Pith. sign in

REVIEW 4 major objections 8 minor 112 references

Ctrl-Z Sampling: Scaling Diffusion Sampling with Controlled Random Zigzag Explorations

T0 review · 4 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Ctrl-Z Sampling claims that diffusion denoising stalls on visually plausible but misaligned outputs, and that rolling back to noisier states with escalating inversion depth until a reward-score improvement is found consistently raises…

desk verdict A plausible adaptive-depth sampler for diffusion models; the core mechanism is new, but the 'consistently improves' claim outruns the evidence. read the letter →

arxiv 2506.20294 v5 pith:I3PMCKTZ submitted 2025-06-25 cs.CV

Shunqi Mao , Wei Guo , Chaoyi Zhang , Jieting Long , Ke Xie , Weidong Cai This is my paper
classification cs.CV
keywords diffusionmodelsinference-timescalingreward-guidedsamplingtest-timecomputelatentinversiontext-to-imagegenerationqualitylandscapeadaptiveexploration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that conditional diffusion sampling can stall in visually plausible but semantically misaligned states, and that such stalls are best treated as plateaus in a surrogate quality landscape. The proposed Ctrl-Z Sampling detects these plateaus with a reward model, then rolls the trajectory back into noisier regions with increasing inversion depth until a candidate that improves the score is found. The authors report that this adaptive, reward-gated exploration consistently beats fixed-depth search baselines such as Search-over-Path at comparable number of function evaluations, and that it works across U-Net and Transformer diffusion backbones.

What carries the argument

The load-bearing mechanism is the combination of three pieces: a plateau detector that compares the reward model score of the current clean estimate to the most recently accepted score plus a threshold $\delta$; a controlled latent inversion operator $\Psi(x_t, \Delta;\epsilon)$ that re-noises the latent back to step $t+\Delta$ while preserving accumulated structure; and an adaptive escalation loop that increases $\Delta$ (up to $d_{\max}$) and samples $N$ candidate continuations per depth until a candidate satisfies the acceptance criterion. The work is done by concentrating exploration at detected stagnation points and deepening it only when shallow attempts fail, rather than perturbing every step uniformly.

What would settle it

Run Ctrl-Z on the same prompts with the reward model's scores shuffled or sign-flipped while keeping the rest of the algorithm fixed; if human-preference or benchmark metrics still improve over DDIM, then the plateau-detection signal is not driving the gains. A second, cleaner test: compare Ctrl-Z against a fixed-depth rollback baseline with matched NFE; if the fixed-depth version matches its quality, then adaptive depth escalation is not the reason for the improvement.

Watch

Extended reading notes

Core claim

The central discovery is that spending extra inference compute on escape attempts only when progress stagnates, and deepening those attempts adaptively, yields better quality per compute than spending compute on many shallow perturbations at every step. Ctrl-Z treats the denoising trajectory as a hill-climbing process on a surrogate reward: at each step the clean estimate $\hat{x}^{t-1}_0$ is scored, and when $R(c,\hat{x}^{t-1}_0) < r_{\text{prev}} + \delta$ (Equation 5) the sampler injects noise via the inversion operator $\Psi(x_t, \Delta; \epsilon)$ (Equation 4), re-runs forward denoising for a set of candidates, and keeps the best-scoring candidate; if none beats the threshold, the inversion depth $\Delta$ is increased and the search repeats up to $d_{\max}$. The paper demonstrates on Pick-a-Pic, DrawBench, and T2I-CompBench that this controlled zigzag trajectory improves alignment and fidelity metrics over DDIM, Resampling, Z-Sampling, and SOP at similar NFE budgets.

Load-bearing premise

The method assumes the reward model evaluated on intermediate clean estimates is a faithful proxy for final image quality, so that a plateau in that score really means the generation is stuck and a score improvement really means a better image.

Editorial extensions

If this is right

  • At roughly 3x NFE, Ctrl-Z (with $\lambda=30$, $N=2$, $d_{\max}=3$) exceeds all compared baselines on most metrics; at roughly 7-9x NFE it improves further, so the method offers a tunable compute-quality trade-off.
  • Because exploration is triggered only on detected plateaus, Ctrl-Z uses fewer function evaluations than 'always explore' strategies while reaching comparable quality, as shown by the initiation-criteria ablations.
  • Depth and width complement each other, and deeper-but-narrower exploration often outperforms wider-but-shallow search at matched NFE, supporting the adaptive-depth design.
  • The method is compatible across U-Net (Stable Diffusion 2.1) and Transformer (Hunyuan-DiT) diffusion backbones, and can be combined with orthogonal samplers such as AYS and CFG++.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The plateau-and-escalate recipe is generic: any generative model with an intermediate-state proxy score and a controllable 'un-doing' operator could use the same trigger-and-deepen loop, so similar schemes may transfer to flow matching, video generation, or even LLM decoding with process reward models.
  • A fixed acceptance threshold $\delta$ is a known weak point the authors flag; a testable extension would replace it with a schedule that tracks reward gradients or normalizes improvements, which could suppress the redundant late-stage explorations the paper observes.
  • The reliance on a single reward model leaves open reward-overoptimization risk; using an ensemble of scorers or a conservatively calibrated threshold could make the quality gains more robust across metrics, particularly aesthetic score.
  • Since the paper only tests moderate budgets, a natural scaling question is whether the depth-escalation strategy continues to beat wider searches at very large NFE, or whether a hybrid tree search would overtake it.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 8 minor

Summary. This paper proposes Ctrl-Z Sampling, an inference-time sampling method for conditional diffusion text-to-image models. A reward model R scores the intermediate clean estimate x̂_0 at each denoising step; when the score fails to improve by a threshold δ, the method inverts the current latent to a noisier state with random noise, re-denoises it, scores N candidates, accepts the best if it clears the threshold, and otherwise escalates the inversion depth up to d_max. Experiments on Pick-a-Pic, DrawBench, and T2I-CompBench with Stable Diffusion 2.1 and Hunyuan-DiT compare Ctrl-Z against DDIM, Resampling, Z-Sampling, and SOP, and ablations cover exploration depth, width, acceptance threshold, window λ, the choice of reward model, and the exploration initiation criterion. The paper claims consistent quality gains across NFE budgets and releases code.

