Pith. sign in

REVIEW 2 major objections 5 minor 45 references

Under real wall-clock budgets, cheap draft candidates beat guided intermediate search for scaling text-to-image diffusion models.

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 · grok-4.5

2026-07-11 18:58 UTC pith:EHFC4BNP

load-bearing objection Wall-clock flips the NFE rankings: cheap multi-knob drafts + multi-stage selection beat guided search, and the gains hold across models and transfer to reflection and RL. the 2 major comments →

arxiv 2607.04461 v1 pith:EHFC4BNP submitted 2026-07-05 cs.CV

Flash-BoN: Instant Drafts for Inference-Time Scaling in Diffusion Models

Ruchit Rawal , Reza Shirkavand , Sayak Paul , Yuxin Wen , Heng Huang , Yizheng Chen , Tom Goldstein , Gowthami Somepalli This is my paper
classification cs.CV
keywords inference-time scalingtext-to-image generationdiffusion modelsbest-of-Nwall-clock evaluationdraft generationverifier overheadflow matching
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.

This paper argues that the usual way of comparing inference-time scaling methods for text-to-image models—counting only denoising steps (NFEs)—hides verifier cost and therefore misranks methods. Once total wall-clock time is the budget, simple Best-of-N already matches or beats guided search that repeatedly checks intermediate trajectories. The authors therefore introduce Flash-BoN: generate a large pool of inexpensive draft images by jointly applying timestep truncation, layer skipping, and activation proxies under one configuration chosen once per model, then use a multi-stage pointwise-plus-pairwise verifier to pick the best draft and finish only that draft at full quality. Across three benchmarks and three model sizes the method leads under fixed runtime, with larger gains on bigger models, and the same draft idea also speeds reflection-style prompt tuning and RL post-training. A sympathetic reader cares because it reframes how compute should be spent at test time: more diverse exploration, not more intermediate verification.

Core claim

Under matched wall-clock budgets, methods that spend compute on broader exploration of independent seeds outperform guided search that verifies and steers intermediate denoising trajectories; Flash-BoN makes that exploration still cheaper by producing many low-cost drafts from a single once-per-model acceleration configuration, selecting the best draft with multi-stage verification, and refining only the winner, consistently raising quality-versus-time AUC over BoN, BFS, DFS, and ZOS.

What carries the argument

The draft configuration ϕ*—a joint setting of early timestep stopping, contiguous layer skipping, and activation-proxy frequency, optimized once per model by dual annealing for speedup subject to LPIPS similarity to full outputs—plus a multi-stage Elo-based verifier that prunes with pointwise scores then refines with sparse and dense pairwise comparisons before selective full-quality refinement.

Load-bearing premise

That one fixed draft configuration, tuned once on a small calibration set so drafts stay roughly similar to full images, is good enough that a verifier’s ranking of those drafts still picks the image that would win after full refinement.

What would settle it

On the same models and benchmarks, measure whether the top-ranked draft under the multi-stage verifier still ranks first after full refinement for most prompts; if draft rankings reverse often, or if random/unoptimized acceleration configs match Flash-BoN under wall-clock, the central transfer claim fails.

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

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

2 major / 5 minor

Summary. The paper argues that NFE-based comparisons of inference-time scaling for text-to-image diffusion/flow models undercount verifier cost and can reverse method rankings under wall-clock budgets; under equal runtime, simple Best-of-N already matches or beats guided search (BFS/DFS/ZOS). It proposes Flash-BoN: a once-per-model draft configuration ϕ* combining timestep truncation, layer skipping, and activation proxies (chosen by dual annealing on a 120-prompt calibration set subject to LPIPS similarity), multi-stage pointwise-then-pairwise Elo selection of the best draft, and full-quality refinement of only that candidate. Across GenAI-Bench, GenEval, and UniGenBench on Wan2.1 1.3B/14B and FLUX.1-dev, Flash-BoN leads all baselines under fixed wall-clock budgets (gains larger at bigger scales), layers onto Reflection-Tuning and BFS, correlates with candidate diversity (Vendi, r=0.75), and accelerates Flow-GRPO via draft-guided rollouts.

