acceptodds
Under review as a conference paper at ICLR 2027

Compositional Diffusion with Hierarchical Latent Cascades

Abstract

Diffusion models are powerful generative tools across images, video, motion planning and beyond. Compositional diffusion extends a pretrained model to horizons longer than those it was trained on by jointly denoising overlapping windows, and a growing body of work keeps the result globally coherent: seamless joints and a consistent layout and style. Yet the composed output often fails the intended global semantics: an object requested once appears in several windows, or in the wrong place. We introduce Hierarchical Latent Cascades (HLC), which routes the global condition down a hierarchy of windows. A forward pass of the pretrained model on a coarse view of the whole output reveals how each part of the condition is distributed over local horizons; we use it to reweight the condition for every child window and pass it down the tree. Each child denoises under its own condition and its parent’s estimate, and overlapping siblings are aggregated level by level into the long-horizon output. HLC is training-free and uses only the model’s own attention, so the same procedure applies to image, video and trajectory models whose condition enters through attention. Experiments on panorama images, long videos and robot motion planning show that it keeps much of the requested global semantics while staying coherent.

Primary area:
generative models
Keywords:
Diffusion models, Long-horizon generation, Compositional Diffusion, Planning

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_qPYGOpenReview ↗
Rating
2
Confidence
2(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies global semantic failures in compositional diffusion. While existing methods can generate coherent long-horizon outputs by jointly denoising overlapping windows, they often repeat objects or events because every window receives the same condition. The paper proposes Hierarchical Latent Cascades (HLC), a training-free method that first generates a coarse global sample, uses its attention maps to assign parts of the condition to different regions, and then refines the result through a hierarchy of overlapping windows. Experiments cover panorama generation, long-video generation, and long-horizon robotic planning.

Critical strengths

1.The paper identifies an important and underexplored problem. The distinction between global coherence and global semantic correctness is clear and practically relevant: existing compositional methods can produce seamless outputs while still violating object counts, positions, or event order. 2.The proposed method is intuitive and broadly applicable. Using the pretrained model's own attention to route conditions to local windows is elegant and does not require additional training, external detectors, or an explicit language-model planner. 3.The cross-domain perspective is valuable. Applying the same general idea to images, videos, and robotic trajectories suggests that semantic repetition is a broader issue in compositional diffusion rather than a domain-specific artifact.

Critical weaknesses

1.The evaluation is too small to support the paper's broad claims: the panorama experiment uses only 25 prompts with one seed each, and the video evaluation contains only two stories and six events, making results sensitive to prompt selection and sampling variance. 2.Absolute semantic performance remains limited. On panoramas, HLC reaches only 0.24 detector-based and 0.48 VLM-based full-prompt success, well below the external planner. It reduces repetition but does not yet reliably preserve the requested global semantics. 3.The evaluation relies heavily on automatic judges, and the sizeable gap between detector- and VLM-based results suggests the metrics are not fully stable. The small human validation does not replace a direct human evaluation of counting, placement, and overall prompt satisfaction. 4.The benefits come with understated trade-offs: HLC worsens panorama style consistency, yields lower video subject consistency than some baselines, and takes roughly 10 minutes per video versus about 3.5 minutes for reported alternatives. Moreover, in planning, root-only results match or exceed the full cascade, leaving the contribution of hierarchical refinement unclear.

Questions for authors

1.Can the authors provide results over a larger prompt set and multiple random seeds, particularly for long videos, where the current conclusions are based on only two stories? 2.How well do the detector and VLM metrics agree with human judgments on the complete evaluation set? Could the authors include a direct human evaluation of count, position, event order, and overall quality? 3.Why is AccDiffusion not included as a quantitative panorama baseline? It appears closely related in its use of attention for local prompt assignment. 4.In the planning experiments, why does the full hierarchy not improve over the root-only solution? Can the authors isolate when hierarchical refinement is actually necessary and provide a clearer ablation of decomposition versus cascading?

2 comments

We sincerely thank the reviewer for the kind and thoughtful review. We are glad that the reviewer finds the distinction between global coherence and global semantic correctness important and underexplored, and that using the model’s own attention without extra training, detectors or a planner is seen as elegant. We share this view, and the revision is built around making the evidence for it stronger.