Significance. The core idea — detecting a surrogate-score plateau and responding with adaptively deepening inversions — is a clear and reasonable progression beyond SOP's fixed-depth search, and the adaptive depth escalation with a bounded candidate budget is a sensible compute-quality trade-off. Strengths include a well-specified algorithm, ablations over all core hyperparameters (Tables 3, 5, 6; Figures 3, 5), a reward-model ablation (Table 3) that makes part of the evaluation non-circular, an analysis of where explorations trigger (Figure 6), integrations with orthogonal techniques (AYS/DPM and CFG++ in the appendix), and a public code release. The manuscript is also candid about its limitations (variable NFE and AES reduction in Supp. C.1; reward overfitting in Supp. D.1). If the empirical claims survive re-measurement with variance reporting and matched budgets, this would be a useful contribution to inference-time scaling for diffusion models; at present, the headline 'consistently improves' claim is stronger than the reported evidence.

major comments (4)
  1. [Section 5.2, Tables 1–2] The abstract's claim that Ctrl-Z Sampling 'consistently improves generation quality over other inference-time scaling samplers across different NFE budgets' is not supported by the reported point estimates. At comparable budgets, SOP-4 scores higher than Ctrl-Z‡ on several cells: Pick-a-Pic PickScore on SD2.1 (21.12 vs 21.02), DrawBench AES on Hunyuan-DiT (5.942 vs 5.908), Hy-DiT Numeracy in Table 2 (57.58 vs 56.69), and Hy-DiT Pick-a-Pic ImageReward (1.444 vs 1.441), with many further cells within 0.01–0.1 points. No error bars, seeds, or significance tests are reported in Tables 1–3, and the budgets are not matched (Ctrl-Z‡ uses 7.72–8.79 NFE against SOP-4's 9.00; Ctrl-Z† uses 2.77–2.85 against SOP-1's 3.00), so the realized cost is also a random variable that the single averages hide. The §5.2 sentence claiming 'higher HPSv2 and PickScore' than SOP is contradicted by two of the four PickScore cells. Please either add variance and paired statistical tests at matched budgets or explicitly restrict the claim to specific metrics and regimes.
  2. [Algorithm 1; Tables 1–3] ImageReward (IR) is simultaneously the search objective (Algorithm 1, lines 8, 22, 27) and the principal headline metric of Tables 1 and 2, so the 'significant gains on ImageReward' highlighted in §5.2 are largely guaranteed by construction: the acceptance rule in lines 27–32 keeps r_prev non-decreasing in R. Table 3 shows the selection effect directly: exploring with PickScore raises evaluation PickScore to 21.69 (vs 21.02 when exploring with IR), exploring with HPSv2 raises evaluation HPSv2 to 27.39 (vs 27.34), and exploring with AES raises evaluation AES to 6.147 (vs 5.705). Given the paper's goal of improved 'generation quality,' the IR columns should be presented as the optimized metric, with the non-circular evidence (CompBench subsets, HPSv2, PickScore, AES) as primary, or the authors should provide a held-out evaluation regime in which the evaluation metric's reward model is never used during search.
  3. [Table 1; Supp. C.1] The paper uses 'generation quality' in the abstract to cover what the experiments actually measure as two distinct quantities: prompt alignment and aesthetic quality. The evidence for aesthetic quality is mixed: on Hunyuan-DiT DrawBench, both Ctrl-Z variants score below DDIM on AES (5.906 and 5.908 vs 5.930), while SOP-4 reaches 5.942, and Supp. C.1 concedes a 'slight reduction in AES' on Hunyuan-DiT. The claims should be decomposed into alignment-sensitive metrics and aesthetic metrics, with the trade-off quantified, so the reader can judge whether 'quality' uniformly improves.
  4. [Eq. (3); Supp. D.1] The plateau detector (Eq. 5) and the candidate acceptance rule both rely on the premise that R evaluated at the intermediate clean estimate x̂_0 (Eq. 3) is a faithful proxy for final-image quality. Supp. D.1 acknowledges that exploration can 'overfit to intermediate states' that score well under R without improving the final output. Because this premise is load-bearing for the method's correctness and is currently only indirectly probed by Table 2, I ask for a direct check: report, per prompt, the correlation between stepwise R values and final-image metrics (e.g., HPSv2 or CompBench subset scores), and quantify how often the accepted candidate improves final metrics versus only the surrogate.
minor comments (8)
  1. [Algorithm 1, line 21] The superscript convention for x̂_0^{t−1} is confusing: Eq. (3) defines it as the clean estimate computed from the current latent x_t, but line 21 computes it from the inverted-and-re-denoised state x̃_t; please clarify whether the estimate is taken from x̃_t or from the re-denoised x̃_{t−1}.
  2. [Algorithm 1] The call signature of the reward model is inconsistent: line 8 calls R([c], x̂_0^{t−1}) while line 22 calls R(c, x̂_0^{t−1}); unify these.
  3. [Abstract; §1; §4.3] There are several grammatical slips, including 'a local optima' (Abstract and §1), 'suboptimal local optimal' (§4.3), and 'alleviate above issues' (§1); these should be corrected.
  4. [Section 4.5] The rationale for the exploration window states that early steps assign 'substantial weight (above 0.02)' to the clean estimate; the 0.02 threshold appears unsubstantiated, so either justify the value or remove it.
  5. [Table 3] The table reuses the same four metric columns under two different block headers, and the bold/underline ranking conventions are not explained for the second block, which makes the ablation results harder to read than necessary.
  6. [§2.2; Supp. A.2] Restart Sampling [100] and diffusion tree search [36] are cited as the closest priors (DFS-style rollbacks) but are not included in any comparison; a short justification for their omission would help the reader calibrate the incremental contribution.
  7. [Table 1] The NFE definition is inconsistent: Table 1 says 'average number of function evaluations' while Supp. B.3 defines it per denoising step; specify the normalization in the caption.
  8. [Figures 3 and 5] The markers rely on color and are small enough that the paper itself says comparisons are 'best viewed zoomed in'; larger, shape-coded markers or numeric labels would make the trends legible in print.

Circularity Check

1 steps flagged · score 4.0 of 10

ImageReward serves as both the search objective and a headline evaluation metric, making the reported IR gains partly by construction; the non-optimized HPSv2, PickScore, and AES metrics keep the central claim partially independent.

  1. other [Algorithm 1 lines 21-28; Section 4.3 Eq. (5); Section 5.1; Tables 1-2]
    "At each iteration, N candidates are generated using distinct noise vectors ϵ, and the one with the highest reward is selected. ... A local maximum is detected when the current reward fails to improve over the most recently accepted score by at least δ: R(c, x̂^{t−1}_0) < r_prev + δ. ... By default, δ = 0 enforces non-decreasing reward over time. ... Evaluations are conducted using four metrics: HPSv2, PickScore, ImageReward (IR), and Aesthetic Score (AES)."

