acceptodds
Under review as a conference paper at ICLR 2027

PreviewDiff: Multimodal Critic-Guided Search over Diffusion Latents

Abstract

Diffusion models can produce striking images and videos, but they still struggle with the compositional details that make a generation faithful to a prompt, such as object counts, attribute binding, spatial relations, and temporally grounded actions. A common way to improve prompt satisfaction is to spend more compute at test time through Best-of-N sampling, but final-sample selection is fixed. Best-of-N can only choose among completed outputs and cannot repair a promising trajectory before it fails. We introduce PreviewDiff, a training-free test-time search method that turns diffusion sampling from scalar search into a multimodal critic-guided search over intermediate latents. At selected denoising checkpoints, PreviewDiff decodes a partial preview, asks a multimodal judge to score and critique it, and uses the resulting natural-language feedback to branch over semantic prompt edits and locally re-noised latent continuations. These branches are scored, pruned, and selectively rolled forward, allowing verifier compute to guide generation while the sample is still editable. Across image and video generation benchmarks, PreviewDiff consistently improves over budget-matched Best-of-N selection and strong scalar-search baselines. Ablations show that earlier interventions and increased search width provide the largest gains, while deeper search and additional semantic variants offer complementary improvements. PreviewDiff demonstrates that multimodal feedback is most useful not only as a final verifier, but as an active controller inside the denoising process.

Primary area:
generative models
Keywords:
visual generation, test-time scaling, diffusion models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_X2MfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PreviewDiff, which is a test-time scaling method. PreviewDiff differs from previous test-time scaling methods for diffusion models in some respects:

  • Unlike Best-of-N, they preview intermediate diffusion generation particles and evaluates on them.
  • The method uses a multimodal judge to score and critique, and utilizes resulting natural-language feedback to improve prompts
  • The method uses tree search to choose promising particles

The method shows better scalability than Best-of-N, EvoSearch, and Particle-Sample on three image-generation benchmarks and also performs better on video generation. The main contribution of this paper is to demonstrate that multimodal feedback is most useful not only as a final verifier, but as an active controller inside the denoising process.

Critical strengths
  • The authors conducted experiments on both image and video generation across five base models, and they showed that the proposed method beats three or four baseline test-time alignment methods in 4 candidate settings.
  • The main message of the paper (multimodal feedback is most useful not only as a final verifier, but as an active controller inside the denoising process) sounds promising for inference-time scaling of multimodal generation.
  • The paper also provided human evaluation results on image and video generation, which show that the gain from the proposed method is not just quantitative, but also qualitative.
Critical weaknesses
  1. Comparison of inference-time scaling may be insufficient.
  • Table 1 and Table 2 show comparisons with baseline methods, but the computational cost settings for the proposed method and the baselines are not obviously shown. The paper should show scaling-efficiency comparisons (like Figure 2) for all baseline methods. Baselines may beat the proposed method when the computational cost scales.
  • Scaling on wall-clock time should be shown. A multimodal judge for intermediate latents may take a lot of time, and the proposed method may not be efficient in wall-clock time. Although the authors note this limitation in the paper, it remains a critical concern for real-world use.
  1. Previewing the intermediate latent does not seem to be a novel idea.
  • Previewing latent reward has already been discussed in previous works [1,2,3]. The authors should discuss how their proposed method differs from these previous works.
  • The novelty seems to lie in the use of a multimodal judge and prompt modification.

[1] Yeh et al. Training-free Diffusion Model Alignment with Sampling Demons. ICLR 2025.
[2] Oshima et al. Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search. NeurIPS 2025.
[3] Wang et al. Image Diffusion Preview with Consistency Solver. CVPR 2026.

Questions for authors

Please see Critical Weakness and address the concerns. If the authors address the concerns sufficiently, I will consider raising the score.

4 comments

We thank the reviewer for their thoughtful comments. We are glad the reviewer appreciated our numerous experiments across image and video models, the message of our paper, and the human evaluation.

W1: Comparison of inference-time scaling may be insufficient.

Figure 2 evaluates scaling against Best-of-N, whereas Tables 1–2 compare additional search methods at selected budget settings. We agree that these results do not establish dominance over every baseline across all compute budgets, and have narrowed the wording accordingly.