W1 / Q1. Evaluation size.
We agree, and we have scaled up the evaluation substantially (see the general response). Panoramas now use 100 prompts for every method and 50 prompts × 3 seeds for the controlled comparison. Videos use 50 stories at 40 s, beyond what LTX-2.5 generates in one pass, plus a seed check with three seeds. Every difference is reported with a paired bootstrap 95% confidence interval.

W2. Absolute semantic performance.
Part of the low absolute numbers came from the judges of the submission. The OWLv2 detector counts malformed fragments as objects and misses small or unusual ones, which also explains the gap between the detector and the VLM that the reviewer noted. We therefore re-scored everything with a stronger judge, Qwen3.8-27B, and checked all three judges against blind human labels (general response, item 2). With the new judge, HLC succeeds on 45% of the 100 prompts, against 0% for coherence-oriented samplers, and is not significantly different from the external planner (−4 [−15, 6]). The change of judge does not change the ranking of the methods (Table 6 in Appendix D), and we release every image, video and judge output so that the results can be checked directly. We agree that the problem is not solved: more than half of the prompts still fail, and we say so in the conclusion.

W3 / Q2. Agreement with human judgement.
We labelled 102 objects on 57 panoramas by hand without seeing any judge output and without the method names, using a rule fixed in advance: an instance counts only if a person recognises it as the requested object. Against these labels, agreement is 82% for OWLv2, 80% for Qwen2.5-VL-7B and 87% for Qwen3.8-27B (Cohen’s κ 0.60 / 0.57 / 0.70). We did not run a full human study of every output, which we list as a limitation. For videos we do not score event order: many stories list events separated by commas, which does not state an order, and several events may legitimately overlap (Appendix D).

W4. Trade-offs.
We report the standard coherence and quality metrics of the field, but they do not measure the failure we study. I-Style and NIQE reward windows that look alike, so a sampler that repeats one scene in every window scores well. A panorama that follows the prompt must show different objects in different places and cannot match a replicated one on these measures. Even so, HLC stays close to SyncDiffusion on all three measures and has the lowest I-LPIPS (Table 2). On video, with the stronger LTX-2.5 backbone, HLC now has the highest subject consistency (0.932 vs 0.923 for one-pass; Table 3). The trade-offs are real, and we now state them in the limitations: the style is less uniform than in samplers that repeat one scene, and on video HLC is about twice as slow as window-based samplers because the levels of the tree are sampled one after another (Table 3, Appendix G).

Q3. AccDiffusion.
We added it. We ported the AccDiffusion selector into our tree, so that only the allocation rule differs from HLC, and tuned it. Its best variant reaches 26% success against 45% for HLC (+19 [11, 28]; Table 2 and Table 10 in Appendix F).

Q4. Hierarchy in planning, and decomposition vs. cascading.
In planning, the diffusion planner run once as the root turned out to be surprisingly robust beyond its trained horizon, and the cascade does not improve on it at these lengths. We do not claim HLC is the best planner for robot manipulation. Planning is included to show that the same procedure transfers unchanged to a third domain, where it clearly beats start–goal conditioning and broadcasting. Section 4.4 now states this scope. To separate decomposition from cascading, we ran both ablations on panoramas:

Decomposition: with the same root and tree, removing the decomposition (full prompt in every window) drops success from 45% to 0% (Table 2). The semantics come from the decomposition. Cascading: at 1:14.5, a binary tree and a depth-1 flat variant (n children) with the same root and allocation do not differ significantly in success (30% vs 38%, −8 [−24, 10]). The tree scores higher on learned quality measures (CLIP-IQA and MANIQA), while the flat variant is more uniform in style (Table 9 in Appendix E.2). The hierarchy is therefore needed for fidelity when the output is refined far beyond the root’s resolution, not for semantics.

Thank you again for the encouraging review. We would be glad to answer any further question.

AuthorsOpenReview ↗

Gentle reminder

Dear Reviewer qPYG,

Thank you again for your time and thoughtful feedback. As the discussion period closes on November 18, we would appreciate any feedback on our response above, particularly whether any concerns remain that we could clarify before the deadline.

We would be happy to address any further questions. Thank you for your consideration.

Reviewer_R3NyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Hierarchical Latent Cascades (HLC), a method to enable wider / longer generations with diffusion models. By using HLC they avoid typical error cases like repeating objects and generate more coherent outputs. The algorithm works by first generating a coarse output (root) for the desired scene, using the attention from the root to ground the text embeddings into the real coordinates in the image and compute a weight. To come to a finer resolution the image is split in two (building a tree with branching factor 2.) Each image gets a noised estimate from the root (parent) and denoise it. The two images overlap and the overlap smoothed. This process can then be repeated to reach the desired depth. The algorithm is itself training free, but the authors state that tuning of hyperparameter is required for decent results.