    The same reward model R (ImageReward) is used to trigger exploration (Eq. 5), to score candidate continuations (Algorithm 1 line 22: r_cand ← R(c, x̂^{t−1}_0)), and to accept updates only when best_score ≥ r_prev + δ (line 27). With δ = 0, the accepted trajectory is selected to have non-decreasing R, so the reported ImageReward gains in Tables 1-2 are in part a restatement of the selection rule rather than an independent measurement. The paper itself warns in Supp. D.1 that exploration may overfit to intermediate states that align well with the reward model. The reduction is only partial because final IR is measured on final images rather than the intermediate x̂_0 used for selection, and because HPSv2, PickScore, and AES are not optimized and mostly improve as well.

full rationale

The paper is an empirical inference-time search method rather than a first-principles derivation, and its main evidence is against external baselines on standard benchmarks. No load-bearing self-citation chain, imported uniqueness theorem, or ansatz-via-citation pattern is present. The one substantive circularity concern is the dual role of ImageReward: it is both the objective optimized by Algorithm 1 (via Eq. 5 and candidate selection) and a principal evaluation metric in Tables 1-2. With δ = 0, the algorithm by construction maintains a non-decreasing reward along the accepted trajectory, so the IR improvements are partly guaranteed. This does not collapse the paper, however, because the method also reports and mostly improves non-optimized metrics (HPSv2, PickScore, AES) on external benchmarks. The separate issue that the 'consistently improves' claim is weakened by a few table cells favoring SOP-4 and by the absence of error bars is a statistical-evidence concern, not a circularity, and is therefore not counted beyond the partial metric-overlap step.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method's effectiveness rests on a handful of tuned hyperparameters (lambda, N, d_max, delta, reward model) and on the behavioral assumption that reward plateaus signal recoverable local optima. The reward model itself is a measurable input and the main source of potential circularity.

free parameters (5)
  • lambda (exploration window) = 40 (default)
    Restricts exploration to the first 40 of 50 steps; Supp. C.3 (Table 6) shows larger lambda improves IR but raises NFEs; 40 chosen as a balance.
  • N (candidates per depth) = 4
    Number of random inversion candidates per search step; Fig. 3 shows diminishing returns for larger N.
  • d_max (max inversion depth) = 3
    Caps escalation to control compute; Fig. 3 and C.2 show deeper search improves quality but costs NFEs.
  • delta (acceptance threshold) = 0
    Non-decreasing reward acceptance; Supp. C.2 (Fig. 5) shows negative/positive delta change compute and quality trade-offs.
  • reward model R = ImageReward
    Surrogate quality scorer; Table 3 shows results vary significantly with choice of R, affecting both metrics and realized NFEs.
assumptions (5)
  • domain assumption DDIM inversion (Eq. 4) with added Gaussian noise preserves enough semantic structure to enable meaningful exploration.
    Algorithm 1, line 17; if inversion destroys accumulated content, rollback cannot recover better layouts.
  • domain assumption The reward model R provides a valid surrogate for final output quality at intermediate steps.
    Section 4.1 posits Q approximate R; Eq. 5 uses R for plateau detection and candidate acceptance; Supp. D.1 concedes overfitting risk.
  • domain assumption A plateau in the reward trajectory (R(c, x_hat_0) < r_prev + delta) marks a local optimum worth escaping.
    Section 4.3 states this; the whole exploration trigger depends on this behavioral premise.
  • domain assumption Early denoising steps dominate global structure, so exploration should be limited to the first lambda steps.
    Section 4.5 and Supp. C.3 justify restricting to high-noise regime; later perturbations have limited effect.
  • standard math Standard DDIM formulation and Tweedie-style clean estimate (Eq. 3) hold for the tested models.
    Section 3.2; uses standard alpha_t schedule and noise prediction from the pretrained model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ctrl-Z Sampling: Scaling Diffusion Sampling with Controlled Random Zigzag Explorations." pith.science (2026). https://pith.science/paper/I3PMCKTZ

@misc{pith2026250620294,
  author       = {Pith},
  title        = {Pith review of: Ctrl-Z Sampling: Scaling Diffusion Sampling with Controlled Random Zigzag Explorations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I3PMCKTZ}},
  note         = {Machine review of arXiv:2506.20294}
}
read the original abstract

Diffusion models generate conditional samples by progressively denoising Gaussian noise, yet the denoising trajectory can stall at visually plausible but low-quality outcomes with conditional misalignment or structural artifacts. We interpret this behavior as local optima in a surrogate quality landscape: Once early denoising commits to a suboptimal global structure, later steps mainly sharpen details and seldom correct the underlying mistake. While existing inference-time approaches explore alternative diffusion states via re-noising with fixed strength or direction, they exhibit limited capacity to escape steep quality plateaus. We propose Controlled Random Zigzag Sampling (Ctrl-Z Sampling),a scalable sampling strategy that detects plateaus in quality landscape via a surrogate score, and allocates exploration only when a plateau is detected. Upon detection, Ctrl-Z Sampling rolls back to noisier states, samples a set of alternative continuations, and updates the trajectory when a candidate improves the score, otherwise escalating the exploration depth to escape the current plateau. The proposed method is model-agnostic and broadly compatible with existing diffusion frameworks. Experiments show that Ctrl-Z Sampling consistently improves generation quality over other inference-time scaling samplers across different NFE budgets, offering a scalable compute-quality trade-off. Code available at: https://github.com/ShunqiM/Ctrl-Z-Sampling.

Figures

Figures reproduced from arXiv: 2506.20294 by the authors.

