acceptodds
Under review as a conference paper at ICLR 2027

SeaMem: Seen-Aware Implicit Memory for Long-Horizon Camera-Controlled World Generation

Abstract

Long-horizon world generation aims to enable controllable exploration of a scene beyond initial observations by generating views along user-specified camera trajectories. As the viewpoint changes, each target view may contain both previously seen content and newly revealed regions. The former must remain consistent with historical observations, while the latter requires plausible synthesis guided by learned world priors. The key challenge is therefore how to represent and use historical observations to preserve previously seen content while allowing plausible synthesis in newly revealed regions. Existing methods typically rely on either explicit 3D memory or implicit context memory. However, 3D reconstruction can accumulate persistent geometric errors when integrating imperfect generated views. In contrast, implicit context memory avoids this reconstruction error, but lacks awareness of seen and unseen regions. Consequently, previously observed content may be unnecessarily resynthesized, introducing hallucinations and degrading long-horizon consistency. To address these issues, we introduce SeaMem, a seen-aware implicit memory for long-horizon camera-controlled world generation. SeaMem couples implicit recovery of historical content with explicit identification of where the recovered content should constrain generation. Specifically, SeaMem queries compact historical representations with the target camera to recover content and predicts an unseen-region mask from attention statistics. Recovered content anchors seen regions, while a pretrained novel view synthesis (NVS) backbone synthesizes unseen regions. Within seen regions, a detail-deficient mask relaxes anchoring where historical detail is insufficient, enabling refinement under historical constraints. This design extends short-range NVS backbones to long-horizon world generation without requiring additional backbone fine-tuning. Extensive experiments demonstrate that SeaMem effectively extends NVS backbones to long-horizon world generation while achieving robust revisit consistency.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Novel View Synthesis, Implicit View Memory, Camera-Controlled Generation, Long-Horizon Consistency

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_1MRzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents SeaMem, a seen-aware memory for long-horizon camera-controlled world generation. The main idea of the paper is to dientangle the reconstruction and generation process. Specifically, it first conducts reconstruction by rendering the scenes reconstructed implicitly from the input tokens into explicit images. Two lightweight predictors then estimate unseen and detail-deficient region masks, which are then used to guide a frozen NVS backbone to go through the generation process. Experiments on RealEstate10K and ACID using SEVA and ViewCrafter demonstrate the proposed metho improves short-range generation quality and long-horizon revisit consistency, with additional component and predictor-feature ablations.

Critical strengths
  • The motivation of disentangling generation and reconstruction is good and insteresting.
  • Experiments demonstrate the proposed method improves short range memory and generation quality for novel view synthesis.
  • The paper is well-written and easy to understand.
Critical weaknesses
  • Although SeaMem stores historical observations as implicit tokens, it still relies on recovered target-view images to condition the downstream NVS backbone, which is explicit, and very similar to those point-cloud-memory based methods. As point cloud could also provide visible regions and invisbile regions, just like the proposed mask prediction module, I would recommend that the authors to provide a comparison with point-cloud-based memory under the same backbone and demonstrate the effectiveness of the proposed method.

  • The other concern of me is that the method appears incremental, mainly combining the original one-stage LVSM network with an extra initialization stage to improve its robustness. However, this two-stage design also involves some possible drawbacks:

    a) The additional recovery and mask-prediction stages introduce computational overhead.

    b) The recovered content in the first stage may have artifacts and errors which could be preserved and propagated to the downstreaming networks to provide some wrong guidance.

    I would recommend that the authors report end-to-end inference time and peak GPU memory usage for different history lengths, evaluate robustness to recovery errors, and clarify how incorrect historical content can be corrected to better justify the proposed two-stage design.

Questions for authors

N.A.

3 comments
AuthorsOpenReview ↗

Clarifications on Implicit Memory, Novelty, Computational Cost, and Recovery Reliability

Thank you for your constructive feedback. We address memory representations, contributions beyond LVSM, computational cost, and recovery-error propagation, with additional experimental results.