Critical strengths

The paper introduces a training free method to enable diffusion models to generated images far beyond their pretraining size. At the same this method solves issues other previous methods had, like repeating objects (with only one desired one). Additionally, this method is more efficient. The paper is well written and easy to understand and tackles a very interesting topic, as this can also be used for other applications than image generation, such as robot planning.

Critical weaknesses

The results missing the statistical significance required for a convincing contribution. Given all results are just based on less than 50 runs. Especially for the video generation 2 samples are not enough for the claims made. Based on that, I will recommend weak reject.

Questions for authors

Q1: Why is the VBench subject score for one pass so high?
Q2: Why not use a linear interpolation between the two children?
Q3: Where are the prompts for the panoramas and videos from? Are the prompts from SyncDiffusion or were they synthetically generated?
Q4: Given Figure 5, the results from LLM planner seems worse than from HLC. It seems like the scores in Table 1 might not capture that. Why is this the case? Did you consider a user preference study?

3 comments

We sincerely thank the reviewer for the review and for several sharp questions, which helped us improve both the evaluation and the presentation.

W. Statistical significance.
We agree, and we have scaled up the evaluation substantially (see also the general response). Panoramas now use 100 prompts for every method and 50 prompts × 3 seeds for the controlled comparison, and videos use 50 stories at 40 seconds long with a seed check. Every difference is reported with a paired bootstrap 95% confidence interval. For example, HLC improves full-prompt success over SyncDiffusion and CoFi by +45 [35, 55] points each, and over the strongest AccDiffusion variant by +19 [11, 28] (Table 2). On videos, HLC reaches 74% against 64% for one-pass generation (Table 3).

Q1. Why is the VBench subject score for one pass so high?
Subject consistency measures how stable the appearance of the subject is over time. It does not check whether the requested events happen, so a clip that barely moves scores near the maximum. One-pass generation with Wan 2.2 produced such nearly static clips. Because we had to scale up the experiments for the revision, we moved all video experiments to LTX-2.5, a stronger backbone, where HLC has the highest subject consistency (0.932 vs 0.923 for one-pass) as well as higher event success (Table 3).

Q2. Why not use a linear interpolation between the two children?
We already do. On the overlap of two sibling windows, the merged estimate is (1−u)A+uB(1-u)A + uB with uu rising linearly across the overlap (Eq. 5 with a linear-ramp weight). It is applied to the clean estimates at every denoising step and written back into both windows, rather than once on the final pixels (Appendix B). We clarified this in Section 3.2.

Q3. Where do the prompts come from?
None are taken from SyncDiffusion. The 100 panorama prompts consist of the 25 prompts of the submission, written by us; 25 prompts selected by a fixed rule from a set we pre-registered before generating any image of it (some adapted from T2I-CompBench); and 50 prompts written for this revision with the help of a language model and frozen before generation. The 50 video stories were written in the same way and frozen before generation. All prompts are listed in Appendix C and released.

Q4. Figure 5 vs. Table 1, and a user study.
The two measure different things. Global semantics only asks whether each requested object appears the right number of times at the right place. The planner places objects well because an external language model writes each window’s prompt from the stated positions. But each window then receives a different prompt and is generated flat with overlap averaging, so neighbouring windows do not agree on one scene. This is the collage effect in Figure 5, and it results in a poor coherence score (I-Style, lower is better: 78.6 for the flat planner vs 31.5 for HLC on 100 prompts).

In short, the LLM planner gets a high semantic score thanks to its external planner module, but a poor coherence score because of the collage effect visible in Figure 5.

Following another reviewer, we now run the planner’s prompts inside HLC’s root and tree. Its semantic success stays about the same (49% vs 48% flat) and its coherence becomes as good as HLC (I-Style 29.7). This shows that our cascade is an effective way to aggregate windows. With this privileged information the planner remains the strongest, as expected, but its advantage over HLC is small and not significant (−4 [−15, 6]), and HLC is the best method that uses no external module. We did not run a user preference study. Instead, we validated the judge against blind human labels (general response, item 2).

We would be glad to answer any further question.

I thank the authors for their detailed answer. All of my concerns are addressed and I will raise my score to weak accept.