Figure 1
Figure 1. We introduce Ctrl-Z Sampling, a diffusion sampling strategy that scales with compute to improve generation quality. Under comparable budgets, it outperforms inference-scaling baselines such as Search-over-Path (SOP). For the prompt “Library on a flying whale’s back”, SOP produces misaligned compositions (red), while Ctrl-Z generate prompt-consistent (green) output as compute increase from 3× to 9× NFEs. image genera… view at source ↗
Figure 2
Figure 2. Illustration of different sampling strategies during the diffusion process. The diffusion trajectory is depicted as ascending a rugged, conceptual quality landscape for sample quality and condition alignment, measured by a surrogate score. Each strat￾egy begins with a standard denoising update (red), followed by either a single or multiple inversion explorations. The selected/executed inversions are marked in blue, … view at source ↗
Figure 3
Figure 3. Effect of exploration depth (dmax) and candidate width (N). Each marker aggregates results over the same prompts and seeds for a fixed (dmax, N). Lighter colors indicate higher realized average NFEs. Ctrl-Z Sampling benefits from both deeper and wider exploration, with increased depth sometimes yielding stronger gains than additional candidates under similar NFEs. Comparisons are best viewed zoomed in. As exploratio… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of different sampling methods with SD-2.1-base. Gen￾erated images that align with the input condition (shown on the left) are highlighted with green bounding boxes, while erroneous or suboptimal images are marked in red. low-frequency components.…
Figure 5
Figure 5. Figure 5: Evaluation metrics across different accept threshold δ = −1, 0, 1 for different max search depth dmax ∈ {1, 2, 5,∞}. requiring the reward to exceed that of the previous step by at least δ (which may be negative to allow slight decreases). We find that increasing either…
Figure 6
Figure 6. Figure 6: Left: Count of explorations initiated at each diffusion step. Right: Count of final exploration depths when adaptive exploration terminates; the yellow bar indicates cases where termination occurs due to reaching the maximum depth limit. Results and Discussion. The qua…
Figure 7
Figure 7. Figure 7: Qualitative comparison of different sampling methods. Generated images that align with the input condition (shown on the left) are highlighted with green bounding boxes, while erroneous or suboptimal images are marked in red. Inconsistently generated concept are also b…
Figure 8
Figure 8. Figure 8: Sampling trajectories over the first 10 steps, comparing DDIM (right) and Ctrl-Z Sampling (left). Prompt: “a yellow dog and a blue apple”. Red boxes mark steps where exploration is triggered, and green boxes mark regular steps. DDIM greedily steps towards higher probab…
Figure 9
Figure 9. Figure 9: Qualitative comparison of the decoded X t−1 0 across the first 30 over 50 gen￾eration steps. Steps with exploration initiated are highlighted in red, while others are shown in green [PITH_FULL_IMAGE:figures/full_fig_p040_9.png]
Figure 10
Figure 10. Figure 10: Qualitative comparison of the decoded Xt−1 0 across the first 30 over 50 generation steps. Steps with exploration initiated are highlighted in red, while others are shown in green [PITH_FULL_IMAGE:figures/full_fig_p041_10.png]
Figure 11
Figure 11. Figure 11: Qualitative comparison of the decoded X t−1 0 across the first 30 over 50 gen￾eration steps. Steps with exploration initiated are highlighted in red, while others are shown in green [PITH_FULL_IMAGE:figures/full_fig_p042_11.png]
Figure 12
Figure 12. Figure 12: Qualitative comparison of the decoded Xt−1 0 across the first 30 over 50 generation steps. Steps with exploration initiated are highlighted in red, while others are shown in green [PITH_FULL_IMAGE:figures/full_fig_p043_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

112 extracted references · 54 canonical work pages

  1. [1]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2024) 2, 4, 7, 9, 26

    Bai, L., Shao, S., Zhou, Z., Qi, Z., Xu, Z., Xiong, H., Xie, Z.: Zigzag diffusion sampling: Diffusion models can self-improve via self-reflection. In: Proceedings of the International Conference on Learning Representations (ICLR) (2024) 2, 4, 7, 9, 26

    2024

  2. [2]

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

    Bansal, A., Chu, H.M., Schwarzschild, A., Sengupta, S., Goldblum, M., Geiping, J., Goldstein, T.: Universal guidance for diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR). pp. 843–852 (2023) 1

    2023

  3. [3]

    In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)

    Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Podstawski, M., Gianinazzi, L., Gajda, J., Lehmann, T., Niewiadomski, H., Nyczyk, P., Hoefler, T.: Graph of thoughts: Solving elaborate problems with large language models. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 38, pp. 17682– 17690 (2024) 23

    2024

  4. [4]

    arXiv preprint arXiv:2311.15127 (2023) 1

    Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023) 1

    arXiv 2023

  5. [5]

    In: International Conference on Machine Learning (ICML) (2025) 23

    Bortoli, V.D., Galashov, A., Gretton, A., Doucet, A.: Accelerated diffusion mod- els via speculative sampling. In: International Conference on Machine Learning (ICML) (2025) 23

    2025

  6. [6]

    arXiv preprint arXiv:2407.21787 (2024) 23

    Brown, B., Juravsky, J., Ehrlich, R., Clark, R., Le, Q.V., Ré, C., Mirhoseini, A.: Large language monkeys: Scaling inference compute with repeated sampling. arXiv preprint arXiv:2407.21787 (2024) 23

    arXiv 2024

  7. [7]

    Advances in Neural Information Processing Systems (NeurIPS)33, 1877–1901 (2020) 23

    Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Nee- lakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few- shot learners. Advances in Neural Information Processing Systems (NeurIPS)33, 1877–1901 (2020) 23

    2020

  8. [8]

    In: International Conference on Machine Learning (ICML)

    Chamberlain, B., Rowbottom, J., Gorinova, M.I., Bronstein, M., Webb, S., Rossi, E.: Grand: Graph neural diffusion. In: International Conference on Machine Learning (ICML). pp. 1407–1418 (2021) 1

    2021

Show all 112 references
  1. [9]

    ACM SIG- GRAPH42(4), 1–10 (2023) 4

    Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., Cohen-Or, D.: Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM SIG- GRAPH42(4), 1–10 (2023) 4

    2023

  2. [10]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2024) 1

    Chen, C., Ding, H., Sisman, B., Xu, Y., Xie, O., Yao, B.Z., Tran, S.D., Zeng, B.: Diffusion models for multi-task generative modeling. In: Proceedings of the International Conference on Learning Representations (ICLR) (2024) 1

    2024

  3. [11]

    Advances in Neural Information Processing Systems (NeurIPS)37, 133661–133709 (2024) 23

    Chen, H., Ren, Y., Ying, L., Rotskoff, G.: Accelerating diffusion models with parallel sampling: Inference at sub-linear time complexity. Advances in Neural Information Processing Systems (NeurIPS)37, 133661–133709 (2024) 23

    2024

  4. [12]

    In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2024) 4

    Choi, Y., Park, J., Kim, H., Lee, J., Park, S.: Fair sampling in diffusion mod- els through switching mechanism. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2024) 4

    2024

  5. [13]

    arXiv preprint arXiv:2406.08070 (2024) 4, 31, 33

    Chung,H.,Kim,J.,Park,G.Y.,Nam,H.,Ye,J.C.:CFG++:Manifold-constrained classifier free guidance for diffusion models. arXiv preprint arXiv:2406.08070 (2024) 4, 31, 33