Significance. If the wall-clock results hold, the paper usefully reframes inference-time scaling for T2I: compute is better spent expanding exploration via cheaper generation than on frequent intermediate verification, and NFE-only rankings can mislead. Strengths include end-to-end wall-clock accounting (generation, verification, refinement, transfer), explicit separation of verifier from evaluation metric with cross-metric heatmaps, progressive knob ablations (Tab. 2), multi-stage selection ablations, diversity analysis, modular combination with Reflection-Tuning (+AUC), and transfer of draft-guided selection to RL post-training (~10× fewer steps to match baseline). The once-per-model ϕ* design is practical. These are solid, reproducible systems contributions for the field.

major comments (2)
  1. [§4.1, App. C, Tab. 2] §4.1 and App. C select ϕ* by maximizing speedup subject to LPIPS similarity ≥0.6 to same-seed full-compute outputs. The load-bearing premise is that verifier rankings on these drafts transfer after refinement. Tab. 2 shows random configs collapse and progressive knobs help, but the paper does not report a direct rank-correlation (e.g., Spearman/Kendall between draft V scores and refined scores or E metrics across the candidate pool on held-out prompts). Adding that measurement would tightly validate ranking transfer rather than only absolute LPIPS fidelity.
  2. [Table 1, Fig. 3, §5] Table 1 and Fig. 3 report normalized AUC and score-vs-time curves without uncertainty (bootstrap CIs over prompts, or prompt-level paired tests beyond the Wilcoxon notes in Tab. 2). The central claim of consistent leadership and +8% AUC at larger scales would be stronger with prompt-level variance and significance for Flash-BoN vs BoN on each model–benchmark pair, especially given the 300-prompt subsamples of GenAI-Bench and UniGenBench.
minor comments (5)
  1. [Fig. 1, App. A.2] Fig. 1 and related NFE-vs-time plots mark time-to-4k-NFEs with dashed lines; stating the exact wall-clock accounting (whether verifier GPU is reserved separately, as in App. A.2) in the main caption would help readers interpret the reversal.
  2. [§4.2, Fig. 14] Default multi-stage parameters (ρ1=ρ2=0.5, κ=1.5) are stated in §4.2; a one-line pointer in the main text to the budget sweep in Fig. 14 would make sensitivity easier to find.
  3. [§5, App. A.1] For FLUX, ODE-to-SDE conversion with a=0.3 is applied only to guided baselines (App. A.1). Briefly noting in the main experiments section that Flash-BoN does not require this conversion would avoid confusion about fairness.
  4. [Eq. (13)] Minor notation: S vs S' for full vs draft steps is clear, but ϕ_full in Eq. (13) is introduced without restating that it disables all three knobs; a short parenthetical would help.
  5. [Fig. 8, §5] Fig. 8 and Fig. 18 usefully illustrate same-metric bias; ensuring figure callouts in §5 explicitly say “false positives/negatives under matched verifier–evaluator” would match the caption language.

Circularity Check

0 steps flagged

No circularity: empirical pipeline with independent calibration metric, external benchmarks, and separated verifier/evaluator.

full rationale

Flash-BoN is an empirical systems paper that combines known acceleration knobs (timestep truncation, layer skipping, activation proxies) into a once-per-model draft configuration ϕ* chosen by dual annealing on a 120-prompt calibration set to maximize wall-clock speedup subject to LPIPS similarity ≥0.6 to full-compute outputs. The resulting drafts are ranked by an off-the-shelf VLM (pointwise + multi-stage pairwise Elo) and only the winner is refined. Downstream claims are measured on held-out benchmarks (GenAI-Bench VQAScore, GenEval, UniGenBench) under matched wall-clock budgets against BoN/BFS/DFS/ZOS, with explicit separation of the inference-time verifier V from the evaluation metric E and cross-metric ablations. LPIPS is an independent perceptual proxy, not the target score; random configurations are shown to fall below BoN, confirming the optimization is not tautological. No equation equates a claimed prediction to a fitted input by construction, no uniqueness theorem is imported from overlapping authors, and no load-bearing premise reduces to self-citation. The derivation chain is therefore self-contained against external data and metrics.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central claim rests on standard diffusion/flow sampling properties, the empirical complementarity of three known acceleration knobs, a one-time black-box optimization of a discrete configuration, and the transferability of VLM rankings from drafts to refined images. No new physical entities are postulated; free parameters are the usual hyper-parameters of search and the per-model ϕ*.