Thank you for reviewing our response. we appreciate your constructive feedback

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

This paper proposes Hierarchical Latent Cascades (HLC), a training-free method that addresses repetition when a global prompt is broadcast to every local window of a long output. It samples a coarse view of the whole output, uses attention to assign phrases or instructions to windows, and carries these assignments through a refinement tree. Parent estimates guide the layout and coherence across windows. In the panorama implementation, routing operates on comma-separated phrases, with scene and style retained globally. Experiments cover SD3 panorama generation, long video generation and instruction-list planning in MuJoCo.

Critical strengths
  • The evaluation directly addresses the stated failure mode. Count and position metrics distinguish semantic correctness from visual coherence, and the detector and VLM judgments are checked against hand labels.

  • The panorama ablation shows a clear routing benefit: with the same root, tree and seed, object correctness rises from 0 to 0.63 with the detector and 0.81 with the VLM judge (Table 1), and Table 5 shows how this benefit changes with horizon length.

Critical weaknesses
  • The image experiments show that routing helps within HLC, but its added value over existing attention-derived prompting is less clear. AccDiffusion [1] already selects and re-encodes local prompts from coarse attention, and the panorama experiments use the same kind of operation, re-encoding a shortened prompt for each window (Appendix B). HLC adds phrase grouping, a different scoring rule, a root that compresses the whole output into one window, and inherited assignments over a tree that extends the horizon. The same-root ablation compares the combined method with broadcasting the full prompt, so it does not separate the proposed allocation rule from the earlier approach. Appendix E’s word-level comparison (count error 2.33→0.722.33 \rightarrow 0.72) could help, but neither the selector nor the evaluation is described in enough detail to interpret the gain. So my main reservation is how much these results show beyond the known benefit of local prompting.

  • The comparison with the planner leaves an important trade-off unresolved. The planner has better semantic scores, while HLC is more coherent (Table 1). Most of the semantic gap is in placement: with the detector, almost a fifth of HLC’s objects have the right count but not the stated position (count 0.81 vs. objects 0.63), which does not happen for the planner (0.84 for both). However, Appendix B describes planner prompts sampled with overlap averaging, without HLC’s root-conditioned cascade, so its lower coherence could come from the generation structure as well as from the prompts. As it stands, the comparison shows a difference between the two complete pipelines, but not that attention-based allocation gives a better balance of semantics and coherence than other allocation choices within HLC’s hierarchy.

  • As a secondary point, the other domains provide useful demonstrations, but offer limited further evidence of the hierarchy’s practical benefit: the video evaluation covers two single-seed stories, while root-only planning matches or exceeds HLC’s all-instruction success in all four settings at much lower cost (Section 4.4; Appendix F).

[1] Lin et al., “AccDiffusion: An Accurate Method for Higher-Resolution Image Generation”, ECCV 2024.

Questions for authors
  • Could you describe the prompts, number of windows, count-error definition and selector used in Appendix E? How does that selector relate to AccDiffusion’s mask-coverage rule? This would clarify what the reported gain says about HLC’s allocation rule.

  • On the Table 1 prompts, could you test either the existing planner’s prompts or routing from the stated positions within HLC’s same root and hierarchy? This would clarify whether the coherence advantage depends on attention-based allocation or mainly on the shared coarse initialization and refinement.

4 comments

We sincerely thank the reviewer for a careful and insightful review, and we are glad the reviewer found that our evaluation targets the stated failure mode and that the same-root ablation shows a clear routing benefit. Your questions on what the allocation rule adds led to the most informative new experiments of the revision.

W1. Added value over AccDiffusion.
We agree the allocation itself should be compared. AccDiffusion is a patch-wise upscaler and cannot produce long panoramas on its own, so we had to port its selector from the official code into our tree. It now shares our root, tree, noise, attention maps and text path, and only the allocation rule differs. We also tuned it over several design choices and report the best-performing AccDiffusion variant: word and phrase units, lower coverage thresholds, and a fallback we added for objects it places in no window.

Best AccDiffusion variant: 26% success vs 45% for HLC on 100 prompts (+19 [11, 28]; Table 2).

W2. Planner vs. HLC inside the same hierarchy.
We ran the planner’s prompts inside HLC’s root and tree (Table 2, last row; Table 10, row F).