    2024 arXiv

  6. [14]

    arXiv preprint arXiv:2110.14168 (2021) 24 16 S

    Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., et al.: Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168 (2021) 24 16 S. Mao et al

    2021 arXiv

  7. [15]

    In: Proceedings of the Interna- tional Conference on Learning Representations (ICLR) (2023) 24

    Creswell, A., Shanahan, M., Higgins, I.: Selection-inference: Exploiting large lan- guage models for interpretable logical reasoning. In: Proceedings of the Interna- tional Conference on Learning Representations (ICLR) (2023) 24

    2023

  8. [16]

    https://r1.deepseek.com/(2025) 23

    DeepSeek-AI: Deepseek-r1: Raising the bar of open-source reasoning models. https://r1.deepseek.com/(2025) 23

    2025

  9. [17]

    In: Ad- vances in Neural Information Processing Systems (NeurIPS)

    Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: Ad- vances in Neural Information Processing Systems (NeurIPS). vol. 34, pp. 8780– 8794 (2021) 4

    2021

  10. [18]

    Journal of the American Statis- tical Association106, 1602 – 1614 (2011),https://api.semanticscholar.org/ CorpusID:232841545

    Efron, B.: Tweedie’s formula and selection bias. Journal of the American Statis- tical Association106, 1602 – 1614 (2011),https://api.semanticscholar.org/ CorpusID:232841545

    2011

  11. [19]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

    Everaert, M.N., Fitsios, A., Bocchio, M., Arpa, S., Süsstrunk, S., Achanta, R.: Ex- ploiting the signal-leak bias in diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 4025–4034 (2024) 4

    2024

  12. [20]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Fan, Y., Watkins, O., Du, Y., Liu, H., Ryu, M., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Lee, K., Lee, K.: Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 36, pp. 79858–79885 (2023) 4

    2023

  13. [21]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Fang, G., Ma, X., Wang, X.: Structural pruning for diffusion models. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 36, pp. 16716–16728 (2023) 4

    2023

  14. [22]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2023) 4, 9, 25

    Feng, W., He, X., Fu, T.J., Jampani, V., Akula, A.R., Narayana, P., Basu, S., Wang, X.E., Wang, W.Y.: Training-free structured diffusion guidance for compo- sitional text-to-image synthesis. In: Proceedings of the International Conference on Learning Representations (ICLR) (20...

    2023

  15. [23]

    In: Conference on Language Modeling (COLM)) (2024) 24

    Gandhi, K., Lee, D.H.J., Grand, G., Liu, M., Cheng, W., Sharma, A., Goodman, N.: Stream of search (sos): Learning to search in language. In: Conference on Language Modeling (COLM)) (2024) 24

    2024

  16. [24]

    In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    Gangi Reddy, R., Doo, J., Xu, Y., Sultan, M.A., Swain, D., Sil, A., Ji, H.: FIRST: Faster improved listwise reranking with single token decoding. In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 8642–8652 (2024) 24

    2024

  17. [25]

    In: Proceedings of the Inter- national Conference on Learning Representations (ICLR) (2023) 4

    Ghosh, A., Lyu, H., Zhang, X., Wang, R.: Implicit regularization in heavy-ball momentum accelerated stochastic gradient descent. In: Proceedings of the Inter- national Conference on Learning Representations (ICLR) (2023) 4

    2023

  18. [26]

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

    Guo, X., Liu, J., Cui, M., Li, J., Yang, H., Huang, D.: Initno: Boosting text- to-image diffusion models via initial noise optimization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR). pp. 9380–9389 (2024) 4

    2024

  19. [27]

    CrossLabs Research Blogs (January

    Guttenberg, N.: Diffusion with offset noise. CrossLabs Research Blogs (January

  20. [28]

    arXiv preprint arXiv:2505.17618 (2025) 2, 4, 24

    He, H., Liang, J., Wang, X., Wan, P., Zhang, D., Gai, K., Pan, L.: Scaling image and video generation via test-time evolutionary search. arXiv preprint arXiv:2505.17618 (2025) 2, 4, 24

    2025 arXiv

  21. [29]

    arXiv preprint arXiv:2104.08718 (2021) 12, 25

    Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference- free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021) 12, 25

    2021 arXiv

  22. [30]

    In: Ad- vances in Neural Information Processing Systems (NeurIPS)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Ad- vances in Neural Information Processing Systems (NeurIPS). vol. 33, pp. 6840– 6851 (2020) 1, 3, 4, 24 Ctrl-Z Sampling 17

    2020

  23. [31]

    arXiv preprint arXiv:2207.12598 (2022) 2, 4

    Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) 2, 4

    2022 arXiv

  24. [32]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video dif- fusion models. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 35, pp. 8633–8646 (2022) 1

    2022

  25. [33]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., de Las Casas, D., Hendricks, L.A., Welbl, J., Clark, A., Hennigan, T., Noland, E., Millican, K., van den Driessche, G., Damoc, B., Guy, A., Osindero, S., Si- monyan, K., Elsen, E., Vinyals, O., R...

    2022

  26. [34]

    In: Advances in Neural Information Processing Sys- tems (NeurIPS)

    Hong, S.: Smoothed energy guidance: Guiding diffusion models with reduced en- ergy curvature of attention. In: Advances in Neural Information Processing Sys- tems (NeurIPS). pp. 66743–66772 (2024) 4

    2024

  27. [35]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Huang, K., Sun, K., Xie, E., Li, Z., Liu, X.: T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 36, pp. 78723–78747 (2023) 9, 24

    2023

  28. [36]

    arXiv preprint arXiv:2506.20701 (2025) 4

    Jain, V., Sareen, K., Pedramfar, M., Ravanbakhsh, S.: Diffusion tree sam- pling: Scalable inference-time alignment of diffusion models. arXiv preprint arXiv:2506.20701 (2025) 4

    2025 arXiv

  29. [37]

    In: International Conference on Machine Learning (ICML) (2022) 1

    Janner, M., Du, Y., Tenenbaum, J.B., Levine, S.: Planning with diffusion for flexible behavior synthesis. In: International Conference on Machine Learning (ICML) (2022) 1

    2022

  30. [38]

    arXiv:2411.11694 (2024) 24

    Jiang,J.,Chen,Z.,Min,Y.,Chen,J.,etal.:Enhancingllmreasoningwithreward- guided tree search. arXiv:2411.11694 (2024) 24

    2024 arXiv

  31. [39]

    In: Chiruzzo, L., Ritter, A., Wang, L

    Jinnai, Y., Morimura, T., Ariu, K., Abe, K.: Regularized best-of-n sampling with minimum Bayes risk objective for language model alignment. In: Chiruzzo, L., Ritter, A., Wang, L. (eds.) Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Associatio...