We also report end-to-end latency over 295 NarrLV prompts on one NVIDIA H100 using LTX-Video and Gemini-3.1-Flash-Lite. Timing includes generation, decoding, critic requests, and required media transfers. Model loading and five warm-up prompts are excluded. Best-of-3 uses two sequential generation passes and up to two concurrent critic requests. Values below are mean ± standard deviation in seconds per prompt.

Method / configurationLatency (s/prompt)
PreviewDiff (86 aggregate denoising steps)9.80 ± 0.99
Best-of-N (N = 3)5.66 ± 1.20
EvoSearch9.01 ± 1.01
Particle sampling6.88 ± 0.87

PreviewDiff is slower at these configurations but improves results substantially. These measurements make the end-to-end latency tradeoff explicit. The protocol, results, and scope of these comparisons are highlighted in blue.

W2: Previewing intermediate latents does not appear novel.

We agree that intermediate previewing and VLM-based evaluation are not themselves novel. Sampling Demons [1] controls denoising noise using reward feedback, including VLM-derived rewards. Diffusion Latent Beam Search [2] uses lookahead reward estimates to select latent trajectories. ConsistencySolver [3] improves low-step preview quality and consistency for preview-and-refine workflows. Our focus is using a critic's natural-language diagnosis to construct corrective conditioning updates within the active diffusion trajectory. Positive and negative updates are combined with local re-noising to create candidate continuations, which are then evaluated and pruned. The contribution is this integration of semantic corrective actions with latent search, not previewing or multimodal scoring alone. We have expanded Section 2 to make these distinctions explicit.

Thank you for the response. W2 is addressed, but W1 is not sufficiently addressed.

I believe that in the inference-time alighment/scaling field, what matters is whether performance scales better than other methods with the compute spent. However, the response narrows the wording instead of showing this.

Therefore, I will keep my score.

Thank you. Apologies for the delay, we needed additional time to run an experiment on wall-clock time, leading us to initially narrow our positioning on the scaling experiment. We have now added a wall-clock scaling comparison using SDXL and Gemini on a single NVIDIA H100. For each PreviewDiff configuration in Figure 2, we measured its per-prompt runtime. For example, 86 aggregate denoising steps correspond to 2.77 seconds. We then used the measured Best-of-N runtime to determine the candidate count corresponding to that time budget, rounding N upward to the next integer. This approximately matches the latency budgets, with upward rounding allowing Best-of-N additional computation. Both methods execute serially under the same single-GPU allocation. PreviewDiff achieves higher mean Gemini scores at every evaluated budget, including approximately 2.86 versus 2.78 at the largest budget. This directly extends the evaluation from denoising-step scaling to quality under measured wall-clock budgets. We include the figure here. We will include this for our video model setting as well in the final submission but we believe this provides strong evidence for the superior scaling of PreviewDiff.

Thank you for the additional experiment and response.

However, Best-of-N (BoN) is the simplest and most naive baseline; demonstrating scalability only against BoN is insufficient to establish your method's scalability advantage over existing inference-time scaling approaches.

Currently, all existing baselines except BoN are evaluated only at a single fixed compute budget. To adequately demonstrate your claims, you must plot scaling curves for all baseline methods and evaluate these comparisons not only against compute budgets (e.g., denoising steps) on the horizontal axis, but also against wall-clock time.

Because these thorough analyses and discussions are missing, I believe the paper does not yet meet the bar for a top-conference work on inference-time alignment. Therefore, I will maintain my current rating.

Reviewer_5pQkOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how to spend test-time compute cost to improve compositional prompt following in diffusion models. It introduces PreviewDiff, a training-free test-time search method for image and video diffusion. At selected denoising checkpoints, it decodes an intermediate latent into a preview. A multimodal critic scores the preview and provides a short correction. The method then branches over these prompt edits, restart depths, and latent seeds. Each child is re-noised and denoised to the next checkpoint. A beam-pruned search keeps the best continuations, and the best-scoring final output is returned. The method is training-free and is evaluated on pretrained generators that expose intermediate latent states and support local restarts.

As for the contributions, the main contribution is using a multimodal model critic as both a verifier and a source of semantic search action during denoising. The method combines intermediate preview evaluation, prompt-level correction, local re-noising, and beam-pruned search. This makes the diffusion trajectory editable before completion. The paper also applies the framework to both image and video generation.

