acceptodds
Withdrawn from ICLR 2027

CLOSER: Closing the Feedback Loop for Inference-Time Scaling of Diffusion Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Inference-time scaling aims to improve diffusion-model outputs by allowing more time and computation during generation. Using task feedback to direct this computation requires turning a terminal assessment into an effective correction. Our controlled diagnostics show the challenge: informative evaluator gradients can still yield harmful updates, and the effective intervention stage varies across objectives. We therefore introduce CLOSER, a training-free controller that turns terminal feedback into verified, position-aware state corrections. At an initial or intermediate sampler state, CLOSER differentiates through the remaining discrete sampler to construct a bounded candidate set. It executes these candidates through the unchanged suffix and accepts a correction only after terminal evaluation, keeping the incoming trajectory available for rollback. After acceptance, refreshed feedback guides the next proposal from the updated sampler state. This separation lets gradients guide the search while terminal criteria determine which trajectory is retained. Across SD3.5-Medium and Z-Image on DPG-Bench, OneIG-EN, and GenEval, CLOSER achieves the strongest reported terminal utility across all evaluated backbone–benchmark settings, exceeding the strongest GenEval baselines by 0.113 and 0.106. The evaluated operating points also outperform the compared utility–cost baselines. Position analysis connects these gains to correction scale: early interventions favor structural changes, whereas late interventions favor preference and alignment refinement in the controlled study.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
diffusion model, inference time scaling, text-to-image generation

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_TviNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CLOSER, a training-free inference-time scaling method for frozen diffusion/flow models with evaluator access. At one or a few fixed sampler positions, CLOSER (i) computes the gradient of a differentiable proposal objective JJ by back-propagating through the exact remaining sampler and decoder, (ii) builds a small candidate set by moving the intermediate state along the normalized gradient with a fixed set of step sizes, (iii) completes each candidate with the unchanged suffix and scores the final image with a terminal selector VV, optionally subject to a preservation guard HH, and (iv) accepts the best feasible candidate or rolls back to the incoming state, recomputing the gradient after each acceptance. The authors separate proposal (J), selection (V), preservation (H), and reporting (M) roles. On SD3.5-Medium and Z-Image, CLOSER achieves the best scores on GenEval, DPG-Bench, and OneIG-EN. Controlled studies on SD3.5M analyze the role of each component and argue that early interventions favor structural changes while late interventions favor preference refinement.

Critical strengths
  • Strong empirical results. CLOSER ranks first in all six backbone–benchmark settings, with large margins on GenEval and DPG-Bench (Table 1), and the paired confidence intervals against Standard are all above zero (Fig. 4d).

  • Interesting position analysis. The link between intervention position, response frequency, and local directional reliability (Figs. 6–7) is an informative observation about where gradient-based corrections are trustworthy.

Critical weaknesses
  • Missing baselines. Particle-based (SMC) steering and path-search methods (e.g., FK steering [1], SVDD [2], DAS [3], and the search-over-paths variant of [4]) are not compared, so the benefit of closed-loop correction over alternative ways of spending the same compute remains unclear for me.

  • Limited generality. Positions, iterations, step-size menus, proxies, and selection rules differ across benchmarks and backbones (Tables 6–9), while baselines use one frozen setting, and the selection procedure is not described for GenEval and DPG. The intervention position must be chosen by the practitioner and does not transfer across backbones (Table 8), and this calibration cost does not appear to be included. Moreover, the method requires a differentiable proxy well aligned with the evaluator and white-box access to the generator, which limits its applicability compared with gradient-free approaches.

  • Cost reporting. Only wall-clock time is reported (up to ~50× on DPG, Fig. 3). NFE, peak memory, and per-prompt cost distributions are missing, even though the method back-propagates through the DiT.

  • Clarity. The paper packs many experiments into brief main-text descriptions, so they are hard to follow.

[1] R. Singhal et al. A General Framework for Inference-time Scaling and Steering of Diffusion Models. ICML 2025. arXiv:2501.06848.

[2] X. Li et al. Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-Based Decoding. NeurIPS 2025. arXiv:2408.08252.

[3] S. Kim et al. Test-time Alignment of Diffusion Models without Reward Over-optimization. ICLR 2025. arXiv:2501.05803.

[4] N. Ma et al. Inference-Time Scaling for Diffusion Models beyond Scaling Denoising Steps. CVPR 2025. arXiv:2501.09732.