1. Implicit Memory versus Explicit Point-Cloud Caching (W1)

We agree that SeaMem and point-cloud memory can both provide target-view content and coverage information. However, SeaMem is an implicit memory framework, fundamentally different from explicit point-cloud memory in how historical information is stored, queried, and utilized. Point-cloud memory stores explicit 3D samples and retrieves content through geometric projection. In contrast, SeaMem stores compressed latent tokens and recovers content through camera-conditioned attention, without constructing or rendering an explicit 3D representation. Historical support is estimated from recovery attention statistics rather than projected point occupancy. Importantly, the image-based interface does not change the implicit nature of SeaMem's stored memory. Instead of serving as an ordinary conditioning image, the recovered content and predicted masks guide the NVS backend to selectively generate or refine content in appropriate regions.

Following your suggestion, we have added a comparison with a Gen3C-style point-cloud cache with the same SEVA backend in Appendix D.2, Table 8 (lines 976–985). SeaMem + SEVA achieves better performance across all three settings. These results demonstrate the effectiveness of SeaMem compared with point-cloud cache.

SettingMemory guidancePSNR ↑\uparrow (dB)SSIM ↑\uparrowLPIPS ↓\downarrow
Smooth MotionGen3C Cache + SEVA18.730.6750.363
Smooth MotionSeaMem + SEVA20.390.7260.310
Random MotionGen3C Cache + SEVA15.600.5930.455
Random MotionSeaMem + SEVA16.220.6130.439
RevisitGen3C Cache + SEVA19.590.7130.324
RevisitSeaMem + SEVA21.170.7430.281

2. Novelty and Contributions Beyond Existing Memory Paradigms (W2)

To our knowledge, SeaMem is the first world-generation memory framework to predict explicit historical-support masks from implicit memory for spatially selective anchoring. Unlike implicit context memory, this guidance reduces unnecessary resynthesis and hallucination-induced inconsistencies. Unlike explicit 3D memory, it avoids reconstruction and fusion that may accumulate geometric errors from inconsistent observations.

Specifically, SeaMem compresses observed views into latent tokens and queries them with the target camera to recover historical content. Recovery attention statistics provide visibility and detail-support cues for lightweight mask predictors. The resulting unseen and detail-deficient masks guide a frozen NVS backend to preserve supported content, refine insufficient details, and synthesize unseen regions.

3. End-to-End Computational Cost (W2)

We have added complete-pipeline measurements at two rollout lengths in Appendix D.1, Table 7 (lines 953–965). We evaluate the 63-view short-range and 243-view long-horizon revisit trajectories, containing 60 and 240 generated targets, respectively. Total generation time increases from 196.1 to 197.5 seconds (+0.7%) for short-range generation and from 797.3 to 824.2 seconds (+3.4%) for long-horizon revisit. Peak GPU memory increases from 11.87 to 12.79 GiB and from 11.87 to 15.10 GiB, respectively. Overall, SeaMem introduces minimal end-to-end runtime overhead and only a modest increase in peak GPU memory usage.

4. Historical Preservation and Error Correction (W2)

Preserving historically supported content is a deliberate design choice in SeaMem. Allowing the generator to freely rewrite previously observed regions may introduce hallucinated changes and cause scene drift during long-horizon generation. SeaMem therefore applies stronge anchoring to well-supported seen regions, weaker anchoring to detail-deficient regions, and no direct anchoring to unseen regions.

To evaluate the reliability of recovered content as a memory anchor, we additionally compare its recovery quality against ground truth in seen, non-detail-deficient regions on both datasets:

DatasetRecovery outputPSNR ↑\uparrow (dB)Regional-crop FID ↓\downarrow
RE10KSeaMem prior25.80618.413
ACIDSeaMem prior27.33411.513

These results support the quality of recovered content as a memory anchor. We appreciate your suggestion and will explore explicit error-detection and correction mechanisms in future work.

Dear Reviewer,

As the discussion phase is approaching its end, we would like to kindly follow up on our previous responses. We have carefully considered your concerns regarding the value of learned generation and the reproducibility of our repair method. If possible, we would sincerely appreciate your feedback on whether our clarifications and additional experiments have adequately addressed your main concerns. Should any questions remain, we would be happy to provide further clarification during the remaining discussion period.

Thank you again for your time and valuable feedback on our work.

