acceptodds
Under review as a conference paper at ICLR 2027

HLA-WM: Hybrid Linear Attention for Long-Horizon Video World Models

Abstract

Long-horizon video world models require persistent memory to preserve scene consistency over extended rollouts. Softmax attention retains the full generation history through a growing KV cache, whereas recurrent linear attention compresses history into fixed-size states with substantially lower memory cost. However, we identify severe long-range forgetting in Gated DeltaNet (GDN), where information from distant but relevant scenes is progressively attenuated by subsequent state updates. To address this limitation, we propose HLA-WM, a training-free hybrid linear-attention framework that combines coarse-grained geometry-guided retrieval with fine-grained recurrent linear-state computation. HLA-WM exploits the affine structure of GDN to cache compact chunk-wise transition summaries, retrieve scene-relevant historical chunks using camera geometry, and recompose them into query-specific recurrent states. On the -second SANA-WM-Bench, HLA-WM improves all six aggregate revisit-consistency and camera-control metrics of the base autoregressive generator without additional training, including a dB PSNR gain and a reduction in rotation error. The improvements persist after downstream refinement and generalize to MBench-A, where HLA-WM consistently improves all three revisit-consistency metrics across all four subsets and all evaluated inference modes over samples. At a -second context, HLA-WM reduces historical-state memory by relative to full KV caching while incurring at most a reduction in inference throughput. These results demonstrate that selectively addressable recurrent memory can improve long-range scene recall while preserving the efficiency advantages of GDN.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Video World Models, Hybrid Linear Attention, Long-Range Memory, Gated DeltaNet, Efficient Inference

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_1QmeOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates a limitation of camera-conditioned video world models based on Gated DeltaNet (GDN). When the camera returns to a previously visited region, the model often fails to maintain consistent scene geometry. The authors attribute this problem to long-range forgetting. As new video chunks are processed, repeated state updates gradually weaken information from earlier chunks, even when the new chunks show unrelated regions. To address this issue, the authors propose HLA-WM, a training-free method that caches an affine summary of each chunk's GDN state transition. Using camera poses, the method identifies relevant historical chunks based on camera frustum overlap and combines their cached summaries in chronological order to reconstruct the recurrent state. Experiments on SANA-WM-Bench and MBench-A show improvements in revisit consistency and camera control.

Critical strengths
  1. The paper clearly identifies and analyzes a limitation of recurrent linear attention. The chunk-level formulation of GDN and the retention analysis provide an intuitive explanation of why relevant information from earlier chunks is gradually lost.

  2. The proposed method makes effective use of camera poses that are already available as model inputs. It requires no additional training and introduces minimal computational overhead.

  3. The experimental results consistently show improved revisit consistency across two benchmarks, two trajectory splits, and three inference modes.

Critical weaknesses
  1. The method relies on camera poses and a scene-level median depth provided in the trajectory metadata. While these inputs are available in the evaluated setting, it is unclear how robust the method would be to inaccurate pose or depth estimates, or how it could be extended to action-conditioned models without explicit camera poses.

  2. The comparison with existing memory methods is limited. Although HLA-WM uses 12 times less memory than full KV caching, the paper does not compare their revisit consistency. It also lacks comparisons with retrieval-based memory methods such as WorldMem and Context-as-Memory, which use similar geometric criteria to select historical information. Without these comparisons, it is difficult to assess the trade-off between memory efficiency and revisit consistency relative to alternative approaches.

Questions for authors
  1. How robust is the retrieval to errors in camera poses or scene-level depth estimates? Can the method work in interactive settings where camera poses must be estimated rather than provided?

  2. How does HLA-WM compare with full-history KV access or geometry-retrieved KV memory under comparable conditions? The KV-only ablation in Table 4 does not provide such a comparison. This would help clarify the trade-off between memory efficiency and revisit consistency.

  3. How much does geometric ranking improve retrieval over alternative selection strategies under the same memory budget? Table 7 shows only a small gain over uniform selection (9.91 vs. 10.00 PSNR), while the retrieval budgets are not strictly matched.

2 comments
AuthorsOpenReview ↗

Response to reviewer 1Qme(1/2)

Q1. How robust is retrieval to inaccurate camera poses or depth estimates, and can it support action-conditioned models without explicit poses?

A1. HLA-WM demonstrates robust historical retrieval under inaccurate camera poses and depth estimates. We conduct additional experiments on the complete Hard-Trajectory split, perturbing only the geometric metadata used for retrieval while keeping the generation-conditioning camera trajectories unchanged. We evaluate translation noise, rotation noise, and depth-scale perturbations.