free parameters (5)
  • ϕ* draft configuration (S', n_skip, L_skip, f_full) = e.g. Wan1.3B: S'=49, n_skip=3, L_skip={22,23}, f_full=5
    Chosen once per model by dual annealing on a 120-prompt calibration set; different values for Wan-1.3B, Wan-14B, FLUX (App. A.2).
  • multi-stage retention ratios ρ1, ρ2 and sparse comparisons κ = 0.5 / 0.5 / 1.5
    Defaults ρ1=ρ2=0.5, κ=1.5 control verification budget; ablated but still free design choices.
  • LPIPS similarity threshold for Pareto selection = 0.6
    Sim(ϕ)≥0.6 used to pick ϕ* from the speed-fidelity frontier.
  • ODE-to-SDE noise scale a for flow models = 0.3
    Controls stochasticity needed for BFS/DFS/ZOS branching on FLUX; set to 0.3.
  • Flash-Flow-GRPO selection ratio and KL β = 6+2 / β=0.1
    Top-6 + bottom-2 of 16 drafts; β raised to 0.1 vs baseline 0.05.
axioms (4)
  • domain assumption Diffusion/flow sampling is Markovian, so a cached latent at the draft truncation point can be resumed with full computation without depending on how the latent was produced.
    Invoked in Stage 3 (Eq. 13) to justify selective refinement.
  • domain assumption Wall-clock time (including verifier and data-transfer cost) is the correct primary efficiency metric for comparing inference-time scaling methods.
    Central methodological premise of §1 and Fig. 1; NFE rankings are treated as distorted.
  • domain assumption Off-the-shelf VLM pointwise and pairwise judgments, after multi-stage Elo filtering, are sufficiently calibrated to identify drafts whose refined versions maximize downstream benchmark metrics.
    Underpins Stage 2; authors show residual calibration issues and same-metric bias but still rely on the assumption for the main results.
  • ad hoc to paper A single configuration ϕ* optimized on a general-purpose 120-prompt calibration set generalizes across prompts and the three evaluation benchmarks without per-prompt retuning.
    Stated in §4.1; App. C.1 shows modest sensitivity to calibration set but claims robustness.
invented entities (2)
  • Flash-BoN draft configuration ϕ and its one-shot discrete optimization independent evidence
    purpose: Unifies three acceleration knobs into a single reusable draft policy that expands candidate diversity under fixed wall-clock.
    The joint configuration space and the dual-annealing selection procedure are introduced by the paper; independent evidence is the empirical Pareto frontier and transfer to multiple benchmarks.
  • Multi-stage pointwise + sparse-pairwise + dense-Elo selection procedure independent evidence
    purpose: Identify the best draft among a large pool at sub-quadratic verifier cost while mitigating score compression.
    New combination of Elo seeding, adjacent sparse pairs, and final tournament; ablated against Swiss, knockout, full pairwise.

pith-pipeline@v1.1.0-grok45 · 29510 in / 3329 out tokens · 32797 ms · 2026-07-11T18:58:06.235423+00:00 · methodology

0 comments
read the original abstract

Inference-time scaling for text-to-image generation has progressed from simple Best-of-$N$ (BoN) sampling to guided search methods that verify and steer candidate trajectories at intermediate denoising steps. These approaches focus on when and how often to verify during denoising but largely treat the cost of generation itself as fixed. Moreover, the standard practice of comparing methods by number of function evaluations (NFEs) counts only denoising forward passes and ignores verifier overhead, which can distort efficiency rankings. We show that under wall-clock evaluation, simple BoN already matches or outperforms several guided search techniques, suggesting that compute is better spent on broader exploration than on repeated intermediate verification. This motivates Flash-BoN, which generates a large pool of inexpensive draft candidates by combining three complementary acceleration knobs: timestep truncation, layer skipping, and activation proxies into a single configuration optimized once per model. An efficient multi-stage verification procedure then identifies the most promising draft, which is refined at full quality. Across three benchmarks and three model scales, Flash-BoN consistently outperforms all baselines under fixed wall-clock budgets, with gains that grow at larger model scales (+8% AUC). We further show that our strategy combines well and improves existing orthogonal techniques such as reflection-based prompt optimization (+16% AUC). The gains correlate with increased candidate diversity, which also enables draft-guided selection to accelerate RL post-training convergence.

Figures

Figures reproduced from arXiv: 2607.04461 by Gowthami Somepalli, Heng Huang, Reza Shirkavand, Ruchit Rawal, Sayak Paul, Tom Goldstein, Yizheng Chen, Yuxin Wen.