Questions for authors
  1. Could the authors compare against particle-based or path-search methods (e.g., FK steering, SVDD, search-over-paths) at matched compute?
  2. Could the authors report performance against NFE, peak memory, and per-prompt cost distributions?
  3. CLOSER’s gains depend strongly on how well the proxy J aligns with the selector V (Table 20). Could the authors evaluate a setting where J is weakly aligned with V (e.g., the mismatched pairs in Table 20) against gradient-free methods such as SMC-based steering at matched compute?
  4. Best-of-K uses the benchmark evaluator for reranking, yet its score decreases with K on SD3.5M GenEval (Fig. 3) and Best-of-4 falls below Standard on SD3.5M DPG (Table 1). Is there an explanation for this?
Reviewer_HfuAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CLOSER uses inference time compute to improve generated images from a frozen diffusion model. The authors address the key challenge of turning terminal assessment / gradients of generated image w.r.t. downstream task/reward into an effective correction signal in the intermediate denoising steps of the diffusion model. To this end, they propose CLOSER, that computes gradients /denoising steps thru the remaining sampling steps and uses them to propose position-aware changes to the initial noise or intermediate states. It then generates the final / terminal denoised image per candidate (candidates from the proposed chnages based on the grads of the remaining steps), and checks whether the change helped per candidate. Accepted changes become the starting point for the next update otherwise the method keeps the previous state. Experiments with SD3.5M and Z-Image on DPG_bench, OneIG-EN, GenEval show significant gains. Further analyses and ablations examine which parts of the method help and when to intervene during sampling.

Critical strengths
  1. The proposed method is grounded in / motivated by a useful and clearly explained problem - gradients can suggest a change without telling us whether the final image will actually be better w.r.t. the task-based reward. CLOSER checks the fully denoised image before keeping the proposed change. To support the claims and the proposed method, the experimental evaluations report substantial gains over the existing state of the art in this space across the two diffusion models (Table 1). In terms of presentation and writing, the clear separation of J, V, H, and M also helps readers clearly understand each modular component.

  2. There are experiemnts to understand why the method works. There are tests on proposal directions, step sizes, gradient refresh, and the choice of selector (Table 2, Figure 5 and Table 20). The stage experiments include both selected corrections and random perturbations, which helps separate the effects of selection from the behavior of the remaining sampler (Figures 6-7, Table 22). I also appreciate that the authors report cases where improving the target hurts other scores, and test guards to reduce this trade-off (Tables 13 and 21). These results help explain both the method's benefits and its limits.

Critical weaknesses
  1. It isn't clear to me how much of the gain comes from CLOSER's acceptance rule and intervention posns. vs. repeatedly improving the initial noise. The ablations and costs weeps are useful, but I would like a comparison that can separate the aforementioned choices. One baseline for comparison could repeatedly update the initial noise using the grads of the final-image score, generate the image after each update, and then return the best image it found / searched for. Unlike Closer, this baseline would continue from each update irrespective of the fact that the update made the score better. Compare it with CLoser restricted to the initial noise and with Closer at its chosen positions. The proposal score J provides grads, and V chooses the best returned image for all three comparison methods/baselines here. That being said, if the existing Fig.3 runs already provide this comparison then explaining that in more details would address this concern. From a practical standpoint of completing experiments in the limited rebuttal time period, one diffusion model and one benchmark would suffice for this experiment check.

  2. The evidence is stronger for improving the optimized score than for improving the image more broadly. Table 21 shows that more search thru increase in K can raise the target score while lowering HPSv2 or Aesthetic scores. On the same prompts and seeds, I would suggest comparing outputs from the final benchmark settings at 2 or 3 runtime budgets. Compare Closer with the baseline that performs best in the paper's cost sweep at each corresponding budget. Have a judge unused by either method assess prompt satisfaction and visual quality separately. This would show whether extra search improves the intended task and what it costs in image quality.

Overall, the method is useful and promising, and the paper includes substantial experiments. My main reservation is about whether the gains hold against repeated initial-noise updates that retain the best image, at the same runtime. Better evidence about point 2 would strengthen the case towards an accept further.

Questions for authors
  1. Point (1) of Critical Weaknesses, and the suggested experiment.
  2. Point (2) of Critical Weaknesses, and the suggested experiment.
Reviewer_jipLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents CLOSER, a training-free controller for improving diffusion-model generation through additional inference-time computation. The method intervenes at initial or intermediate sampler states while keeping the generator and evaluators frozen. It differentiates a task objective through the exact remaining sampling steps and uses the resulting gradients to construct a finite set of candidate state corrections. Candidates are assessed by running the remaining sampler without changing its settings. A terminal selector accepts an improving candidate that satisfies any configured preservation constraints, or retains the incoming state. After acceptance, the method recomputes feedback from the updated trajectory before proposing further corrections.

