances in Neural Information Processing Systems36, 52132–52152 (2023)

    Ghosh, D., Hajishirzi, H., Schmidt, L.: Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems36, 52132–52152 (2023)

    2023

  15. [15]

    In: IEEE/CVF Winter Conference on Appli- cations of Computer Vision, WACV 2026, Tucson, AZ, USA, March 6-10, 2026

    Guimarães, R., Xiao, F., Perona, P., Marks, M.: Diffusion-based action recognition generalizes to untrained domains. In: IEEE/CVF Winter Conference on Appli- cations of Computer Vision, WACV 2026, Tucson, AZ, USA, March 6-10, 2026. pp. 5919–5933. IEEE (2026).https://doi.org/10.1109/WACV61042.2026.00573, https://doi.org/10.1109/WACV61042.2026.00573

    arXiv 2026

  16. [16]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

    2020

  17. [17]

    In: Artificial intelligence and statistics

    Jamieson, K., Talwalkar, A.: Non-stochastic best arm identification and hyperpa- rameter optimization. In: Artificial intelligence and statistics. pp. 240–248. PMLR (2016)

    2016

  18. [18]

    CoRRabs/2512.16853(2025).https://doi.org/10.48550/ARXIV.2512.16853, https://doi.org/10.48550/arXiv.2512.16853

    Kamath, A., Chang, K., Krishna, R., Zettlemoyer, L., Hu, Y., Ghazvinine- jad, M.: Geneval 2: Addressing benchmark drift in text-to-image evaluation. CoRRabs/2512.16853(2025).https://doi.org/10.48550/ARXIV.2512.16853, https://doi.org/10.48550/arXiv.2512.16853

    doi:10.48550/arxiv.2512.16853 2025

  19. [19]

    In: Dasgupta, S., McAllester, D

    Karnin, Z., Koren, T., Somekh, O.: Almost optimal exploration in multi-armed bandits. In: Dasgupta, S., McAllester, D. (eds.) Proceedings of the 30th inter- national conference on machine learning. Proceedings of machine learning re- search, vol. 28, pp. 1238–1246. PMLR, Atlanta, Georgia, USA (Jun 2013),https: //proceedings.mlr.press/v28/karnin13.html, number: 3

    2013

  20. [20]

    Advances in Neural Information Processing Systems38, 30830–30864 (2026)

    Kim, J., Yoon, T., Hwang, J., Sung, M.: Inference-time scaling for flow models via stochastic generation and rollover budget forcing. Advances in Neural Information Processing Systems38, 30830–30864 (2026)

    2026

  21. [21]

    Advances in neural information processing systems36, 36652–36663 (2023)

    Kirstain,Y.,Polyak,A.,Singer,U.,Matiana,S.,Penna,J.,Levy,O.:Pick-a-pic:An open dataset of user preferences for text-to-image generation. Advances in neural information processing systems36, 36652–36663 (2023)

    2023

  22. [22]

    Lee, G., Bao, T.N.N., Yoon, J., Lee, D., Kim, M., Bengio, Y., Ahn, S.: Adap- tive Inference-Time Scaling via Cyclic Diffusion Search (Oct 2025).https:// doi.org/10.48550/arXiv.2505.14036,http://arxiv.org/abs/2505.14036, arXiv:2505.14036 [cs]

    doi:10.48550/arxiv.2505.14036 2025

  23. [23]

    CoRRabs/2503.02039(2025).https://doi.org/10.48550/ARXIV.2503.02039, https://doi.org/10.48550/arXiv.2503.02039

    Li, X., Uehara, M., Su, X., Scalia, G., Biancalani, T., Regev, A., Levine, S., Ji, S.: Dynamic search for inference-time alignment in diffusion models. CoRRabs/2503.02039(2025).https://doi.org/10.48550/ARXIV.2503.02039, https://doi.org/10.48550/arXiv.2503.02039

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.02039 2025

  24. [24]

    In: Belgrave, D., Zhang, C., Montoya, L.N., Lin, H., Pascanu, R., Koniusz, P., Ghassemi, M., Chen, N., Ruíz, I.V.M., Loaiza-Bonilla, A

    Li, X., Zhao, Y., Wang, C., Scalia, G., Eraslan, G., Nair, S., Biancalani, T., Ji, S., Regev, A., Levine, S., Uehara, M.: Derivative-free guidance in continuous and discrete diffusion models with soft value-based decoding. In: Belgrave, D., Zhang, C., Montoya, L.N., Lin, H., Pascanu, R., Koniusz, P., Ghassemi, M., Chen, N., Ruíz, I.V.M., Loaiza-Bonilla, A...