Dear Reviewer,

Thank you again for your time and constructive feedback. As the discussion period is nearing its end, we would greatly appreciate your consideration of our latest response above. We hope the additional clarifications and evidence address your concerns, and would be grateful if you could reconsider your assessment in light of them. Please let us know if any concerns remain unresolved; we remain available to clarify them. Thank you for your consideration.

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

SeaMem augments pretrained novel-view synthesis models with a queryable implicit memory of historical views. An LVSM-based compressor and recovery network reconstruct content at a requested camera pose. Lightweight predictors use attention statistics and auxiliary features to identify unseen regions and seen regions lacking sufficient detail. These masks control latent anchoring during generation: preserve supported content, release detail-deficient areas earlier, and synthesize unseen areas freely. The NVS backbones remain frozen, while the memory and mask-prediction components are trained separately. Experiments with SEVA and ViewCrafter on RealEstate10K and ACID evaluate 60-view synthesis and 240-frame out-and-back consistency, demonstrating substantial improvements over the unaugmented backbones.

Critical strengths
  1. The memory-to-generator interface is useful and supported across two backbones (§3; Table 1). Recovering historical content together with spatial support is a concrete way to control where generation may overwrite memory. With SEVA, RE10K revisit PSNR improves from 14.61 to 21.17 dB, while smooth-motion target-view PSNR improves from 18.92 to 20.39 dB. The fixed-backbone comparisons make the contribution of the added memory system reasonably clear.

  2. The paper examines the preservation–synthesis trade-off rather than treating consistency alone as quality (§4.3; Table 2; Appendices A–B). The separate unseen/detail masks, component ablations, predictor analyses, and explicit scope of the memory-capacity test help explain the design. In particular, acknowledging that indiscriminate anchoring can improve revisit scores while degrading synthesis is an important strength.

Critical weaknesses
  1. The long-horizon evaluation does not establish that improved consistency retains good synthesis quality on the same rollouts (§4.1.2; Tables 1–2). Revisit metrics compare two generated images, and the no-unseen-mask variant achieves much higher revisit PSNR than the full model, 32.79 versus 21.17 dB, despite the authors' explanation that it reproduces under-detailed content. The ground-truth quality measurements concern separate, shorter trajectories. Thus, the evaluation demonstrates consistency gains but does not adequately quantify the quality–consistency trade-off during prolonged memory updates. This limits the strength of the long-horizon claim; it does not negate the improvements over the frozen backbones.

  2. Visibility supervision can differ from actual visibility of historical content (Appendix B.2, p. 16). The visibility labels come from forward-projecting historical points with a nearest-depth buffer, explicitly without target-depth agreement. A historical surface can therefore project behind a newly revealed occluder while still contributing coverage. The detail labels, in contrast, require depth-consistent visibility. Because the predicted seen mask determines where generation is anchored, this mismatch is consequential: a false-seen label can preserve unsupported content. The paper should explain this distinction, characterize its limitations, an

Questions for authors
  1. Given the 32.79 dB revisit result without unseen masking, what evidence from the existing long-rollout evaluations supports the full model's quality–consistency trade-off? Are any quality measurements for those same rollouts already available?
  2. Does the visibility-label pipeline contain any additional occlusion test beyond the source-point depth buffer? If not, how should a source surface projecting behind an unobserved foreground surface be interpreted, and how does this relate to the different validity rule for detail labels?
  3. Please specify the actual anchoring/release schedules and mask thresholds used in Table 1, and the number of evaluated trajectories per dataset and setting. These details would clarify the reproducibility and breadth of the reported gains.
5 comments
AuthorsOpenReview ↗

Additional Quality Evaluation and Clarifications on Supervision and Inference

We sincerely thank you for your positive assessment and constructive feedback. We address your concerns regarding long-horizon generation quality, the two supervision signals, and inference reproducibility below.

1. Visual Quality of Long-Horizon Generation (W1, Q1)

We have additionally evaluated the visual quality of long-horizon revisit generation on RE10K and ACID using three reference-free VBench metrics. Furthermore, we conducted long-horizon inference on the 50 RE10K test scenes with the largest pairwise camera rotations and added quantative and qualitive results in Appendix D.4 (lines 1037–1087).