Semantics are unchanged (49% vs 48% flat), and coherence becomes as good as HLC (I-Style 29.7 vs 78.6 flat; HLC 31.5). So the flat planner’s collages came from flat generation not from its prompts, and our cascade is what keeps the windows consistent. With privileged information (it reads the stated positions), the llm planner remains the strongest as expected but its edge over HLC is small and not significant (−4 [−15, 6]). Among methods without an external module, HLC is best.

W3. Other domains.
In planning, the diffusion planner run once as the root turned out to be surprisingly robust beyond its trained horizon, and the cascade does not improve on it at these lengths. We do not claim HLC is the best planner in robot manipulation. Planning is included to show that the same procedure transfers unchanged to a third domain, where it clearly beats start–goal conditioning and broadcasting. Section 4.4 now states this scope. For video, the number of stories in the submission was too small to show the advantage of HLC. We now evaluate 50 stories at 40 seconds long, beyond what LTX-2.5 generates in one pass.: HLC 74% vs one-pass 64% and window-based samplers 48% (FreeNoise) and 30% (Gen-L-Video) (Table 3).

Q1. The Appendix E number.
Our description was too brief. It came from an early pilot of a method we later dropped: instead of attention, we deleted one word at a time and checked which window of the coarse view the model denoised worse, and treated that window as the word’s place. The pilot used 3 scenes × 2 seeds at 1:16 (8 windows), a root picked with the stated positions and an OWL-ViT detector. It is unrelated to AccDiffusion’s rule and does not test the paper’s method, so we removed the number. Word vs. phrase units now have their own controlled study (Table 10): for HLC’s rule they perform the same (39% vs 45%, +7 [−1, 14]).

Q2. Planner prompts and stated positions within HLC’s hierarchy.
We ran both on 50 prompts × 3 seeds (Table 10, App. F):

Planner prompts in our tree (F): 49%, not significantly different from HLC (−4 [−15, 6]). Stated positions (G): 26%, 19 [7, 33] points below HLC. A stated position helps only where it matches what the root drew, and objects with no stated position (counts, relations, attributes) cannot be placed this way.

Coherence comes from the shared root and cascade, while semantic gains come from the allocation.

We would be glad to answer any further question.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer JHYX,

Thank you again for your time and thoughtful feedback. As the discussion period closes on November 18, we would appreciate any feedback on our response above, particularly whether any concerns remain that we could clarify before the deadline.

We would be happy to address any further questions. Thank you for your consideration.

Thank you for the clarifications and the additional experiments. The new comparisons with AccDiffusion and the LLM planner resolve my main concerns about the original experiments. I also appreciate the correction to the Appendix E result. The new word/phrase comparison is convincing and answers my question.

However, I have one remaining concern about describing the planner as using "privileged information." As I understand it, HLC receives the same original prompt, including the stated positions. The difference seems to be whether an external LLM interprets the prompt and assigns its contents to windows. Could the authors clarify what information they consider privileged in this setting?

I agree that not requiring an external LLM is an advantage of HLC. However, the planner achieves slightly higher semantic scores with comparable coherence, so the practical trade-off depends on the planner's overhead. If the planner adds little overhead, HLC's practical advantage over it would be less clear to me. I know there is not much time left for discussion, but could the authors report the additional time and cost of the LLM planner, including its effect on end-to-end generation time?

Thank you for the constructive review. We are glad the new experiments resolved the main concerns.

About privileged information.
The planner cannot see the image or the model, so it has to be told the setting in words: our instruction states that the panorama is drawn as four side-by-side windows, each from its own prompt, and which window each position word refers to. Without this it could not write per-window prompts at all.

Even so, we fully understand your concern about our use of the word privileged. It is a reasonable doubt, and other readers of the paper may well have the same question. We have therefore renamed these rows External LLM planner throughout the paper.

Overhead.
We used Claude Opus 5.5 as the planner in our main experiment. The planner can also be run locally with an open-source model such as Qwen3.8-27B, which we found strong at visual reasoning when using it as the judge, so we measured its overhead as well.

MethodDecomposition (s)Generation: root, windows, decoding (s)Total per panorama (s)Extra cost per prompt
HLC (ours)0.5 (attention probe)22.923.4–
LLM planner in HLC's root and tree, Claude Opus 5.5 (API)5.423.128.5$0.014
LLM planner in HLC's root and tree, Qwen3.8-27B (local, one A100)9.723.132.8–

Mean seconds on one A100. Decomposition is the step that decides which content goes to which window: HLC's attention probe, or one planner call. Generation times are means over the 100 prompts. Loading the Qwen model (330 s, once) is not included.