    2025

  25. [25]

    In: The Eleventh International Conference on Learning Rep- resentations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023

    Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: The Eleventh International Conference on Learning Rep- resentations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net (2023), https://openreview.net/forum?id=PqvMRDCJT9t

    2023

  26. [26]

    In: The Eleventh International Conference on Learn- ing Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023

    Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and trans- fer data with rectified flow. In: The Eleventh International Conference on Learn- ing Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net (2023),https://openreview.net/forum?id=XVjTT1nw5z

    2023

  27. [27]

    In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S

    Luo, G., Dunlap, L., Park, D.H., Holynski, A., Darrell, T.: Diffusion hyperfea- tures: Searching through time and space for semantic correspondence. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems 36: Annual Conference on Neural Informa- tion Processing Systems 2023, NeurIP...

    2023

  28. [28]

    CoRRabs/2501.09732(2025).https://doi.org/10.48550/ ARXIV.2501.09732,https://doi.org/10.48550/arXiv.2501.09732

    Ma, N., Tong, S., Jia, H., Hu, H., Su, Y., Zhang, M., Yang, X., Li, Y., Jaakkola, T.S., Jia, X., Xie, S.: Inference-time scaling for diffusion models beyond scaling denoising steps. CoRRabs/2501.09732(2025).https://doi.org/10.48550/ ARXIV.2501.09732,https://doi.org/10.48550/arXiv.2501.09732

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.09732 2025

  29. [29]

    CoRRabs/2601.03267(2026).https: //doi.org/10.48550/ARXIV.2601.03267,https://doi.org/10.48550/arXiv

    OpenAI: Openai GPT-5 system card. CoRRabs/2601.03267(2026).https: //doi.org/10.48550/ARXIV.2601.03267,https://doi.org/10.48550/arXiv. 2601.03267

    Pith/arXiv arXiv doi:10.48550/arxiv.2601.03267 2026

  30. [30]

    Advances in Neural Information Processing Systems38, 13170–13216 (2026)

    Oshima, Y., Suzuki, M., Matsuo, Y., Furuta, H.: Inference-time text-to-video align- ment with diffusion latent beam search. Advances in Neural Information Processing Systems38, 13170–13216 (2026)

    2026

  31. [31]

    In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S

    Park, Y., Kwon, M., Choi, J., Jo, J., Uh, Y.: Understanding the latent space of diffusion models through the lens of riemannian geometry. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neu- ral Information Processing Systems 36: Annual Conference on Neural Informa- tion Processing Systems 2023, NeurIPS 2023, N...

    2023

  32. [32]

    In: The Twelfth International Conference on Learning Represen- tations, ICLR 2024, Vienna, Austria, May 7-11, 2024

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: SDXL: improving latent diffusion models for high-resolution im- age synthesis. In: The Twelfth International Conference on Learning Represen- tations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net (2024), https://openreview.net/forum?id=di52zR8xgf

    2024

  33. [33]

    CoRRabs/2407.14041(2024).https://doi.org/ 10.48550/ARXIV.2407.14041,https://doi.org/10.48550/arXiv.2407.14041

    Qi, Z., Bai, L., Xiong, H., Xie, Z.: Not all noises are created equally:diffusion noise selection and optimization. CoRRabs/2407.14041(2024).https://doi.org/ 10.48550/ARXIV.2407.14041,https://doi.org/10.48550/arXiv.2407.14041

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.14041 2024

  34. [34]

    Advances in Neural Information Processing Systems38, 87284–87317 (2026)

    Ramesh, V., Mardani, M.: Test-time scaling of diffusion models via noise trajec- tory search. Advances in Neural Information Processing Systems38, 87284–87317 (2026)

    2026

  35. [35]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022)

    2022

  36. [36]

    In: Singh, A., Fazel, M., Hsu, D., Lacoste-Julien, S., Berkenkamp, F., Maharaj, T., Wagstaff, K., Zhu, J

    Singhal, R., Horvitz, Z., Teehan, R., Ren, M., Yu, Z., Mckeown, K., Ranganath, R.: A general framework for inference-time scaling and steering of diffusion models. In: Singh, A., Fazel, M., Hsu, D., Lacoste-Julien, S., Berkenkamp, F., Maharaj, T., Wagstaff, K., Zhu, J. (eds.) Proceedings of the 42nd international conference on Inference-Time Scaling of Di...

    2025

  37. [37]

    Advances in Neural Information Processing Systems35, 9460–9471 (2022)

    Skalse, J., Howe, N., Krasheninnikov, D., Krueger, D.: Defining and characterizing reward gaming. Advances in Neural Information Processing Systems35, 9460–9471 (2022)

    2022

  38. [38]

    In: International conference on machine learning

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsuper- vised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256–2265. pmlr (2015)

    2015

  39. [39]

    In: 9th Inter- national Conference on Learning Representations, ICLR 2021, Virtual Event, Aus- tria, May 3-7, 2021

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: 9th Inter- national Conference on Learning Representations, ICLR 2021, Virtual Event, Aus- tria, May 3-7, 2021. OpenReview.net (2021),https://openreview.net/forum? id=St1giarCHLP

    2021

  40. [40]

    In: 9th Inter- national Conference on Learning Representations, ICLR 2021, Virtual Event, Aus- tria, May 3-7, 2021

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. In: 9th Inter- national Conference on Learning Representations, ICLR 2021, Virtual Event, Aus- tria, May 3-7, 2021. OpenReview.net (2021),https://openreview.net/forum? id=PxTIG12RRHS

    2021

  41. [41]

    In: Singh, A., Fazel, M., Hsu, D., Lacoste-Julien, S., Berkenkamp, F., Maharaj, T., Wagstaff, K., Zhu, J

    Tang, Z., Peng, J., Tang, J., Hong, M., Wang, F., Chang, T.: Inference-time align- ment of diffusion models with direct noise optimization. In: Singh, A., Fazel, M., Hsu, D., Lacoste-Julien, S., Berkenkamp, F., Maharaj, T., Wagstaff, K., Zhu, J. (eds.) Forty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19,...