RE10K

MethodAesthetic Quality ↑\uparrowTemporal Flickering ↑\uparrowMotion Smoothness ↑\uparrow
ViewCrafter39.1896.3297.83
SEVA40.7997.2198.63
LIVE39.8697.1898.08
VMem43.9298.0799.15
Gen3C48.4797.4299.40
SeaMem + ViewCrafter44.8698.9099.53
SeaMem + SEVA50.5198.4899.42

ACID

MethodAesthetic Quality ↑\uparrowTemporal Flickering ↑\uparrowMotion Smoothness ↑\uparrow
ViewCrafter40.2394.6897.04
SEVA41.8996.0997.94
LIVE39.6497.9598.30
VMem47.0998.6799.15
Gen3C51.4998.0199.38
SeaMem + ViewCrafter43.8199.3499.56
SeaMem + SEVA50.3899.0299.43

Across both datasets, SeaMem + SEVA and SeaMem + ViewCrafter consistently outperform their respective standalone backends and achieve competitive results compared with other baselines.

Figure 5 (lines 1037–1079) additionally pairs SeaMem + SEVA outputs from the outgoing and returning passes at matching camera poses. These examples show generated scene appearance and revisit consistency within the same long-horizon rollouts, rather than relying only on separate short-range quality examples.

2. Unseen Region Mask and Detail-Deficient Region Mask Supervision (W2, Q2)

Our unseen-region mask supervision already includes a target-surface consistency check based on VGGT point maps. We have revised the supervision procedures for both masks in Appendix B.2 (lines 810–825).

During offline pseudo-label construction for the unseen-region mask, we use reference and target VGGT point maps. Reference 3D points are projected into the target view, excluding invalid and out-of-image projections. At each target pixel, we compare the distances of projected reference points to the target camera with that of the corresponding target-view point. Only surface-consistent reference points provide evidence of historical coverage. Occluded reference points lying behind the target surface do not contribute coverage, and pixels without valid reference support are treated as unseen. Thus, the occlusion case you described is correctly identified as unseen.

The unseen-region and detail-deficient-region masks are constructed using different VGGT outputs: point maps and depth maps, respectively. For the detail-deficient-region mask, target-view surface points are projected into reference views, retaining in-image correspondences with depth agreement within 5%. Magnification is estimated from local projection scales using the valid reference with the strongest detail support. Valid seen patches with magnification of at least 1.3 are labeled detail-deficient, while unseen and invalid patches are excluded from supervision.

3. Inference Settings and Evaluation Coverage (Q3)

We have added the detailed anchoring/release schedules, mask thresholds, and evaluation sample counts to Appendix C (lines 925–933).

For the SeaMem-augmented backends in Table 1, sampling uses 50 denoising iterations. Counting iterations in execution order, sufficiently supported seen regions are anchored for the first 40 iterations and released for the final 10; detail-deficient seen regions are anchored for the first 20 and released for the remaining 30; unseen regions receive no direct memory anchoring. The final 10 unanchored steps promote smooth transitions between preserved and newly synthesized regions while minimizing changes to anchored content. The visibility predictor outputs seen probability: values below 0.60 form the unseen mask. A predicted detail-deficiency probability of at least 0.80 defines the detail-deficient mask, which is restricted to predicted seen regions (lines 925–930).

For both datasets, Smooth Motion, Random Motion, and Long-Horizon Revisit use 100, 100, and 50 trajectories, respectively. Within each setting, every trajectory comes from a different scene, rather than drawing multiple trajectories from the same scene (lines 931–933). Short-range trajectories contain three initial references and 60 generated targets; long-horizon round trips contain three initial references and 240 generated targets (lines 291–298).

Dear Reviewer,

As the discussion phase is approaching its end, we would like to kindly follow up on our previous responses. We have carefully considered your concerns regarding the value of learned generation and the reproducibility of our repair method. If possible, we would sincerely appreciate your feedback on whether our clarifications and additional experiments have adequately addressed your main concerns. Should any questions remain, we would be happy to provide further clarification during the remaining discussion period.

Thank you again for your time and valuable feedback on our work.

Reviewer_pwWZOpenReview ↗

Resolved concerns

Thank you for the detailed response and revisions. The target-surface consistency check fully resolves my visibility-supervision concern, and the added schedules, thresholds, and sample counts address the reproducibility question.