Figure 1
Figure 1. Figure 1: [T2I Model: Wan2.1 1.3B] Comparison under matched NFE and matched wall-clock budgets. Left: At a fixed NFE budget, Breadth-First Search (BFS) is marginally best at high NFEs. Right: Under equal runtime, the trend reverses. Dashed vertical lines mark the wall-clock time each method needs to reach 4k NFEs (the maxi￾mum NFE shown on the left). Methods with frequent verifier calls take longer to reach the same… view at source ↗
Figure 2
Figure 2. Figure 2: [T2I Model: Wan2.1 1.3B] Acceleration strategies for draft generation: (1) activation proxies exploit redundancies in activations across timesteps, (2) layer skip￾ping omits redundant layers within a timestep, and (3) early stopping truncates the denoising trajectory. The bottom row compares drafts that over-index on each strategy with our Flash-BoN draft (far right), which combines all three through a lea… view at source ↗
Figure 3
Figure 3. Figure 3: GenAI-Bench under fixed wall-clock budgets. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-category GenAI-Bench scores on Wan2.1 1.3B under wall-clock time bud￾gets. Flash-BoN leads across all ten categories. in every category, with the largest gains on compositional prompts (Spatial Re￾lation, Comparison, Part Relation), where the primary failure mode is incorrect layout or missing structure; a broader draft pool increases the chance of sam￾pling a valid composition. The gap narrows for fin… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison across methods. Each row shows the best image selected by each technique under a fixed 300 s wall-clock budget. Prompts span spatial relations, negation, counting, and complex scene composition. Flash-BoN consistently produces images that more faithfully satisfy the prompt. from cheaper drafts, while those that refine within a fixed region see diminishing returns. Why does broader ex… view at source ↗
Figure 6
Figure 6. Figure 6: Combining the Flash draft strategy with existing inference￾time scaling methods on Wan2.1 1.3B (GenAI-Bench). We report AUC/Time, the area under the score-vs￾wall-clock curve normalized by total time, as a summary efficiency metric. Applying Flash drafts to Reflection-Tuning yields a large improvement (0.46 to 0.62), while the gain on BFS is more modest (0.49 to 0.55). 2 4 6 8 10 12 Diversity of Generated … view at source ↗
Figure 8
Figure 8. Figure 8: Failure modes when verifier and evaluator coincide (Wan 2.1 1.3B). [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Percent improvement over no-scaling baseline (Wan 2.1 1.3B). Each cell reports relative improvement when using a given verifier (rows) and evaluating with a specific metric (columns). Diagonal entries, where veri￾fier matches evaluation metric, show the largest gains, while off-diagonal entries re￾flect cross-metric generalization. 0 60 120 180 240 300 360 420 480 540 600 660 Steps 0.684 0.686 0.688 0.690 … view at source ↗
Figure 11
Figure 11. Figure 11: [T2I Model: Wan2.1 1.3B, Dataset: GenEval] Comparison under matched NFE and matched wall-clock budgets. Left: At a fixed NFE budget, Breadth-First Search (BFS) is marginally best at high NFEs. Right: Under equal runtime, the trend reverses. Dashed vertical lines mark the wall-clock time each method needs to reach 4k NFEs (the maximum NFE shown on the left). Methods with frequent verifier calls take longer… view at source ↗
Figure 12
Figure 12. Figure 12: [T2I Model: Wan2.1 1.3B, Dataset: UniGenBench] Comparison under matched NFE and matched wall-clock budgets. Left: At a fixed NFE budget, Breadth￾First Search (BFS) is marginally best at high NFEs. Right: Under equal runtime, the trend reverses. Dashed vertical lines mark the wall-clock time each method needs to reach 4k NFEs (the maximum NFE shown on the left). Methods with frequent verifier calls take lo… view at source ↗
Figure 14
Figure 14. Figure 14: Effect of verification budget on selection quality (Wan 2.1 1.3B, GenAI-Bench). Higher pairwise bud￾gets consistently improve performance over pointwise-only scoring (dashed), with the gap widening as the candidate pool grows at larger time budgets. B.3 Verification and Selection Ablations Effect of Verification Budget on Multi-Stage Filtering Our main experiments use the exhaustive budget for multi-stage… view at source ↗
Figure 15
Figure 15. Figure 15: Comparison of pairwise ranking strategies (Wan 2.1 1.3B, GenAI-Bench). Strategies that com￾bine pointwise screening with pairwise refinement (Hybrid, Multi-Stage) consis￾tently lead. Full Pairwise is hampered at small budgets by its quadratic cost, while Knockout and Top-k Maintenance lack sufficient comparison structure for reli￾able selection. 30 60 90 120 180 300 Total Time (s) 0.60 0.64 0.68 0.72 0.76… view at source ↗
Figure 17
Figure 17. Figure 17: [T2I Model: Wan2.1 1.3B] Pareto frontier of draft configurations showing speedup versus LPIPS similarity to full-compute outputs. The selected configuration ϕ ∗ attains high similarity with substantial speedup across a wide array of discrete optimizers Effect of Discrete Optimizer We compare dual annealing against three alternative black-box optimizers: Bayesian optimization (surrogate-based sequential mo… view at source ↗
Figure 18
Figure 18. Figure 18: Example of same-metric bias when the verifier and evaluator are both Im￾ageReward. For the prompt “a plate without a banana but with two apples,” each verifier selects its top-ranked image, and all selected images are then evaluated us￾ing ImageReward. The image selected by the ImageReward verifier contains a banana and therefore violates the prompt, yet it receives the highest ImageReward score. This sho… view at source ↗
Figure 19
Figure 19. Figure 19: Expanded rollouts improve training signal for difficult prompts. [PITH_FULL_IMAGE:figures/full_fig_p033_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Diverse candidate exploration as inference-time budget grows [PITH_FULL_IMAGE:figures/full_fig_p034_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: BFS particles vs. BoN-style candidates. Top: BFS particles converge to near-identical outputs despite starting from independent seeds; pixel differences (left, scaled 4×) are imperceptible. Bottom: BoN-style candidates explore diverse composi￾tions, layouts, and styles within the same prompt. Reasoning: ⟨1-2 sentences⟩ Verdict: ⟨yes|no⟩ Score: ⟨0-10⟩ System Prompt: Pairwise Scoring You are comparing two A… 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