Retrieval metadataPerturbationPSNR ↑SSIM ↑LPIPS ↓RotErr ↓
HLA-WMNone10.110.19970.584815.0887
Translation noise1% depth10.100.19900.583514.6183
Translation noise5% depth10.060.19840.586414.1660
Rotation noise1°10.110.19830.583914.9830
Rotation noise3°10.060.19710.583015.8757
Rotation noise5°10.080.19840.585215.9813
Depth-scale perturbation×0.5010.060.19660.586414.3419
Depth-scale perturbation×0.7510.110.19890.584414.3960
Depth-scale perturbation×1.2510.080.19960.584315.8286
Depth-scale perturbation×1.5010.030.19820.585715.6138

HLA-WM maintains stable revisit consistency across all tested perturbations, with PSNR ranging from 10.03 to 10.11 dB compared with 10.11 dB without perturbation. These results demonstrate that geometry-guided retrieval remains effective under the evaluated pose and depth inaccuracies. For action-conditioned world models, motion-related actions can similarly provide spatial cues for retrieving relevant historical information. We have included the robustness analysis in Appendix A.9 of the revised manuscript.

Q2. Comparisons with full-history KV access, geometry-retrieved KV memory, WorldMem, and Context-as-Memory are missing.

A2. HLA-WM addresses long-range forgetting in fixed-size GDN states through selective historical-state recomposition, a limitation not addressed by existing memory retrieval methods. We have clarified this distinction in Sec. 2 and added comparisons in Appendix A.10.

  1. Additional comparisons demonstrate the effectiveness and memory efficiency of GDN-state recomposition. We evaluate alternative memory mechanisms on the same SANA-WM backbone. The KV-based baselines retain historical softmax KV while keeping the native GDN recurrent states unchanged.

    MethodPSNR ↑SSIM ↑LPIPS ↓Peak Memory (GB) ↓
    Native SANA-WM9.370.17740.604432.40
    Full-history softmax KV9.910.19360.601740.64
    Geometry-retrieved KV-only9.380.17700.602140.64
    WorldMem9.800.18610.599940.64
    HLA-WM (GDN-only recomposition)9.780.18900.592933.72
    HLA-WM (Ours)10.110.19970.584841.52

    GDN-only recomposition improves all three revisit metrics with minimal additional memory. Full HLA-WM further achieves the best revisit consistency with only 0.88 GB more peak memory than the KV-based alternatives, demonstrating a favorable memory–consistency trade-off.

  2. Existing retrieval methods do not address recurrent GDN-state forgetting. WorldMem retrieves historical visual memory rather than recomposing GDN states. Our adapted KV-retrieval baseline achieves 9.80 dB PSNR compared with 10.11 dB for HLA-WM. Context-as-Memory's official implementation has not been publicly released, preventing a reliable quantitative comparison. We therefore discuss it qualitatively.

  3. HLA-WM preserves substantial memory efficiency while achieving the best revisit consistency. At 60 seconds, GDN summaries require only 2.15 GB compared with 25.82 GB for full-history KV in the corresponding layers. Including softmax KV, HLA-WM requires 10.76 GB of historical cache, 3.20× less than the all-layer full-history KV reference (34.43 GB). It also maintains comparable throughput to native SANA-WM (22.053 vs. 22.403 FPS). Complete memory accounting is provided in Appendix A.11.

AuthorsOpenReview ↗

Response to reviewer 1Qme(2/2)

Q3. How much does geometric ranking help under the same historical-memory budget?

A3. Geometry-guided retrieval provides meaningful benefits over uniform selection, although its advantage depends on the evaluation horizon and scene characteristics.

  1. Matched-budget experiments demonstrate the effectiveness of geometric ranking. As shown in the result table in our response to Reviewer cE3k (Q3), under the default K=1 setting, Geometry-1 improves PSNR by 0.45 dB over Uniform-1 (10.11 vs. 9.66) and 0.53 dB over Random-1, outperforming both across all six metrics. These results demonstrate the benefits of geometric ranking beyond simply retaining additional history. We have also incorporated this comparison into the Experiments section of the revised manuscript.

  2. The limited evaluation horizon and scene complexity can reduce the advantage of geometric retrieval. The 60-second rollouts contain only 40 chunks, allowing uniform selection with six chunks to cover much of the history. Occlusions and restricted viewpoints in indoor scenes, together with dynamic objects and repetitive structures in outdoor city scenes, further reduce retrieval effectiveness. We discuss these limitations in the conclusion.