For about the validation, the main baseline is Best-of-N under the paper’s reported verifier-call or denoising-step matching protocols. EvoSearch, particle sampling, and Video-T1 are used as additional baselines. In Table 1, PreviewDiff outperforms Best-of-N on GenEval2 across three image backbones: 2.292 vs 2.090 on SDXL, 1.965 vs 1.754 on SD-3.5, and 3.112 vs 2.775 on FLUX.2. Table 2 reports the same numerical pattern for the rows attributed in the text to the proposed method: 3.021 vs. 2.716 on LTX-Video and 3.541 vs. 3.362 on Wan 2.2 on T2V-CompBench. However, these rows are labeled “MCTS-CGD” and require clarification. The paper also reports higher scores under Qwen-3, BLIP-VQA, LLaVA, and Qwen-2.5 evaluation, as well as favorable human preferences from nine raters. Figure 5 ablates checkpoint timing, tree depth, semantic variants, and search width. Table 3 ablates language feedback, prompt editing, and the number of roots.

In conclusion, I lean toward weak rejection. The method is understandable and the evaluation is broad. However, the ablations do not isolate the semantic-feedback mechanism that the paper claims as its main contribution, and the comparison against the strongest video baseline is not budget-matched. Both concerns can be addressed with additional experiments.

Critical strengths
  1. The paper introduces a clear search method and formulation for intermediate diffusion states. PreviewDiff allows the critic to identify an error while the latent is still editable, then uses a correction prompt and local restart to continue the same trajectory. This is a meaningful distinction from Best-of-N, which only ranks completed samples. The design gives the paper a concrete algorithmic contribution that can be tested through budget comparisons and ablation studies.

  2. The paper evaluates the proposed search across multiple generator backbones and test-time budgets. Table 1 reports higher Gemini P@4 scores than Best-of-N on GenEval2 for SDXL (2.292 vs. 2.090), SD-3.5 (1.965 vs. 1.754), and FLUX.2 (3.112 vs. 2.775). Figure 2 shows that the gap remains at every tested denoiser-step budget: on SDXL, the score increases from 2.17 to 2.86 for PreviewDiff compared with 1.91 to 2.71 for Best-of-N, while on LTX-Video it increases from 1.99 to 2.39 compared with 1.75 to 2.13. Table 3 further shows that removing language feedback lowers the reported score from 2.292 to 2.063 on SDXL and from 2.167 to 2.014 on LTX-Video. Together, these results provide evidence that intermediate critic feedback can improve prompt following across multiple generators.

Critical weaknesses
  1. The ablation studies do not isolate the contribution of semantic prompt correction from the other components of the search. The paper presents language feedback and semantic prompt correction as central parts of the search mechanism. In Table 3, removing prompt editing changes the SDXL GenEval2 score only from 2.292 to 2.285, while removing language feedback entirely drops it to 2.063. The paper gives only a high-level description of these two ablations, and Table 3 reports no variance. Therefore,I am concerned that this leaves unclear what language feedback contributes when prompt editing is removed. Figure 5 raises the same concern: the largest gain comes from search width, which increases the number of explored and retained trajectories. On SDXL, the reported score rises from 1.97 to 2.74, but this gain is not shown at a fixed end-to-end cost. In addition, the paper does not compare against a baseline that applies the same critic feedback only after completion. An iterative critique-and-regenerate baseline under the same reported verifier-call or denoising-step protocol would directly test whether feedback is more effective during denoising, but this baseline is not included.

  2. The comparison against Video-T1 is not controlled by a explicitly matched budget. The margin is smaller under the LLaVA-Eval column, which uses a different evaluator, while the VBench 2.0 column still uses Gemini P@4. Section 4 states that the Video-T1 results come from "reported verifier-guided video search settings where available," so Table 2 does not compare against Video-T1 under a matched budget. On the Gemini metric, PreviewDiff leads Video-T1 on LTX-Video by a clear margin (3.021 vs 2.861). Under LLaVA-Eval, the margin narrows to 0.4523 vs 0.4329. On VBench 2.0, it is 3.162 vs 3.124. The headline metric is produced by the same model that guides the search. Thus my concern is that, without a budget-matched Video-T1 run, it is hard to tell how much of the reported advantage is a real gain in prompt following.

Questions for authors

Clarification Requests:

  1. It would be appreciated if you can add a final-stage feedback baseline: critique the completed image or video with the same judge, append the fix notes to the prompt, and regenerate, under the same critic-call and denoiser-step budget as PreviewDiff. This is the direct control for the paper's central claim.