    2025

  32. [40]

    arXiv preprint arXiv:2001.08361 (2020) 23

    Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., Amodei, D.: Scaling laws for neural language mod- els. arXiv preprint arXiv:2001.08361 (2020) 23

    2020 arXiv

  33. [41]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2022) 11, 23

    Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. In: Advances in Neural Information Processing Systems (NeurIPS) (2022) 11, 23

    2022

  34. [42]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2024) 23

    Katzir, O., Patashnik, O., Cohen-Or, D., Lischinski, D.: Noise-free score distilla- tion. In: Proceedings of the International Conference on Learning Representations (ICLR) (2024) 23

    2024

  35. [43]

    In: Proceedings of the International Conference on Learn- ing Representations (ICLR) (2025) 4, 26

    Kim, S., Kim, M., Park, D.: Test-time alignment of diffusion models without re- ward over-optimization. In: Proceedings of the International Conference on Learn- ing Representations (ICLR) (2025) 4, 26

    2025

  36. [44]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    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. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 36, pp. 36652–36663 (2023) 9, 24, 25

    2023

  37. [45]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2021) 1 18 S

    Kong, Z., Ping, W., Huang, J., Zhao, K., Catanzaro, B.: Diffwave: A versatile dif- fusion model for audio synthesis. In: Proceedings of the International Conference on Learning Representations (ICLR) (2021) 1 18 S. Mao et al

    2021

  38. [46]

    Advances in Neural Information Processing Systems (NeurIPS)37, 85203–85240 (2024) 4

    Li, S., Hu, T., van de Weijer, J., Khan, F.S., Liu, T., Li, L., Yang, S., Wang, Y., Cheng, M.M., Yang, J.: Faster diffusion: Rethinking the role of the encoder for diffusion model inference. Advances in Neural Information Processing Systems (NeurIPS)37, 85203–85240 (2024) 4

    2024

  39. [47]

    arXiv preprint arXiv:2502.17437 (2025) 3

    Li, T., Sun, Q., Fan, L., He, K.: Fractal generative models. arXiv preprint arXiv:2502.17437 (2025) 3

    2025 arXiv

  40. [48]

    In: Advances in Neural Information Pro- cessing Systems (NeurIPS)

    Li, X., Thickstun, J., Gulrajani, I., Liang, P.S., Hashimoto, T.B.: Diffusion-lm improves controllable text generation. In: Advances in Neural Information Pro- cessing Systems (NeurIPS). vol. 35, pp. 4328–4343 (2022) 1

    2022

  41. [49]

    arXiv preprint arXiv:2503.02039 (2025) 4

    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. arXiv preprint arXiv:2503.02039 (2025) 4

    2025 arXiv

  42. [50]

    In: British Machine Vision Conference (BMVC) (2023) 4

    Li, Y., Keuper, M., Zhang, D., Khoreva, A.: Divide & bind your attention for improved generative semantic nursing. In: British Machine Vision Conference (BMVC) (2023) 4

    2023

  43. [51]

    arXiv preprint arXiv:2405.08748 (2024) 9, 25

    Li,Z.,Zhang,J.,Lin,Q.,Xiong,J.,Long,Y.,Deng,X.,Zhang,Y.,Liu,X.,Huang, M., Xiao, Z., Chen, D., He, J., Li, J., Li, W., Zhang, C., Quan, R., Lu, J., Huang, J., Yuan, X., Zheng, X., Li, Y., Zhang, J., Zhang, C., Chen, M., Liu, J., Fang, Z., Wang, W., Xue, J., Tao, Y., Zhu, J., Li...

    2024 arXiv

  44. [52]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2024) 24

    Lightman, H., Kosaraju, V., Burda, Y., Edwards, H., Baker, B., Lee, T., Leike, J., Schulman, J., Sutskever, I., Cobbe, K.: Let’s verify step by step. In: Proceedings of the International Conference on Learning Representations (ICLR) (2024) 24

    2024

  45. [53]

    In: Proceedings of the IEEE/CVF winter conference on applica- tions of computer vision (WACV)

    Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. In: Proceedings of the IEEE/CVF winter conference on applica- tions of computer vision (WACV). pp. 5404–5411 (2024) 4

    2024

  46. [54]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2023) 24

    Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: Proceedings of the International Conference on Learning Representations (ICLR) (2023) 24

    2023

  47. [55]

    arXiv preprint arXiv:2502.06703 (2025) 24

    Liu, R., Gao, J., Zhao, J., Zhang, K., Li, X., Qi, B., Ouyang, W., Zhou, B.: Can 1b llm surpass 405b llm? rethinking compute-optimal test-time scaling. arXiv preprint arXiv:2502.06703 (2025) 24

    2025 arXiv

  48. [56]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2023) 24

    Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. In: Proceedings of the International Conference on Learning Representations (ICLR) (2023) 24

    2023

  49. [57]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 35, pp. 5775–5787 (2022) 4, 23, 31

    2022

  50. [58]

    Machine Intelligence Research pp

    Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. Machine Intelligence Research pp. 1–22 (2025) 23

    2025

  51. [59]

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

    Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Re- paint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR). pp. 11461–11471 (2022) 2, 4, 6, 9, 26 ...

    2022

  52. [60]

    In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR)

    Ma, N., Tong, S., Jia, H., Hu, H., Su, Y.C., Zhang, M., Yang, X., Li, Y., Jaakkola, T., Jia, X., Xie, S.: Scaling inference time compute for diffusion models. In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR). pp. 2523–2534 (June 20...