The new long-horizon VBench results and paired visualizations substantially strengthen the evaluation. One concern remains: these reference-free metrics assess perceptual quality and temporal smoothness, but not scene fidelity or recovered detail. Moreover, the new table does not compare the full model against the no-unseen-mask variant on the same rollouts. Thus, the large revisit-PSNR gap (32.79 vs. 21.17 dB) still leaves the preservation–synthesis trade-off only partially quantified. A same-rollout quality comparison with that ablation, or appropriately tempering the corresponding claim, would fully address this point.

Overall, the rebuttal meaningfully improves the paper and resolves most of my concerns.

Reviewer_7GnQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes seamem, a seen-aware implicit memory for long-horizon camera-controlled world generation. Built on lvsm, historical views are encoded into compressed tokens and queried by the target camera to render a recovered target view. Attention statistics from the same query are used to predict an unseen mask and a detail-deficient mask. The recovered content is then used to anchor the seen regions in the latent space of a frozen nvs backbone while the unseen regions are generated by the backbone and the detail-deficient regions are refined. Generated views are written back into the memory for later segments. Experiments on re10k and acid show improvements on short-range generation and revisit consistency over the backbones.

Critical strengths
  • paper is not hard to understand.

  • Predicting the unseen and detail-deficient regions from the attention statistics is an interesting design.

Critical weaknesses
  • The framing feels a bit off to me. Since seva is frozen and the seen regions are directly anchored to the lvsm rendering, the final content mostly comes from lvsm, and seva only works as a refiner for the blurry or unseen regions. So it is more like using seva to help lvsm, rather than a memory that extends seva to long-horizon generation. From this view the choice of seva is also not very essential, any video generator could be used as the refiner here, especially considering the camera moves in a very short range. Overall the idea is closer to lvsm + a video refiner, and the current story is a bit misleading.

  • pipeline is quite heavy, and many simple engineering level baselines are missing. For example, one can first use seva to generate sparse keyframes along the trajectory, and then use lvsm to interpolate the remaining views with the camera poses. seva itself also already supports long trajectories with its procedural two-pass sampling (Sec. 3.2 in seva) and a memory bank of generated anchor views (Sec. 3.3 in seva), but its not clear whether these are used. In the supplementary videos, seva already collapses when returning to the first pose, e.g., LongTermRevisit sample001. vmem is also almost the same as seva on re10k revisit, while the vmem paper reports around 6 dB gain over seva on similar cycle trajectories. Overall the baselines look weak, and how they are run is not well-described in the paper.

  • The scope feels relatively small. The whole trajectory is very short in the supp videos. Emphasizing revisit consistency and memory under such a short and simple setting is not very meaningful and convincing to me. For long-horizon world generation, the memory should matter when the camera explores a much larger area with large viewpoint changes and revisits a region after a long time, which is not evaluated here

Questions for authors

See weakness

6 comments
AuthorsOpenReview ↗

Clarifications and Additional Evidence on SeaMem's Design and Evaluation

Thank you for your constructive review. We address the memory-generation roles, efficiency, baseline fairness, and spatial evaluation scope below.

1. Memory Recovery and Plausible Synthesis (W1)

We agree that SeaMem is intentionally designed to use LVSM for historical content recovery and a generative NVS backend for refining insufficient details and synthesizing unseen regions.

LVSM provides accurate recovery of historically supported content. On RE10K, its recovery achieves 25.81 dB PSNR and 18.41 regional-crop FID in sufficiently supported seen regions. In contrast, unseen regions achieve only 18.89 dB PSNR and 202.53 FID, motivating the use of a generative NVS backend to synthesize plausible unseen content.

Predicted regionRecovery PSNR ↑\uparrow (dB)FID ↓\downarrow
Seen, not detail-deficient25.8118.41
Unseen18.89202.53

This design preserves consistency with historical memory in seen regions while leveraging the generative priors of NVS backends to refine insufficient details and synthesize plausible unseen content.

SEVA is not the only applicable NVS backend. As demonstrated in Table 1 (lines 368 and 377), SeaMem works with both SEVA and ViewCrafter without additional backbone fine-tuning.