  2. In Table 3, what exactly remains when language feedback is kept but prompt editing is removed? Please define both ablations and explain why removing prompt editing costs only 0.007 on GenEval2 if semantic prompt edits are the main search action.

  3. Are the Video-T1 rows in Table 2 your own runs under the matched-budget protocol, or numbers taken from the original paper? If they are reported numbers, can you rerun Video-T1 with the same verifier-call and denoising-step protocols on LTX-Video and Wan 2.2?

  4. Do the critic-call budgets include the calls that generate the correction proposals and the VAE decodes of the previews? What is the wall-clock latency of the largest Figure 2 configuration compared with Best-of-N running in parallel?

  5. In Table 2, is “MCTS-CGD” a typo for PreviewDiff, or does it denote a different method? Please map every reported row to an exact algorithm and use the same name in the text, tables, and implementation.

Suggestion for future work: A useful follow-up would measure how well preview-time critic scores predict final scores, and test judge diversification: search with one judge family and evaluate with a different, held-out family. This would quantify possible judge-specific overoptimization when the critic and evaluator use the same model family.

5 comments

We thank the reviewer for their thoughtful review. We are glad the reviewer found the method “clear” and appreciated our evaluation across “multiple generator backbones”. We hope to clarify and address concerns here.

W1: The ablation studies do not isolate the contribution of semantic prompt correction from the other components of the search.

We apologize for the confusion. We would first like to clarify the distinction between the two language ablations in Table 3, as the current description does not sufficiently distinguish them. PreviewDiff uses two forms of language feedback. (1) Positive fix notes, which are appended to the original prompt and (2) negative critique notes, which are incorporated into negative conditioning. In "w/o prompt editing," we remove positive prompt updates while retaining negative critique conditioning. In "w/o language feedback," we remove both forms of conditioning updates, leaving scalar-guided branching and pruning.

This distinction explains why removing positive prompt editing produces only a small reduction on SDXL (2.292 to 2.285), while removing all language feedback leads to a larger reduction (2.292 to 2.063). The former retains semantic feedback through negative conditioning, while the latter reduces PreviewDiff to scalar-guided search without language-based corrections. We will rename the first ablation "w/o positive prompt updates" to make this distinction clearer. We also agree that Figure 5 should not be interpreted as isolating the contribution of semantic correction. Search width increases the number of trajectories explored and retained, and therefore increases the available denoising compute. The purpose of this experiment is to characterize how PreviewDiff scales with search width, rather than to attribute the entire improvement to language feedback. We distinguish this scaling result from our fixed-budget comparisons. Figure 2 compares PreviewDiff against Best-of-N across denoising budgets, while Tables 1 and 2 include scalar-guided search baselines such as EvoSearch and particle sampling. These comparisons provide evidence that PreviewDiff benefits from intermediate interventions beyond simply increasing the number of final samples.

We report an additional final-output critique-and-regenerate comparison in our response to Q1, while distinguishing it from an exactly resource-matched control. We also agree that Table 3 does not establish the statistical significance of the 0.007 difference between the full method and removal of positive updates. We therefore interpret these as differences in reported means, not statistically established effects. Figure 5 separately characterizes search scaling: increasing width also increases computation, so its gain should not be attributed entirely to semantic correction at fixed cost.

W2: The comparison against Video-T1 is not controlled by a explicitly matched budget. The margin is smaller under the LLaVA-Eval column

The Video-T1 rows report our matched-budget evaluations, not scores copied from the original paper. Our earlier wording, “reported verifier-guided video search settings,” was imprecise. Within each backbone, Video-T1 and PreviewDiff use the same nominal aggregate denoising-step budget, counting work across all expanded continuations, including discarded branches and repeated denoising after local restarts. This does not imply equal critic-request counts or wall-clock latency, and we have clarified the matched resource in the manuscript.

The improvement also appears beyond the Gemini evaluator. On LTX-Video, PreviewDiff obtains 0.4523 versus 0.4329 under LLaVA-Eval and 1.987 versus 1.954 under Qwen3-VL. These results support transfer beyond the search critic, without establishing the absence of judge-specific bias. Because the metrics have different scales, their raw margins are not directly comparable. We also clarify that the VBench 2.0 column reports Gemini scoring on VBench 2.0 prompts, not an independent evaluation using the official VBench metric suite.