A central contribution is the separation of proposal generation from terminal acceptance, together with explicit roles for preservation criteria and reporting metrics. The paper also investigates intervention timing. Controlled studies on SD3.5-Medium indicate that early corrections favor structural changes, whereas later corrections favor preference and text–image alignment. These findings motivate selecting intervention positions according to the objective.

Validation uses SD3.5-Medium and Z-Image on DPG-Bench, OneIG-EN, and GenEval. CLOSER achieves the highest reported scores among the compared methods in all six settings, reaching GenEval scores of 0.871 and 0.905. These exceed the strongest compared baselines by 0.113 and 0.106, respectively. The evaluation includes comparisons of quality against wall-clock cost, component ablations, paired bootstrap intervals, independent automated judging, and experiments on candidate budgets, gradient refresh, and preservation constraints. The main benchmark results allow access to the corresponding evaluators during generation and should be interpreted within that protocol. The formal non-decrease guarantee applies to the configured selector; transfer to other measures of image quality is assessed separately.

Critical strengths
  1. A practical feedback loop with explicit acceptance criteria. The main strength is that CLOSER separates proposing a correction from deciding whether to keep it. Gradients are computed through the actual remaining sampler, but acceptance depends on the completed image and any preservation constraints. Keeping the incoming state available for rollback guarantees that the configured selection score does not decrease under the stated deterministic setup. Refreshing the gradient after an accepted correction also makes subsequent proposals respond to the updated trajectory. This combination addresses the practical problem demonstrated in the paper: a useful gradient direction can still lead to a harmful finite update. The component ablations support the roles of informed directions, multiple displacement choices, and refreshed feedback. This supports acceptance because it provides an implementable approach to using evaluator feedback without retraining the generator.

  2. Consistent benchmark gains supported by useful diagnostic experiments. CLOSER achieves the highest reported scores among the compared methods across both generators and all three benchmark families. The GenEval improvements of 0.113 and 0.106 over the strongest included baselines are substantial, and the accompanying wall-clock measurements make the computational tradeoff explicit. Paired analyses provide uncertainty estimates for improvements over standard generation. Beyond aggregate scores, the controlled SD3.5-Medium experiments show that intervention position changes which objectives benefit: early corrections favor structure, while later corrections favor preference and alignment. This gives the reader a practical reason to choose where to intervene instead of treating every sampler state as equivalent. These results strengthen my assessment because the method is supported by repeated benchmark gains and experiments that help explain how to deploy it.

Critical weaknesses
  1. The competitiveness of the optimization baselines is not sufficiently established. Table 11 specifies one optimization update for DOODL and ReNO, whereas CLOSER permits multiple accepted corrections and searches over candidate displacements. Figure 3 reports wall-clock sweeps, but the connection between those curves and the listed baseline settings is unclear. In particular, the manuscript does not clearly document how optimization iterations and hyperparameters were varied to obtain the baseline frontiers. This does not show that the comparisons are unfair, but it leaves an important alternative explanation for the gains unresolved. Since efficient use of additional computation is a central claim, clearer sweep configurations and comparisons against competitively tuned iterative baselines at matched cost are needed to establish the advantage of the proposed control loop.

  2. Evidence of improvement beyond the optimized evaluators remains limited. Access to benchmark evaluators is an explicit part of the protocol, so using them during generation is not itself a flaw. However, improvements on these scores provide limited evidence of broader image-quality gains. Table 21 shows that increasing the candidate budget can improve the target score while worsening independent HPSv2 or Aesthetic scores, particularly for GenEval and CLIPScore selection. The InternVL3 audit in Figure 4(b) is helpful, but its sample size, judging prompts, and scoring procedure are not clearly documented. The preservation experiments demonstrate control over registered metrics on a 40-prompt confirmation set, but do not establish independent quality improvements for the main benchmark configurations. This is consequential because the practical value of spending more inference compute depends on whether the resulting images improve beyond the particular evaluator used for selection.

Questions for authors
  1. Clarification of the baseline comparisons. Table 11 lists one optimization update for DOODL and ReNO, whereas CLOSER can perform several accepted corrections. How were the baseline frontiers in Figure 3 constructed from these settings? In particular, which update counts and learning rates were swept for each backbone and benchmark, and how were configurations selected? Were the iterative baselines allowed additional updates at wall-clock budgets comparable to CLOSER’s reported operating points? Please specify the configurations behind the points nearest CLOSER’s operating costs, together with their measured runtimes. If additional iterations were tested, how did their scores change? This information would help establish whether the gains reflect a more effective feedback loop or differences in how the available computation was allocated.