45 extracted references · 11 linked inside Pith

  1. [1]

    Agarwal, S., Mitra, S., Chakraborty, S., Karanam, S., Mukherjee, K., Saini, S.K.: ApproximatecachingforefficientlyservingText-to-Imagediffusionmodels.In:21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24). pp. 1173–1189. USENIX Association, Santa Clara, CA (Apr 2024),https: //www.usenix.org/conference/nsdi24/presentation/agarwal-shubham

    2024

  2. [2]

    Journal of Machine Learning Research26(209), 1–80 (2025),http://jmlr.org/papers/v26/23-1605.html

    Albergo, M., Boffi, N.M., Vanden-Eijnden, E.: Stochastic interpolants: A unifying framework for flows and diffusions. Journal of Machine Learning Research26(209), 1–80 (2025),http://jmlr.org/papers/v26/23-1605.html

    2025

  3. [3]

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J.: Qwen2.5-vl technical report (2025),https://arxiv.org/abs/2502.13923

    Pith/arXiv arXiv 2025

  4. [4]

    Accessed: June 28, 2026

    Black Forest Labs: FLUX.1 [dev].https://huggingface.co/black-forest-labs/ FLUX.1- dev(2024), 12 billion parameter rectified flow transformer for text-to- image generation. Accessed: June 28, 2026

    2024

  5. [5]

    In: 2025 IEEE/CVF International Conference on Computer Vision (ICCV)

    Chen, G., Zhao, X., Zhou, Y., Qu, X., Chen, T., Cheng, Y.: Towards stabilized and efficient diffusion transformers through long-skip-connections with spectral constraints. In: 2025 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 17708–17718 (2025).https://doi.org/10.1109/ICCV51701.2025. 01645

    doi:10.1109/iccv51701.2025 2025

  6. [6]

    Chen, P., Shen, M., Ye, P., Cao, J., Tu, C., Bouganis, C.S., Zhao, Y., Chen, T.:δ- dit: A training-free acceleration method tailored for diffusion transformers (2024), https://arxiv.org/abs/2406.01125

    Pith/arXiv arXiv 2024

  7. [7]

    Arco Publishing, New York (1978)

    Elo, A.E.: The Rating of Chessplayers, Past and Present. Arco Publishing, New York (1978)

    1978

  8. [8]