2. Pipeline Complexity and Computational Cost (W2)

SeaMem follows a streamlined three-stage pipeline consisting of memory recovery, region-mask prediction, and region-guided synthesis. The target camera queries memory to obtain recovered content and attention statistics. Two predictors then estimate region masks from these statistics. The recovered content and masks guide a frozen NVS backend, and newly generated views are written back to memory.

Our additional efficiency analysis shows that SeaMem introduces minimal end-to-end runtime overhead with only a modest increase in peak GPU memory usage. We have added complete-pipeline runtime, peak GPU memory, and component-level timings in Appendix D.1, Table 7 (lines 953–965). For the 63-view short-range and 243-view long-horizon revisit trajectories, total runtime increases from 196.1 to 197.5 seconds (+0.7%) and from 797.3 to 824.2 seconds (+3.4%), respectively, while peak GPU memory increases from 11.87 to 12.79 GiB and from 11.87 to 15.10 GiB.

3. Comparison Fairness and Baseline Configuration (W2)

SEVA's two-stage sampling and keyframe-first interpolation are designed for fully specified trajectories, whereas generating future keyframes in advance would violate the sequential-exploration constraint. We therefore adopt one-pass autoregressive SEVA sampling without future-keyframe-first generation or native anchor retrieval. Similar sequential-generation settings are also adopted in other long-horizon world generation methods, such as VMem and WorldPlay.

VMem's reported results cannot be directly extrapolated to our evaluation due to differences in generation horizon and sampling protocol. VMem samples one generated view every ten source-video frames, yielding at most 54 round-trip views given RE10K's maximum sequence length of 279 views, compared with 240 in our evaluation. Our generation horizon is also comparable to WorldPlay's round trips of at least 250 frames and Geometry Forcing's 256-frame full-rotation evaluation with revisits.

We have now added the baseline configurations to Appendix C (lines 910–923).

4. Scope of Long-Horizon Evaluation (W3)

To address your concern about limited viewpoint variation, we additionally evaluated the 50 RE10K test scenes with the largest pairwise camera rotations. Quantitative and qualitative results have been added in Appendix D.4 (lines 1037–1087).

RE10K and ACID are widely used datasets that generally exhibit limited viewpoint rotations. Thus, the small viewpoint changes in our original evaluation reflect the dataset characteristics rather than deliberate selection of easy scenes. According to our statistics, only 0.73% of RE10K sequences and 0.98% of ACID sequences exceed 90∘90^\circ in maximum azimuth change. The 50 selected RE10K test scenes exhibit maximum pairwise camera rotations ranging from 127.83∘127.83^\circ to 179.99∘179.99^\circ. The table below compares their results with those from the original random selection for SeaMem + SEVA and SeaMem + ViewCrafter.

Figure 5 (lines 1037–1079) shows SeaMem + SEVA outputs for four scenes from this subset, pairing outgoing and returning views at matching camera poses to illustrate generation beyond the initial reference coverage and preservation of scene content during revisits.

MethodScene selectionRevisit PSNR ↑\uparrow (dB)SSIM ↑\uparrowLPIPS ↓\downarrow
SeaMem + ViewCrafterTop 50 by camera rotation19.210.7170.339
SeaMem + ViewCrafterRandomly selected20.790.7450.307
SeaMem + SEVATop 50 by camera rotation20.020.7440.293
SeaMem + SEVARandomly selected21.170.7430.281

Dear Reviewer,

As the discussion phase is approaching its end, we would like to kindly follow up on our previous responses. We have carefully considered your concerns regarding the value of learned generation and the reproducibility of our repair method. If possible, we would sincerely appreciate your feedback on whether our clarifications and additional experiments have adequately addressed your main concerns. Should any questions remain, we would be happy to provide further clarification during the remaining discussion period.

Thank you again for your time and valuable feedback on our work.

Thank you for the response + additional exps. runtime analysis and baseline configurations are helpful, but my main concerns remain: rebuttal confirms that the seen content comes from LVSM and the backbone mainly refines the remaining regions, and that it can be replaced by ViewCrafter, so the method is still closer to LVSM plus a video refiner than a memory for the generator. Also SEVA is run without its own anchor memory which keeps the baseline weak. I still lean towards keeping my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.