Q1: It would be appreciated if you can add a final-stage feedback baseline

Thank you for the suggestion. This baseline can be incorporated although this is not perfectly compute matched. We have our multimodal judge provide feedback on the final generated image/video, reconstruct the prompt, and then regenerate the video/image again. For a budget of 86 denoiser steps with SDXL and LTX-Video, this amounts to 4 feedback loops.

DatasetModelMethodDenoising StepsScore
GenEval2SDXLPreviewDiff862.292
GenEval2SDXLFinal Feedback801.766
NarrLVLTX-VideoPreviewDiff862.167
NarrLVLTX-VideoFinal Feedback801.669

As expected, we find that PreviewDiff performs significantly better than incorporating final feedback even when compared to our smallest budgets. This motivates providing judgements during generation by decoding early.

Q2: Do the critic-call budgets include the calls that generate the correction proposals and the VAE decodes of the previews? What is the wall-clock latency of the largest Figure 2 configuration compared with Best-of-N running in parallel?

Correction-proposal requests are included in our critic-call budgets. VAE decodes are separate operations, but their cost is included in end-to-end timing. Across 295 NarrLV prompts on one H100, PreviewDiff at 86 denoising steps takes 9.80 ± 0.99 seconds per prompt, versus 5.66 ± 1.20 for Best-of-3 (mean ± standard deviation). Under our GPU-memory constraints, Best-of-3 uses two sequential generation passes, with up to two concurrent critic requests. These measurements cover the 86-step setting. This is not the largest setting reported in Figure 2. We also do not compare against fully parallel Best-of-N. We therefore do not claim a wall-clock speedup.

Q3: In Table 2, is “MCTS-CGD” a typo for PreviewDiff, or does it denote a different method? Please map every reported row to an exact algorithm and use the same name in the text, tables, and implementation.

“MCTS-CGD” is a labeling error for PreviewDiff, not a separate method. Both rows refer to the PreviewDiff algorithm described in Section 3. We have corrected the labels in Table 2 and standardized the naming in the manuscript.

Thank you for the detailed responses and additional experiments. The distinction between positive prompt updates and negative critique conditioning resolves my question about the Table 3 ablations. The clarification that Video-T1 was evaluated using your own runs under a matched aggregate denoising-step budget addresses W2. The added compute accounting and timing measurements also clarify the inference overhead.

However, I still have two remaining concerns about the new final-feedback comparison:

  1. Appendix L states that the baseline returns the last regenerated output, whereas PreviewDiff selects its best-scoring final candidate. This makes it difficult to attribute the full score gap to feedback timing. Could you also report a best-so-far version of the final-feedback baseline, using the same selection critic and including its selection calls in the budget?

  2. The denoising budget does not add up for me. The response reports 80 steps, but Appendix L describes four refinement rounds after the initial generation, with 20 steps per complete sample. Under the stated formula, this gives 100 steps, while the appendix reports 120. Please clarify which configuration produced the reported scores and give the actual denoising-step and critic-request totals for both methods.

The response has alleviated several of my original concerns, but I will keep my current score pending clarification of this comparison.

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

PreviewDiff intervenes before diffusion sampling finishes. A multimodal critic inspects a decoded preview and proposes prompt corrections; the method then branches over those corrections and local re-noising, scores the continuations, and keeps selected branches. Experiments compare it with Best-of-N and scalar search on image and video generation, including independent evaluators and human judgments.

Critical strengths
  • The critic’s feedback changes the next continuation rather than merely selecting a finished output. Prompt corrections and local restarts give the method a way to revisit a flawed intermediate state while keeping some of its structure. The algorithm explains how those continuations are scored and pruned.
  • The gains are not confined to the evaluator used during search: they also appear under BLIP-VQA and additional multimodal evaluators. Testing both images and videos gives evidence that the approach works beyond one generator or output type. The examples also expose remaining counting and temporal failures.
Critical weaknesses
  • I cannot tell what separates the two feedback ablations. On SDXL, removing prompt editing changes the score from 2.292 to 2.285, while removing language feedback lowers it to 2.063. The method description says language feedback acts through prompt edits. What, then, is still passed to the generator when prompt editing is disabled? This distinction is central: the current results do not show clearly how much semantic correction adds beyond intermediate scoring, branching, and local restarts.
  • The budget-matching claim is also difficult to check. The paper describes matching critic calls or denoiser steps, but the main tables do not specify which criterion each comparison uses or give the actual counts. Equal denoising compute can still involve different amounts of critic computation. The comparison needs both counts, including calls that produce corrections, so readers can see what is held fixed and what additional computation each method uses.
Questions for authors
  • What remains enabled in each feedback ablation, and are their search budgets the same?
  • For each main comparison, can the authors report denoiser steps and critic calls, indicating which budget is matched and how correction generation is counted?
2 comments