    2025

  53. [61]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Madaan, A., Tandon, N., Gupta, P., Hallinan, S., Gao, L., Wiegreffe, S., Alon, U., Dziri, N., Prabhumoye, S., Yang, Y., et al.: Self-refine: Iterative refinement with self-feedback. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 36, pp. 46534–46594 (2023) 23

    2023

  54. [62]

    In: Proceedings of the 31st ACM International Conference on Multimedia (MM)

    Mao, J., Wang, X., Aizawa, K.: Guided image synthesis via initial image editing in diffusion model. In: Proceedings of the 31st ACM International Conference on Multimedia (MM). pp. 5321–5329 (2023) 4

    2023

  55. [63]

    arXiv preprint arXiv:2501.19393 (2025) 23

    Muennighoff, N., Yang, Z., Shi, W., Li, X.L., Fei-Fei, L., Hajishirzi, H., Zettle- moyer, L., Liang, P., Candès, E., Hashimoto, T.: s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393 (2025) 23

    2025 arXiv

  56. [64]

    OpenAI: Learning to reason with llms.https://openai.com/index/learning- to-reason-with-llms/(2024) 23

    2024

  57. [65]

    arXiv preprint arXiv:2501.19252 (2025) 4

    Oshima, Y., Suzuki, M., Matsuo, Y., Furuta, H.: Inference-time text-to-video alignment with diffusion latent beam search. arXiv preprint arXiv:2501.19252 (2025) 4

    2025

  58. [66]

    In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR)

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR). pp. 4195–4205 (2023) 3

    2023

  59. [67]

    arXiv preprint arXiv:2307.01952 (2023) 1, 3

    Podell,D., English, Z., Lacey, K.,Blattmann, A., Dockhorn, T.,Müller, J., Penna, J.,Rombach,R.:Sdxl:Improvinglatentdiffusionmodelsforhigh-resolutionimage synthesis. arXiv preprint arXiv:2307.01952 (2023) 1, 3

    2023 arXiv

  60. [68]

    In: Proceedings of the International Conference on Learning Repre- sentations (ICLR) (2023) 1

    Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: Proceedings of the International Conference on Learning Repre- sentations (ICLR) (2023) 1

    2023

  61. [69]

    In: International Conference on Machine Learning (ICML)

    Radford,A.,Kim,J.W.,Hallacy,C.,Ramesh,A.,Goh,G.,Agarwal,S.,Sastry,G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (ICML). pp. 8748–8763 (2021) 25

    2021

  62. [70]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2025) 2, 4

    Ramesh, V., Mardani, M.: Test-time scaling of diffusion models via noise trajec- tory search. In: Advances in Neural Information Processing Systems (NeurIPS) (2025) 2, 4

    2025

  63. [71]

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

    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 (CVPR). pp. 10684–10695 (2022) 1, 3

    2022

  64. [72]

    In: International Conference on Machine Learning (ICML) (2024) 4, 26, 31

    Sabour, A., Fidler, S., Kreis, K.: Align your steps: optimizing sampling schedules in diffusion models. In: International Conference on Machine Learning (ICML) (2024) 4, 26, 31

    2024

  65. [73]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: Advances in Neural Information Processing Systems...

    2022

  66. [74]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2022) 23 20 S

    Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion mod- els. In: Proceedings of the International Conference on Learning Representations (ICLR) (2022) 23 20 S. Mao et al

    2022

  67. [75]

    In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2024) 4

    Samuel, D., Ben-Ari, R., Raviv, S., Darshan, N., Chechik, G.: Generating images of rare concepts using pre-trained diffusion models. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) (2024) 4

    2024

  68. [76]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: Advances in Neural Information Processing Systems (...

    2022

  69. [77]

    arXiv preprint arXiv:1707.06347 (2017) 23

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017) 23

    2017 arXiv

  70. [78]

    In: International Conference on Machine Learning (ICML) (2025) 4

    Seo, H., Jeong, W., sun Seo, J., Chun, S.Y.: Skrr: Skip and re-use text encoder layers for memory efficient text-to-image generation. In: International Conference on Machine Learning (ICML) (2025) 4

    2025

  71. [79]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2025) 24

    Setlur, A., Nagpal, C., Fisch, A., Geng, X., Eisenstein, J., Agarwal, R., Agar- wal, A., Berant, J., Kumar, A.: Rewarding progress: Scaling automated process verifiers for LLM reasoning. In: Proceedings of the International Conference on Learning Representations (ICLR) (2025) 24

    2025

  72. [80]

    Advances in Neural Information Processing Systems (NeurIPS) 36, 4263–4276 (2023) 23

    Shih, A., Belkhale, S., Ermon, S., Sadigh, D., Anari, N.: Parallel sampling of diffusion models. Advances in Neural Information Processing Systems (NeurIPS) 36, 4263–4276 (2023) 23

    2023

  73. [81]

    In: International Conference on Machine Learning (ICML) (2025) 4, 24

    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 mod- els. In: International Conference on Machine Learning (ICML) (2025) 4, 24

    2025

  74. [82]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2025) 23

    Snell, C.V., Lee, J., Xu, K., Kumar, A.: Scaling LLM test-time compute optimally can be more effective than scaling parameters for reasoning. In: Proceedings of the International Conference on Learning Representations (ICLR) (2025) 23

    2025

  75. [83]

    In: International Confer- ence on Machine Learning (ICML)

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsuper- vised learning using nonequilibrium thermodynamics. In: International Confer- ence on Machine Learning (ICML). pp. 2256–2265. pmlr (2015) 1, 3, 24

    2015

  76. [84]

    In: Proceed- ings of the International Conference on Learning Representations (ICLR) (2021) 1, 3, 4, 5, 26

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: Proceed- ings of the International Conference on Learning Representations (ICLR) (2021) 1, 3, 4, 5, 26

    2021

  77. [85]

    In: Interna- tional Conference on Machine Learning (ICML)

    Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. In: Interna- tional Conference on Machine Learning (ICML). pp. 32211–32252 (2023) 23

    2023

  78. [86]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 32 (2019) 24

    2019

  79. [87]

    Advances in Neural Information Processing Systems (NeurIPS)37, 32630–32652 (2024) 24

    Sun, H., Haider, M., Zhang, R., Yang, H., Qiu, J., Yin, M., Wang, M., Bartlett, P., Zanette, A.: Fast best-of-n decoding via speculative rejection. Advances in Neural Information Processing Systems (NeurIPS)37, 32630–32652 (2024) 24

    2024

  80. [88]

    arXiv preprint arXiv:2405.18881 (2024) 4

    Tang, Z., Peng, J., Tang, J., Hong, M., Wang, F., Chang, T.H.: Inference-time alignment of diffusion models with direct noise optimization. arXiv preprint arXiv:2405.18881 (2024) 4

    2024 arXiv

  81. [89]

    Transactions on Machine Learning Research (TMLR)2024(2024) 24

    Tong, A., Fatras, K., Malkin, N., Huguet, G., Zhang, Y., Rector-Brooks, J., Wolf, G., Bengio, Y.: Improving and generalizing flow-based generative models with minibatch optimal transport. Transactions on Machine Learning Research (TMLR)2024(2024) 24

    2024

  82. [90]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2025) 23 Ctrl-Z Sampling 21

    Tong, V., Hoang, D.T., Liu, A., den Broeck, G.V., Niepert, M.: Learning to dis- cretize denoising diffusion ODEs. In: Proceedings of the International Conference on Learning Representations (ICLR) (2025) 23 Ctrl-Z Sampling 21

    2025

  83. [91]

    arXiv preprint arXiv:2211.14275 (2022) 24

    Uesato, J., Kushman, N., Kumar, R., Song, F., Siegel, N., Wang, L., Creswell, A., Irving, G., Higgins, I.: Solving math word problems with process-and outcome- based feedback. arXiv preprint arXiv:2211.14275 (2022) 24

    2022 arXiv

  84. [92]

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

    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 (CVPR). pp....