Reviewer_cE3kOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses long-range forgetting in GDN-based video world models such as SANA-WM: content from a scene seen long ago is washed out by later state updates, so the scene changes when the camera comes back. HLA-WM is training-free. It caches each chunk's affine effect on the GDN state as a summary (A_i, B_i). For every new chunk it picks a few past chunks: the first chunk, the most recent ones, and the top-K by camera-frustum overlap computed from the given trajectory. It composes their summaries in time order into a query-specific GDN state and applies the same selection to the KV cache of the softmax layers. On 60 s SANA-WM-Bench rollouts, Stage-1 revisit PSNR improves by about 0.7 dB and rotation error drops by 26–31%, with under 2% throughput loss. The revisit gains also hold on MBench-A and after both refinement modes, with some camera-metric trade-offs under AR refinement.

Critical strengths
  1. The core idea is simple and exact. Because the GDN update is affine in the state, a chunk can be summarized once by (A_i, B_i), and any time-ordered subset of chunks can be composed without running those chunks through the model again. Selecting all chunks gives back the original recurrence. It needs no training and costs little in throughput: Stage-1 throughput goes from 22.40 to 22.05 FPS. I see this as a useful middle ground between a single recurrent state and a full KV cache. The GDN recomposition itself should apply to other GDN-based world models, although hybrid ones may also need the KV alignment (A.4).

  2. The revisit gains are consistent, with inputs, trajectories and prompts held fixed. At Stage 1 all six metrics improve on both splits (PSNR 9.18 to 9.91 and 9.37 to 10.11). On MBench-A, PSNR, SSIM and LPIPS improve for every subset in all three modes (Table 11). I also find it convincing that the gain grows with the revisit interval, from about 0.4 dB for short gaps to about 1.0 dB for long ones at Stage 1 on the Hard split (Fig. 6). The w/o Select A ablation keeps all intermediate transitions and drops PSNR from 10.11 to 9.61 (Table 3), which fits the claim that unrelated transitions wash out distant content.

Critical weaknesses
  1. My main concern is that the paper's account of its gains and memory cost does not fully match its evidence. The abstract reports a 28.5% reduction in rotation error, and the main text explains the Stage-1 camera gains by better-preserved geometry (pp. 2 and 8). Table 4 shows that they come mostly from the KV path, which A.4 treats as a complementary stabilizer and the main text describes in a single sentence (p. 7). On the Hard split, Stage-1 RotErr is 20.38 for the baseline, 20.08 with GDN recomposition only, 13.40 with KV selection only and 15.09 for the full method. For memory, past chunks are never re-evaluated (A.2.1), so KV selection requires keeping their KV, yet Fig. 4 does not say whether this KV is counted in the 2.15 GiB. The reported memory also grows with rollout length (2.15 GiB at 60 s against 0.359 GiB for native GDN), so "preserving the efficiency advantages of GDN" should be qualified.

  2. At the default setting, the evaluation cannot tell geometric retrieval apart from simpler choices. With K=1 the model keeps the sink, the most recent chunk and one chunk chosen by camera overlap, but there is no geometry-free control at or below this budget, such as sink and recent chunks only (K=0), or the top-1 chunk replaced by a random or uniformly spaced one. Geometry-guided selection is compared with other strategies only at K=4, with budgets that are not strictly equal (A.6), and there it beats uniform selection by 0.09 dB PSNR (10.00 vs. 9.91, Table 7). The gains are also uneven across scene types: on the Hard split, the Stage-1 PSNR gain is 1.57 dB for Game Style and 0.39 to 0.61 dB for the other three categories (Table 8), so results in other domains may differ.

Questions for authors

Q1. In native SANA-WM, do the softmax layers attend to the full history or only to a sink and a local window? If a window, how is the KV of older chunks kept for KV alignment? In either case, is that KV counted in the 2.15 GiB, and in the 0.359 GiB GDN line, of Fig. 4?

Q2. At the default K=1, how does HLA-WM compare with sink and recent chunks only (K=0), and with the same budget when the top-1 chunk is chosen at random or uniformly?

4 comments
AuthorsOpenReview ↗

Response to reviewer cE3k(1/2)

Q1. The camera-control gains appear to come mainly from KV selection rather than GDN-state recomposition.

A1. Both GDN-state recomposition and KV-cache selection contribute to HLA-WM, but their effects differ across evaluation metrics. Importantly, KV selection alone does not resolve the long-range forgetting problem in recurrent GDN states.