    Transactions on Machine Learning Research (2023),https:// openreview.net/forum?id=g97OHbQyk1

    Friedman, D., Dieng, A.B.: The vendi score: A diversity evaluation metric for machine learning. Transactions on Machine Learning Research (2023),https:// openreview.net/forum?id=g97OHbQyk1

    2023

  9. [9]

    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

  10. [10]

    censorship

    Guo, W.M., Qian, Q., Hasan, K., Du, S.: Aesthetic alignment risks assimilation: How image generation and reward models reinforce beauty bias and ideological" censorship". arXiv preprint arXiv:2512.11883 (2025)

    Pith/arXiv arXiv 2025

  11. [11]

    In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025),https://openreview.net/forum?id=nqVLJTPoQC

    Guo, Y., Yang, Y., Yuan, H., Wang, M.: Training-free guidance beyond differen- tiability: Scalable path steering with tree search in diffusion and flow models. In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025),https://openreview.net/forum?id=nqVLJTPoQC

    2025

  12. [12]

    In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion(CVPR).pp.8352–8361(2024).https://doi.org/10.1109/CVPR52733.2024

    Habibian, A., Ghodrati, A., Fathima, N., Sautiere, G., Garrepalli, R., Porikli, F., Petersen, J.: Clockwork diffusion: Efficient generation with model-step distilla- tion. In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion(CVPR).pp.8352–8361(2024).https://doi.org/10.1109/CVPR52733.2024. 00798

    doi:10.1109/cvpr52733.2024 2024

  13. [13]

    arXiv preprint arXiv:2505.17618 (2025)

    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)

    Pith/arXiv arXiv 2025

  14. [14]

    Lee, G., Truong, B.N.N., Yoon, J., Lee, D., Kim, M., Bengio, Y., Ahn, S.: Adaptive inference-time scaling via cyclic diffusion search. In: The Thirty-ninth Flash-BoN: Instant Drafts for Inference-Time Scaling in Diffusion Models 17 Annual Conference on Neural Information Processing Systems (2025),https: //openreview.net/forum?id=jkCH6iwFZm

    2025

  15. [15]

    arXiv preprint arXiv:2406.13743 (2024)

    Li, B., Lin, Z., Pathak, D., Li, J., Fei, Y., Wu, K., Ling, T., Xia, X., Zhang, P., Neubig, G., et al.: Genai-bench: Evaluating and improving compositional text-to- visual generation. arXiv preprint arXiv:2406.13743 (2024)

    Pith/arXiv arXiv 2024

  16. [16]

    In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV)

    Li, L., Li, H., Zheng, X., Wu, J., Xiao, X., Wang, R., Zheng, M., Pan, X., Chao, F., Ji, R.: Autodiffusion: Training-free optimization of time steps and architectures for automated diffusion model acceleration. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7082–7091 (2023). https://doi.org/10.1109/ICCV51070.2023.00654

    doi:10.1109/iccv51070.2023.00654 2023

  17. [17]

    arXiv preprint arXiv:2503.02039 (2025)

    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)

    Pith/arXiv arXiv 2025

  18. [18]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Liu, J., Zou, C., Lyu, Y., Chen, J., Zhang, L.: From reusing to forecasting: Ac- celerating diffusion models with taylorseers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15853–15863 (October 2025)

    2025

  19. [19]

    In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025),https:// openreview.net/forum?id=oCBKGw5HNf

    Liu, J., Liu, G., Liang, J., Li, Y., Liu, J., Wang, X., Wan, P., ZHANG, D., Ouyang, W.: Flow-grpo: Training flow matching models via online rl. In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025),https:// openreview.net/forum?id=oCBKGw5HNf

    2025

  20. [20]

    arXiv preprint arXiv:2209.03003 (2022)

    Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003 (2022)

    Pith/arXiv arXiv 2022

  21. [21]

    In: Pro- ceedings of the Computer Vision and Pattern Recognition Conference (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 Computer Vision and Pattern Recognition Conference (CVPR). pp. 2523–2534 (June 2025)

    2025

  22. [22]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Ma, Y., Wu, X., Sun, K., Li, H.: Hpsv3: Towards wide-spectrum human preference score. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15086–15095 (October 2025)

    2025

  23. [23]

    Transactions on Ma- chineLearningResearch(2024),https://openreview.net/forum?id=a68SUt6zFt, featured Certification

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., HAZIZA, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning robust visual feat...