    2024

  85. [93]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2023) 23

    Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language mod- els. In: Proceedings of the International Conference on Learning Representations (ICLR) (2023) 23

    2023

  86. [94]

    In: Proceedings of the Interna- tional Conference on Learning Representations (ICLR) (2022) 4

    Watson, D., Chan, W., Ho, J., Norouzi, M.: Learning fast samplers for diffusion models by differentiating through sample quality. In: Proceedings of the Interna- tional Conference on Learning Representations (ICLR) (2022) 4

    2022

  87. [95]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., Zhou, D.: Chain-of-thought prompting elicits reasoning in large language models. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 35, pp. 24824–24837 (2022) 23

    2022

  88. [96]

    arXiv preprint arXiv:2306.09341 (2023) 9, 25

    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. arXiv preprint arXiv:2306.09341 (2023) 9, 25

    2023 arXiv

  89. [97]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    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. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 36, pp. 15903–15935 (2023) 9, 25

    2023

  90. [98]

    arXiv preprint arXiv:2405.14828 (2024) 2, 4

    Xu, K., Zhang, L., Shi, J.: Good seed makes a good crop: Discovering secret seeds in text-to-image diffusion models. arXiv preprint arXiv:2405.14828 (2024) 2, 4

    2024 arXiv

  91. [99]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2022) 1

    Xu, M., Yu, L., Song, Y., Shi, C., Ermon, S., Tang, J.: Geodiff: A geometric diffusion model for molecular conformation generation. In: Proceedings of the International Conference on Learning Representations (ICLR) (2022) 1

    2022

  92. [100]

    Xu, Y., Deng, M., Cheng, X., Tian, Y., Liu, Z., Jaakkola, T.S.: Restart sampling forimprovinggenerativeprocesses.In:AdvancesinNeuralInformationProcessing Systems (NeurIPS) (2023) 4

    2023

  93. [101]

    arXiv preprint arXiv:2502.18873 (2025) 24

    Yang, S., Li, Y., Lam, W., Cheng, Y.: Multi-llm collaborative search for complex problem solving. arXiv preprint arXiv:2502.18873 (2025) 24

    2025 arXiv

  94. [102]

    In: Ad- vances in Neural Information Processing Systems (NeurIPS)

    Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T.L., Cao, Y., Narasimhan, K.: Tree of thoughts: Deliberate problem solving with large language models. In: Ad- vances in Neural Information Processing Systems (NeurIPS). vol. 36, pp. 11809– 11822 (2023) 23

    2023

  95. [103]

    In: Proceedings of the In- ternational Conference on Learning Representations (ICLR) (2023) 23

    Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., Cao, Y.: React: Synergizing reasoning and acting in language models. In: Proceedings of the In- ternational Conference on Learning Representations (ICLR) (2023) 23

    2023

  96. [104]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Zelikman, E., Wu, Y., Mu, J., Goodman, N.D.: Star: Bootstrapping reason- ing with reasoning. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 35, pp. 15476–15488 (2022) 23

    2022

  97. [105]

    Advances in Neural Information Processing Systems (NeurIPS)37, 64735–64772 (2024) 24

    Zhang, D., Zhoubian, S., Hu, Z., Yue, Y., Dong, Y., Tang, J.: Rest-mcts*: Llm self- training via process reward guided tree search. Advances in Neural Information Processing Systems (NeurIPS)37, 64735–64772 (2024) 24

    2024

  98. [106]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2025) 24 22 S

    Zhang, L., Hosseini, A., Bansal, H., Kazemi, M., Kumar, A., Agarwal, R.: Gener- ative verifiers: Reward modeling as next-token prediction. In: Proceedings of the International Conference on Learning Representations (ICLR) (2025) 24 22 S. Mao et al

    2025

  99. [107]

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

    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 (ICCV). pp. 3836–3847 (2023) 1

    2023

  100. [108]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2023) 23

    Zhang, Q., Chen, Y.: Fast sampling of diffusion models with exponential integra- tor. In: Proceedings of the International Conference on Learning Representations (ICLR) (2023) 23

    2023

  101. [109]

    arXiv preprint arXiv:2503.24235 (2025) 23

    Zhang, Q., Lyu, F., Sun, Z., Wang, L., Zhang, W., Guo, Z., Wang, Y., King, I., Liu, X., Ma, C.: What, how, where, and how well? a survey on test-time scaling in large language models. arXiv preprint arXiv:2503.24235 (2025) 23

    2025 arXiv

  102. [110]

    arXiv preprint arXiv:2505.23614 (2025) 2, 4, 24

    Zhang, X., Lin, H., Ye, H., Zou, J., Ma, J., Liang, Y., Du, Y.: Inference-time scal- ing of diffusion models through classical search. arXiv preprint arXiv:2505.23614 (2025) 2, 4, 24

    2025

  103. [111]

    In: Advances in Neural Informa- tion Processing Systems (NeurIPS)

    Zhao, W., Bai, L., Rao, Y., Zhou, J., Lu, J.: Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. In: Advances in Neural Informa- tion Processing Systems (NeurIPS). vol. 36, pp. 49842–49869 (2023) 4, 26

    2023

  104. [112]

    Revise the grammar and academic wording of this paragraph, and list aspects to be improved, including suggestions on how to improve them

    Zhou, Z., Shao, S., Bai, L., Xu, Z., Han, B., Xie, Z.: Golden noise for diffusion models: A learning framework. arXiv preprint arXiv:2411.09502 (2024) 4 Ctrl-Z Sampling 23 Ctrl-Z Sampling: Scaling Diffusion Sampling with Controlled Random Zigzag Explorations Appendix A Extende...

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.