  1. KV selection contributes substantially to camera-control improvements, but does not fully explain HLA-WM's gains. As the reviewer correctly observes, on Hard Stage 1, RotErr is 20.38° for the baseline, 20.08° for GDN-only, 13.40° for KV-only, and 15.09° for the full method. Thus, KV-only achieves a larger RotErr reduction in this setting. The 28.5% reduction reported in the abstract refers to the average relative improvement of the full method across the Simple and Hard splits, rather than GDN recomposition alone.

  2. GDN-state recomposition is essential for robust long-range memory recovery beyond KV selection alone. On Hard Stage 1, GDN-only improves revisit PSNR from 9.37 to 9.78 dB, compared with 9.52 dB for KV-only, while the full method reaches 10.11 dB. Moreover, under Hard AR refinement, KV-only reduces PSNR from 14.12 to 13.54 dB, whereas GDN-only improves it to 14.35 dB and the full method further reaches 14.47 dB. Under bidirectional refinement, the full method also achieves better camera control than KV-only (10.53° vs. 11.47° RotErr). These results demonstrate that KV selection alone provides limited and sometimes inconsistent benefits, whereas combining it with GDN-state recomposition yields stronger overall performance. This supports our central finding that selective historical-state recomposition effectively alleviates long-range forgetting in GDN.

Q2. Does HLA-WM retain historical softmax KV, and is this storage included in Fig. 4?

A2. HLA-WM explicitly retains historical softmax KV for selective retrieval. We clarify this implementation and provide complete memory accounting in Appendix A.11.

  1. HLA-WM retains historical softmax KV without re-evaluating past chunks. Native SANA-WM uses a sink and local sliding window for softmax attention. HLA-WM additionally archives the KV entries of completed chunks, allowing relevant historical KV to be retrieved beyond the native attention window. The retrieved KV entries are aligned with the selected GDN states without re-evaluating past chunks, as clarified in section 4.4.

  2. HLA-WM achieves the best revisit consistency while retaining substantial memory efficiency. Fig. 4 originally reports only GDN-related memory, excluding softmax KV storage. We therefore provide complete historical-cache and peak GPU memory measurements at a 60-second rollout, together with revisit-consistency results on Hard Stage 1.

    MethodPSNR ↑SSIM ↑LPIPS ↓Layers 0–2, 4–6, … (GB)Layers 3, 7, 11, … (GB)Total (GB) ↓Peak (GB) ↓
    Native SANA-WM9.370.17740.60440.360.480.8432.40
    HLA-WM (GDN-only)9.780.18900.59292.150.482.6333.72
    Full-history softmax KV9.910.19360.60170.368.618.9740.64
    HLA-WM (Full)10.110.19970.58482.158.6110.7641.52
    All-layer full-history KV (reference)———25.828.6134.4367.00

    GDN-only recomposition already improves all three revisit metrics with minimal memory overhead. Full HLA-WM further achieves the best revisit consistency with only 0.88 GB additional peak memory over full-history softmax KV. Its GDN summaries require only 2.15 GB compared with 25.82 GB for full-history KV in the corresponding layers. Including both memory pathways, HLA-WM requires 10.76 GB of historical cache, approximately 3.20× less than the all-layer KV reference (34.43 GB). These results demonstrate a favorable trade-off between long-range memory recovery and storage efficiency.

AuthorsOpenReview ↗

Response to reviewer cE3k(2/2)

Q3. The default K=1 setting lacks geometry-free controls with matched retrieval budgets.

A3. Geometry-guided retrieval provides clear benefits over geometry-free selection under the same retrieval budget. We conducted additional experiments on the Hard-Trajectory split using the default K=1 setting, keeping the sink memory, recent context, turn-aware rules, and KV alignment identical across selection policies. Only the selection of the additional historical chunk differs. Random-1 is evaluated over three random seeds, while Uniform-1 selects the temporal midpoint of eligible history. K=0 serves as a control without additional historical retrieval.

Selection policyKPSNR ↑SSIM ↑LPIPS ↓RotErr ↓TransErr ↓CamMC ↓
Sink + recent only09.62000.19450.595217.67311.97732.0889
Uniform-119.66000.19420.593318.70431.93271.9743
Random-1 (3 seeds)19.5767 ± 0.02080.1887 ± 0.00100.5941 ± 0.000217.3928 ± 0.23411.9514 ± 0.03501.9884 ± 0.0144
Ours110.11000.19970.584815.08871.83881.9482

Geometry-1 achieves the best results across all six metrics, improving PSNR by 0.45 dB over Uniform-1 and 0.53 dB over Random-1. Notably, Random-1 performs worse than K=0 in PSNR, indicating that retrieving additional history does not necessarily improve performance. These results demonstrate that the relevance of retrieved historical information, rather than retrieval budget alone, is critical to effective long-range memory recovery. We have incorporated this matched-budget comparison into the Experiments section of the revised manuscript.