    2024

  24. [24]

    Peebles,W.,Xie,S.:Scalablediffusionmodelswithtransformers.In:Proceedingsof the IEEE/CVF international conference on computer vision. pp. 4195–4205 (2023)

    2023

  25. [25]

    In: International Conference on Medical image computing and computer-assisted intervention

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi- cal image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234–241. Springer (2015)

    2015

  26. [26]

    In: Forty-second International Conference on Machine Learning (2025),https: //openreview.net/forum?id=Jp988ELppQ

    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: Forty-second International Conference on Machine Learning (2025),https: //openreview.net/forum?id=Jp988ELppQ

    2025

  27. [27]

    Rawal et al

    Song,J.,Meng,C.,Ermon,S.:Denoisingdiffusionimplicitmodels.In:International Conferenceon LearningRepresentations(2021),https://openreview.net/forum? id=St1giarCHLP 18 R. Rawal et al

    2021

  28. [28]

    In: Interna- tional Conference on Learning Representations (2021),https://openreview.net/ forum?id=PxTIG12RRHS

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. In: Interna- tional Conference on Learning Representations (2021),https://openreview.net/ forum?id=PxTIG12RRHS

    2021

  29. [29]

    Tsallis, C.: Possible Generalization of Boltzmann-Gibbs Statistics. J. Statist. Phys. 52, 479–487 (1988).https://doi.org/10.1007/BF01016429

    doi:10.1007/bf01016429 1988

  30. [30]

    arXiv preprint arXiv:2501.09685 (2025)

    Uehara, M., Zhao, Y., Wang, C., Li, X., Regev, A., Levine, S., Biancalani, T.: Inference-time alignment in diffusion models with reward-guided generation: Tu- torial and review. arXiv preprint arXiv:2501.09685 (2025)

    Pith/arXiv arXiv 2025

  31. [31]

    Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.W., Chen, D., Yu, F., Zhao, H., Yang, J., Zeng, J., Wang, J., Zhang, J., Zhou, J., Wang, J., Chen, J., Zhu, K., Zhao, K., Yan, K., Huang, L., Feng, M., Zhang, N., Li, P., Wu, P., Chu, R., Feng, R., Zhang, S., Sun, S., Fang, T., Wang, T., Gui, T., Weng, T., Shen, T., Lin, W., Wang, W., Wang, W., Zhou, W.,...

    Pith/arXiv arXiv 2025

  32. [32]

    In: The Thirteenth Interna- tional Conference on Learning Representations (2025),https://openreview.net/ forum?id=wmmDvZGFK7

    Wang, G., Cai, Y., lijiang Li, Peng, W., Su, S.Z.: PFDiff: Training-free acceleration of diffusion models combining past and future scores. In: The Thirteenth Interna- tional Conference on Learning Representations (2025),https://openreview.net/ forum?id=wmmDvZGFK7

    2025

  33. [33]

    arXiv preprint arXiv:2508.18265 (2025)

    Wang, W., Gao, Z., Gu, L., Pu, H., Cui, L., Wei, X., Liu, Z., Jing, L., Ye, S., Shao, J., et al.: Internvl3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265 (2025)

    Pith/arXiv arXiv 2025

  34. [34]

    arXiv preprint arXiv:2508.20751 (2025)

    Wang, Y., Li, Z., Zang, Y., Zhou, Y., Bu, J., Wang, C., Lu, Q., Jin, C., Wang, J.: Pref-grpo: Pairwise preference reward-based grpo for stable text-to-image rein- forcement learning. arXiv preprint arXiv:2508.20751 (2025)

    Pith/arXiv arXiv 2025

  35. [35]

    In: The Fourteenth International Conference on Learning Representations (2026),https: //openreview.net/forum?id=dnJEHl6DI1

    Whitehouse, C., Wang, T., Yu, P., Li, X., Weston, J.E., Kulikov, I., Saha, S.: J1: Incentivizing thinking in LLM-as-a-judge via reinforcement learning. In: The Fourteenth International Conference on Learning Representations (2026),https: //openreview.net/forum?id=dnJEHl6DI1

    2026

  36. [36]

    In: The Thirteenth International Conference on Learning Representations (2025),https: //openreview.net/forum?id=VNckp7JEHn