    2025

  42. [42]

    Advances in neural information processing systems36, 75993–76005 (2023)

    Valmeekam, K., Marquez, M., Sreedharan, S., Kambhampati, S.: On the plan- ning abilities of large language models-a critical investigation. Advances in neural information processing systems36, 75993–76005 (2023)

    2023

  43. [43]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wallace, B., Dang, M., Rafailov, R., Zhou, L., Lou, A., Purushwalkam, S., Ermon, S., Xiong, C., Joty, S., Naik, N.: Diffusion model alignment using direct preference optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8228–8238 (2024)

    2024

  44. [44]

    Advances in Neural Infor- mation Processing Systems36, 31372–31403 (2023)

    Wu, L., Trippe, B., Naesseth, C., Blei, D., Cunningham, J.P.: Practical and asymp- totically exact conditional sampling in diffusion models. Advances in Neural Infor- mation Processing Systems36, 31372–31403 (2023)

    2023

  45. [45]

    CoRRabs/2306.09341(2023).https://doi.org/10.48550/ARXIV

    Wu, X., Hao, Y., Sun, K., Chen, Y., Zhu, F., Zhao, R., Li, H.: Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. CoRRabs/2306.09341(2023).https://doi.org/10.48550/ARXIV. 2306.09341,https://doi.org/10.48550/arXiv.2306.09341

    Pith/arXiv arXiv doi:10.48550/arxiv.2306.09341 2023

  46. [46]

    Advances in Neural Information Processing Systems36, 15903–15935 (2023)

    Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagere- ward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems36, 15903–15935 (2023)

    2023

  47. [47]

    In: 2025 IEEE/CVF Winter Conference on Ap- plications of Computer Vision (WACV)

    Xu, K., Zhang, L., Shi, J.: Good seed makes a good crop: Discovering secret seeds in text-to-image diffusion models. In: 2025 IEEE/CVF Winter Conference on Ap- plications of Computer Vision (WACV). pp. 3024–3034. IEEE (2025)

    2025

  48. [48]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Yang, K., Tao, J., Lyu, J., Ge, C., Chen, J., Shen, W., Zhu, X., Li, X.: Using human feedback to fine-tune diffusion models without any reward model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8941–8951 (2024) 20 R. Guimarães and P. Perona

    2024

  49. [49]

    Advances in Neural Information Processing Systems35, 20744–20757 (2022)

    Yao, S., Chen, H., Yang, J., Narasimhan, K.: Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems35, 20744–20757 (2022)

    2022

  50. [50]

    Advances in neural information processing systems36, 11809–11822 (2023)

    Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T., Cao, Y., Narasimhan, K.: Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems36, 11809–11822 (2023)

    2023

  51. [51]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3836–3847 (2023)

    2023

  52. [52]

    prompt_id

    Zhang, X., Lin, H., Ye, H., Zou, J.Y., Ma, J., Liang, Y., Du, Y.: Inference-time scal- ing of diffusion models through classical search. CoRRabs/2505.23614(2025). https://doi.org/10.48550/ARXIV.2505.23614,https://doi.org/10.48550/ arXiv.2505.23614 Inference-Time Scaling of Diffusion Models via Progressive Seed Pruning 21 Supplemental Material S1 Analysis ...

    doi:10.48550/arxiv.2505.23614 2025

  53. [53]

    Keep constraints identical to original entry: - same object classes - same counts - same color requirements - same positional relations (left/right/above/below) when present - never add contradictory constraints

  54. [54]

    Create exactly 24 expansions for each prompt_id

  55. [55]

    Use unique prompt_expansion_id values 0..23

  56. [56]

    Make expansions meaningfully different from each other: - vary composition, camera framing, distance, angle, lighting, scene context, background style - do NOT change required semantic constraints

  57. [57]

    Prompt length: - target 35-45 words - hard cap 50 words per expanded prompt

  58. [58]

    a photo of a cow left of a stop sign

    Output strictly valid JSON only (no markdown, no commentary). Diversity guidance: - Mix close-up, medium, wide shots. - Vary environment (studio, outdoor, indoor, urban, natural) where compatible. - Vary lighting (soft daylight, overcast, warm indoor, dramatic side light), without changing semantics. - Vary descriptive style while preserving all required ...