Q4. Improvements vary across scene categories, raising concerns about generalization.

A4. HLA-WM consistently improves revisit consistency across diverse scene categories, with gains varying according to scene characteristics. On Hard Stage 1, PSNR improves across all four categories: Game Style (+1.57 dB), Outdoor Nature (+0.61 dB), Indoor (+0.41 dB), and Outdoor City (+0.39 dB). Game Style scenes typically contain stable layouts and distinctive structures that facilitate geometric retrieval. In contrast, vegetation in Outdoor Nature, walls and furniture in Indoor scenes, and building occlusions and dynamic objects in Outdoor City make historical observations less reliable across viewpoints, offering a concrete explanation for the smaller gains. As acknowledged in our conclusion, occlusion is a limitation of geometry-based retrieval. Nevertheless, the improvements across all four categories and all MBench-A subsets and inference modes demonstrate the broad effectiveness of HLA-WM in alleviating GDN's long-range memory limitations.

Thank you to the author for providing the additional experimental details. I believe all my questions have been addressed, I maintain my score of 3.

AuthorsOpenReview ↗

Thanks for the feedback

Thank you for your follow-up and for acknowledging our responses. We appreciate your constructive feedback and careful consideration. We have incorporated the corresponding clarifications and additional experimental results into the revised PDF.

Reviewer_ac16OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

World models depend on long-range attention modules, and to manage this computational complexity a recurrent linear-attention architecture named Gated DeltaNet (GDN) maintains a state that evolves over time. The authors identify that GDN suffers from a long-range forgetting problem as the state is updated according to temporal progression rather than spatial and geometric proximity. To remedy this problem, the authors propose computing the current state using only the relevant historical neighbors on the fly by updating a base state with only the top-k most geometrically relevant neighbors, as determined by camera position and camera view. This can be done as each historical chunk’s relevance is computed using a low-cost symmetric frustum-overlap score (Eq 12) and the state can be updated efficiently using their composition rule (Eq 9). The authors demonstrate their method improves upon existing works on standard benchmarks (Tables 1 and 2).

Critical strengths
  1. The authors have clearly articulated the limitation of existing works (GDN) and how their method addresses this limitation. The authors identify that GDN updates its current state using temporal progression rather than spatial and geometric proximity. The proposed solution directly addresses this problem by explicitly updating the network’s state using spatial and geometric information.

  2. The proposed method appears to be novel. My literature search related to GDN did not find any paper with the same ideas. Additionally, the proposed method addresses a fundamental limitation of the GDN module within World Models, which was proposed in May 2026 (according to arXiv), so it seems likely this limitation has not yet been addressed in related works.

Critical weaknesses
  1. The significance of the finding is questionable due to the low values in the quantitative evaluation. While the qualitative improvement is impressive, the quantitative improvement seems to be relatively small. The SANA-WM-Bench appears to be very difficult, as both methods report very low fidelity metrics (ie a PSNR of about 10 is generally considered very bad). In this low-PSNR/low-SSIM regime, the reported numerical differences do not usually convey much meaning. Perhaps this is related to the difficulty of the problem, but unfortunately, this still makes the current quantitative evaluation less significant.
Questions for authors
  1. Is it possible for the authors to report a quantitative example where both methods (SANA-WN and HLA-WM) achieve a higher PSNR (or SSIM) value? If the base PSNR is at least around 27, then improvements in PSNR are more strongly correlated with improvements in image quality. Perhaps there is a toy problem that can establish the utility of the proposed method?
3 comments
AuthorsOpenReview ↗

Response to reviewer ac16

Q1. The reported PSNR/SSIM values are low, making the quantitative improvements difficult to interpret.

A1. This concern stems from a misunderstanding of revisit PSNR, which compares the first-visit and revisit frames within the same generated video, rather than measuring reconstruction quality against ground truth. We clarify this distinction through qualitative evidence and additional experiments.