    Wu, Y., Sun, Z., Li, S., Welleck, S., Yang, Y.: Inference scaling laws: An em- pirical analysis of compute-optimal inference for LLM problem-solving. In: The Thirteenth International Conference on Learning Representations (2025),https: //openreview.net/forum?id=VNckp7JEHn

    2025

  37. [37]

    Physics Letters A233(3), 216–220 (1997).https://doi.org/https://doi.org/10.1016/S0375-9601(97)00474-X, https://www.sciencedirect.com/science/article/pii/S037596019700474X

    Xiang, Y., Sun, D., Fan, W., Gong, X.: Generalized simulated annealing algorithm and its application to the thomson model. Physics Letters A233(3), 216–220 (1997).https://doi.org/https://doi.org/10.1016/S0375-9601(97)00474-X, https://www.sciencedirect.com/science/article/pii/S037596019700474X

    doi:10.1016/s0375-9601(97)00474-x 1997

  38. [38]

    In: Forty-second International Conference on Machine Learning (2025),https://openreview.net/ forum?id=27hOkXzy9e

    Xie, E., Chen, J., Zhao, Y., YU, J., Zhu, L., Lin, Y., Zhang, Z., Li, M., Chen, J., Cai, H., Liu, B., Zhou, D., Han, S.: SANA 1.5: Efficient scaling of training- time and inference-time compute in linear diffusion transformer. In: Forty-second International Conference on Machine Learning (2025),https://openreview.net/ forum?id=27hOkXzy9e

    2025

  39. [39]

    In: Proceedingsofthe37thInternationalConferenceonNeuralInformationProcessing Systems

    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: Proceedingsofthe37thInternationalConferenceonNeuralInformationProcessing Systems. pp. 15903–15935 (2023) Flash-BoN: Instant Drafts for Inference-Time Scaling in Diffusion Models 19

    2023

  40. [40]

    Advances in Neural Information Processing Systems36, 76806–76838 (2023)

    Xu, Y., Deng, M., Cheng, X., Tian, Y., Liu, Z., Jaakkola, T.: Restart sampling for improving generative processes. Advances in Neural Information Processing Systems36, 76806–76838 (2023)

    2023

  41. [41]

    In: Christodoulopoulos, C., Chakraborty, T., Rose, C., Peng, V

    Xuan, W., Zeng, Q., Qi, H., Wang, J., Yokoya, N.: Seeing is believing, but how much? a comprehensive analysis of verbalized calibration in vision-language mod- els. In: Christodoulopoulos, C., Chakraborty, T., Rose, C., Peng, V. (eds.) Pro- ceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. pp. 1408–1450. Association for ...

    doi:10.18653/v1/2025.emnlp-main.74 2025

  42. [42]

    In: Forty-second International Conference on Machine Learning (2025),https://openreview.net/forum?id=XrCbBdycDc

    Yoon, J., Cho, H., Baek, D., Bengio, Y., Ahn, S.: Monte carlo tree diffusion for system 2 planning. In: Forty-second International Conference on Machine Learning (2025),https://openreview.net/forum?id=XrCbBdycDc

    2025

  43. [43]

    In: The Fourteenth International Con- ference on Learning Representations (2026),https://openreview.net/forum?id= b7Ftp6U78i

    Zhang, X., Lin, H., Ye, H., Zou, J., Ma, J., Liang, Y., Du, Y.: Inference-time scaling of diffusion models through classical search. In: The Fourteenth International Con- ference on Learning Representations (2026),https://openreview.net/forum?id= b7Ftp6U78i

    2026

  44. [44]

    In: Glober- son, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., Zhang, C

    Zhu, H., Tang, D., Liu, J., Lu, M., Zheng, J., Peng, J., Li, D., Wang, Y., Jiang, F., Tian, L., Tiwari, S., Sirasao, A., Yong, J., Wang, B., Barsoum, E.: Dip-go: A diffusion pruner via few-step gradient optimization. In: Glober- son, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., Zhang, C. (eds.) Advances in Neural Information Processing ...

    2024

  45. [45]

    yes” to the question “Does this image match the description: {prompt}?

    Zhuo, L., Zhao, L., Paul, S., Liao, Y., Zhang, R., Xin, Y., Gao, P., Elhoseiny, M., Li, H.: From reflection to perfection: Scaling inference-time optimization for text- to-image diffusion models via reflection tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15329–15339 (October 2025) 20 R. Rawal et al. Flash...

    arXiv 2025