  2. Clarification of the independent evaluation. Could you provide the protocol for the InternVL3 audit in Figure 4(b), including the number and selection of prompts, the judging instructions, the scoring rule, and the uncertainty calculation? Were these images produced using the same CLOSER configurations reported in Table 1, and was InternVL3 excluded from configuration selection as well as generation? Table 21 shows that higher target scores can accompany lower independent HPSv2 or Aesthetic scores. Do the available independent evaluations show whether the main benchmark configurations exhibit a similar tradeoff? Please distinguish evidence for those configurations from the smaller diagnostic and preservation studies using 40 confirmation prompts. This clarification would affect how broadly I interpret the reported improvements: gains on the chosen evaluator are supported, while improvements in other aspects of image quality require separate evidence.

Reviewer_8xLpOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

CLOSER improves the outputs of a frozen text-to-image model at inference time. At an initial or intermediate sampler state, it backpropagates a reward through the remaining sampler. It then tries a few step sizes along that gradient, runs each candidate to completion, and keeps the best candidate if the terminal evaluator score improves; otherwise it keeps the incoming state. The loop repeats from the accepted state. On SD3.5-Medium and Z-Image, CLOSER is best on DPG-Bench, OneIG-EN and GenEval, with the benchmark's own evaluator as the selector. Smaller studies ablate the loop's components and report that early interventions mainly fix structure, while late ones mainly improve preference and alignment scores.

Critical strengths

"Gradients propose; completed rollouts decide" is a simple, useful idea.

The ablations (random vs. gradient directions, step-size menu, rollback, gradient refresh, guards) are informative and well designed.

Large gains on two modern backbones and three benchmarks, concentrated on the hardest prompts.

The configurations and protocols are documented in detail.

Critical weaknesses
  1. A blind human evaluation against the strongest baseline at matched compute, judging both prompt faithfulness and artifacts, is needed. Held-out automatic judges would help but cannot replace it. The main results optimize the test metric. The selector is the official benchmark evaluator, and the GenEval search stops once that evaluator reports success. Table 1 therefore measures how well each method maximizes the scorer, not prompt following. I would like to see a more independent evaluation.

  2. Compute and baselines.

    • The costs of the Table 1 configurations are not reported. From Fig. 3 they are roughly 10× base generation on GenEval, 40–64× on DPG-Bench and 3–4× on OneIG, and the gains are largest where CLOSER spends most (on OneIG the lead is about 0.01).
    • In Table 1 the search baselines use only 3–4 candidates. ReNO and DOODL get a single update (ReNO normally runs 50 iterations), and no baseline gets rollback.
    • Memory and latency are not reported. CLOSER's iterations are sequential, whereas Best-of-K runs in parallel.
    • The comparison with seed-search baselines is only empirical. How were partial trajectories scored for pruning? Why does local correction beat selection among seeds? Are the two complementary?
  3. Novelty. Without the terminology, CLOSER is greedy hill-climbing on the terminal reward: a gradient direction, a line search over step sizes, accept-if-improve. Each ingredient has precedents: optimizing noise through the sampler (DOODL; DNO, Tang et al., 2024; D-Flow, Ben-Hamu et al., 2024), accept-if-improve search (Ma et al., 2025), and reward-based selection of intermediate states (FK steering, Singhal et al., 2025). DNO, D-Flow and FK steering are not cited. Proposition E.1 holds by construction.

  4. Tuning protocol. Where do the development and confirmation prompts and the 10 OneIG position-search examples come from? How were the GenEval and DPG configurations chosen? The confirmation prompts are scored with GenEval's internal score, so they are likely GenEval prompts. Any overlap with the evaluation sets should be stated, and results on held-out prompts reported.

Questions for authors
  1. Can you provide a blind human evaluation (faithfulness and artifacts) against the strongest baseline at matched compute. Can you show an evaluation with held-out automatic judges? Which three InternVL3 settings were significant?
  2. What are the cost, peak memory and per-image latency of each Table 1 configuration? How do the search baselines perform at the same cost? On DPG-Bench, why does Best-of-K, reranking with the DPG evaluator itself, not improve with more candidates?
  3. How do ReNO and DOODL perform with their usual iteration counts and with rollback? How does CLOSER compare with DNO, D-Flow and FK steering?
  4. How were partial trajectories scored for the pruning baselines? Does seed search followed by CLOSER beat either method alone at matched compute?
  5. Where do the development, confirmation, position-search and audit prompts come from, and do they overlap the evaluation sets? How much remains with one configuration shared across benchmarks?

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

Related papers

Discussion (0)

Sign in to comment.