  1. Our qualitative results and additional analysis demonstrate substantial improvements in revisit consistency across different PSNR ranges. As shown in Fig. 5 and Figs. 7–9, HLA-WM better preserves scene structure, object arrangements, and spatial relationships across long-range revisits. Notably, Fig. 8 shows a 236-frame revisit where PSNR improves from 14.59 to 17.73 dB (+3.14 dB), SSIM from 0.4800 to 0.5882, and LPIPS from 0.5784 to 0.4116. Another example in Fig. 9 improves PSNR from 12.15 to 14.97 dB (+2.82 dB). To further verify that these gains are not limited to selected examples, we group the Stage 1 results from Simple80 and Hard80 by baseline PSNR.
Baseline PSNR (dB)Below 88–1010–1212–1414–16≥16
Average gain (dB)+0.62+0.81+0.61+0.88+1.51+2.69

HLA-WM achieves positive average improvements across all PSNR ranges. Together with the qualitative results, this demonstrates that our gains are not limited to low-PSNR scenarios or isolated examples.

  1. Conventional reconstruction PSNR thresholds are not directly applicable to long-horizon revisit evaluation. PSNR values around 27 dB are reported for aligned image reconstruction, such as the VQ/KL autoencoders in LDM [1]. In contrast, revisit evaluation compares frames generated at different time steps, where small viewpoint differences can substantially reduce absolute PSNR. To illustrate this sensitivity, we simulate horizontal camera rotations on a single image and compare the transformed images against the original.

    Rotation0°0.1°0.5°1°2°3°4°5°10°
    PSNR (dB)∞31.4825.4122.7620.4518.4517.2216.9212.92

    Even a simulated 0.5° rotation reduces PSNR to 25.41 dB, while a 3° rotation yields only 18.45 dB, despite using the same source image. Importantly, such factors affect the absolute metric values without invalidating relative comparisons. HLA-WM and the baseline use identical prescribed camera trajectories and revisit-frame pairs under the same evaluation protocol. Thus, although minor viewpoint differences can lower absolute PSNR, improvements over the baseline remain meaningful measures of long-range scene consistency.

    These results demonstrate that requiring 26–27 dB imposes an inappropriate pixel-alignment threshold on realistic long-horizon revisits. Our evaluation instead measures improvements in scene consistency under matched conditions. We have clarified the metric interpretation in the Evaluation Metrics section and added this analysis in Appendix A.8.

Q2. Could the authors provide a higher-PSNR/SSIM example or a toy experiment demonstrating the method's effectiveness?

A2. Our existing qualitative examples and additional controlled experiment address this concern. As shown in Fig. 8, HLA-WM improves revisit PSNR from 14.59 to 17.73 dB (+3.14 dB) and SSIM from 0.4800 to 0.5882, with visibly better preservation of scene structure. Furthermore, the controlled experiment in A1 demonstrates that even minor viewpoint differences substantially reduce PSNR without changing the underlying scene content. These results explain why achieving 26–27 dB is not an appropriate requirement for long-horizon revisit evaluation. We therefore focus on demonstrating improved scene consistency under realistic benchmark conditions rather than constructing artificially aligned high-PSNR examples.

References

[1] Rombach et al. High-Resolution Image Synthesis with Latent Diffusion Models. CVPR 2022.

I appreciate the authors' response and I maintain my original rating.

However, I take issue with the claim that my concerns in Q1 stem "from a misunderstanding of revisit PSNR". Differences in PSNR values, particularly those that fall between 10 and 18, do not correlated strongly with image quality. The authors have actually demonstrated this point in the camera-rotation experiment in the rebuttal. The fact that PSNR is applied to first-visit and revisit frames does not change this point, and I feel using this as an image quality metric in this work remains a weakness. These several limitations of PSNR suggest that an alternative image quality metric may be more appropriate when comparing first-visit and revisit frames.

AuthorsOpenReview ↗

Follow-up comments

Thank you for your clarification and thoughtful feedback. We apologize for our earlier wording, which did not accurately reflect your concern.

We agree that PSNR alone cannot fully capture perceptual revisit quality, particularly when small viewpoint differences affect pixel-level alignment. Our evaluation follows the official SANA-WM benchmark protocol, which uses PSNR, SSIM, and LPIPS to assess revisit consistency. We adopt these established metrics to ensure a fair and consistent comparison with the baseline, rather than introducing a separate evaluation protocol. Importantly, the improvements are supported not only by PSNR, but also by SSIM, LPIPS, and extensive qualitative comparisons.

We appreciate your comments and will clarify the limitations of PSNR in the revised manuscript. Thank you again for your time and careful consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.