We thank the reviewer for their thoughtful review. We are glad the reviewer found the paper recognized the strength of our method with applying “changes to the next continuation” and our demonstration that our method is “not confined to specific evaluators”. We hope to address concerns below.

W1: I cannot tell what separates the two feedback ablations.

Thank you for identifying this ambiguity. PreviewDiff uses two language-conditioning channels: positive fix notes appended to the original prompt and negative critique notes appended to the negative conditioning. “W/o prompt editing” disables only the positive updates, whereas “w/o language feedback” disables both. Intermediate scalar scoring, local restarts, branching, and beam pruning remain enabled. We have renamed the former “w/o positive prompt updates.” Thus, 2.292 versus 2.285 measures the incremental contribution of positive updates while negative feedback remains available, 2.292 versus 2.063 compares the full method with search without either language-conditioning update. The 0.007 difference should not be interpreted as the effect of removing all semantic feedback.

W2: The paper describes matching critic calls or denoiser steps, but the main tables do not specify which criterion each comparison uses or give the actual counts.

We apologize for the incomplete accounting. We distinguish two matching criteria:

  1. Critic-call matching**:** Best-of-N uses N equal to PreviewDiff’s total number of critic requests, including both scoring and correction generation.
  2. Approximate denoiser-step matching**:** N is chosen so that the total denoising steps for Best-of-N approximately match PreviewDiff’s partial rollouts, repeated steps after local re-noising, and final continuations. Matching is approximate because N must be an integer.

These criteria control different resources. We agree that each comparison should identify the matched quantity and report both quantities. Correction generation is included in our critic-call accounting. Each expanded parent receives one correction-generation request, which returns the requested set of correction variants together. We separately count the requests used to score root previews, child previews, and final candidates. Thus, for the two-root, width-two, one-expansion configuration with two correction variants per parent, one restart depth, and one latent branch per correction, the nominal allocation is:

  • 2 root-preview scoring requests
  • 2 correction-generation requests
  • 4 child-preview scoring requests
  • 2 final-candidate scoring requests

for 10 total critic requests per prompt, comprising eight scoring requests and two correction-generation requests.

Method/configurationScoring callsCorrection callsTotal critic callsDenoiser steps
PreviewDiff8210104
Best-of-10: critic-call matched10010400
Best-of-3: approximately denoiser-step matched303120

We will add this accounting to the experimental settings and identify the matching criterion in the relevant table and figure captions. We will also use the specific terms “critic-call matched” and “approximately denoiser-step matched” where appropriate, so that the comparisons do not imply simultaneous equality of both resources.

Reviewer_uDTCOpenReview ↗

Official comment by reviewer uDTC

The ablation is clear now: removing positive prompt updates leaves negative critique conditioning in place. The 0.007 difference therefore does not measure the value of all language feedback.

The best-so-far final-feedback results also make the comparison stronger. PreviewDiff remains ahead on both benchmarks after the baseline is allowed to select its best output. The budgets are close, though: 86 denoising steps and eight critic requests for PreviewDiff, versus 80 and seven for final feedback. I would describe the result with those counts rather than call it exactly budget-matched.

My remaining concern is the accounting for Tables 1–2. The representative example does not identify the actual counts and matching criterion for each comparison. Those should be added, and Appendix L should reflect the corrected configuration and best-so-far results.

I am keeping my rating at 3. The response gives better evidence for intermediate feedback, but the incomplete budget reporting still limits what I can conclude about efficiency.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for their thoughtful comments and for recognizing the value of intermediate semantic feedback, the breadth of our experiments, and our human evaluation. We have clarified the language-feedback ablations, distinguished critic-call matching from denoiser-step matching, explained the Video-T1 comparison, and expanded the discussion of prior work on intermediate diffusion evaluation. We have also added single-GPU end-to-end latency measurements and clarified the scope of our scaling and independent-evaluation claims. All additions and substantive revisions are highlighted in blue in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.