So the overhead is not negligible, but it is not critical either, and we do not claim an overhead advantage as the main point.

Our claim is the following.

MethodSuccess (%) ↑Objects (%) ↑Malformed (%) ↓I-LPIPS ↓I-Style ↓NIQE ↓
No external information
SyncDiffusion (NeurIPS’23)00.039.50.48435.84.15
CoFi (NeurIPS’26)00.022.60.50517.83.77
AccDiffusion selector‡2648.014.70.47630.64.29
HLC w/o decomposition (same root)00.638.40.48326.84.01
HLC (ours)4565.013.00.47631.54.29
External LLM planner
LLM planner, flat windows4865.515.80.58478.64.33
LLM planner in HLC's root and tree4967.211.90.47729.74.24
  • The planner alone does not give a usable panorama. With flat windows and overlap blending, its windows do not agree on one scene and the result is a collage (I-Style 78.6 vs 31.5 for HLC). The comparable coherence you mention appears only when its prompts are run inside HLC's root and tree (29.7).
  • The best row is therefore the planner combined with our cascade. Even when an external model is available, our aggregation is what should be used with it.
  • On the decomposition itself, HLC needs no external module and no training, and it is not significantly behind the external planner (−4 [−15, 6]).
  • Among methods that use no external planner, HLC is the best: 45% success against 0% for SyncDiffusion and CoFi and 26% for the strongest AccDiffusion variant (+19 [11, 28]).

Thank you again for your careful reading and valuable feedback throughout the discussion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We sincerely thank the reviewers for their careful and constructive reviews. We have revised the paper along these lines and summarise the changes below.

1. A larger evaluation with statistics.
We agree that the original evaluation was too small to support a firm conclusion. We have scaled it up.

Panoramas: 100 prompts for every method, including SyncDiffusion and CoFi, and 50 prompts × 3 seeds for the controlled comparison. Every difference is reported with a paired bootstrap 95% confidence interval. Videos: 50 stories at 40 s, twice the length LTX-2.5 generates in one pass, against one-pass generation, Gen-L-Video and FreeNoise (single and multi-prompt), plus a seed check with three seeds.

Revised Table 2 (100 prompts, Qwen3.8 judge):

MethodSuccess (%)Objects (%)Malformed (%)
SyncDiffusion00.039.5
CoFi00.022.6
AccDiffusion selector in our tree2648.014.7
HLC w/o decomposition (same root)00.638.4
HLC (ours)4565.013.0
LLM planner, flat windows (privileged reference)4865.515.8
LLM planner in our tree (privileged reference)4967.211.9

Among methods that use no external information, HLC is significantly best (e.g. +19 [11, 28] points over the strongest AccDiffusion variant). It is not significantly different from the external planner (−3 [−17, 11] flat, −4 [−15, 6] in our tree). On videos, HLC reaches 74% success against 64% for one-pass generation and 48% and 30% for the window-based samplers FreeNoise and Gen-L-Video (Table 3).

2. A stronger judge, checked against human labels.
The submission used an OWLv2 detector and Qwen2.5-VL-7B. We chose a detector because it is deterministic and easy to reproduce. As the reviewers noted, the two judges disagreed substantially. Checking them against our hand labels showed why: the detector counts malformed fragments as objects and misses small or unusual ones. We therefore score global semantics with Qwen3.8-27B, which inspects the output window by window at native resolution and re-checks every instance on a zoomed crop. On 102 hand-labelled objects (57 panoramas), agreement with the human labels is 82% for OWLv2, 80% for Qwen2.5-VL-7B and 87% for Qwen3.8-27B (Cohen’s κ 0.60 / 0.57 / 0.70). Re-scoring with the original judges orders the methods the same way (Table 6 in Appendix D). All judge prompts are given verbatim in Appendix D, and we release every generated panorama and video together with the judge outputs in the supplementary material (downscaled because of the size limit).

3. Improved writing.
We rewrote the main text for readability without changing its content: experimental settings are collected in one table (Table 1 in Section 4.1), long sentences are split and setting details are moved to the appendix. While re-checking every reported number against the original logs we also corrected a few descriptions, listed in the individual responses.

We hope these revisions address the reviewers’ concerns. We would be glad to run further analyses if any question remains.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.