acceptodds
Under review as a conference paper at ICLR 2027

Dandelion: Towards Real-time and Consistent Long Video Depth Estimation

Abstract

Recently, video depth estimation methods have successfully exploited the priors of pre-trained video generative models to enhance geometric perception, but the costly inference limits low-latency deployment. In this paper, we present Dandelion, the first generative streaming video depth estimation model, which reduces first-prediction latency and enables real-time estimation. However, directly adapting traditional frame-wise streaming strategies remains suboptimal due to the repeated per-frame inference and the lack of bidirectional temporal context. To address these issues, through systematic analysis, we introduce the streaming chunk strategy and tailor a dedicated chunk length to achieve both high depth accuracy and low-latency inference. Moreover, a stable reference is indispensable for temporal consistency. To achieve this, we design a two-stage training protocol, which first establishes a geometric anchor from GT depth history and extend its robustness throughout autoregressive rollout. To further accelerate inference, we recognize that details in distant depth history are less critical to current estimates. Thus, we design Multi-scale History Attention to compress distant context. With these advances, Dandelion achieves both state-of-the-art zero-shot performance and superior effeciency. Using only 0.2% depth supervision used by prior methods (VDA), Dandelion achieves a 12.9% performance gain and supports real-time depth estimation over long videos at 24 FPS.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Depth Estimation, geometry perception, generative model

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

This paper proposes Dandelion, a generative framework for streaming video depth estimation. To address challenges in inference latency, temporal consistency, and historical information utilization, the authors adapt several techniques from autoregressive video generation, including chunk-wise causal modeling, two-stage training, and multi-scale memory.

Overall, the paper achieves strong results in depth estimation accuracy, temporal consistency, and inference efficiency, demonstrating practical value.

However, my main concern is that most core components closely resemble existing autoregressive video generation methods, yet their relationships and technical differences receive little or no discussion.

My final rating will depend on the authors' response.

Critical strengths
  1. The paper systematically analyzes several key challenges in streaming video depth estimation. The overall methodology is clearly presented.

  2. The paper achieves strong depth estimation accuracy and temporal consistency while maintaining real-time inference efficiency, demonstrating its practical value.

Critical weaknesses

Most core components of this paper are closely related to existing autoregressive video generation methods. However, the authors do not adequately discuss these prior works.

Specifically:

  1. Two-stage Training Strategy: The second stage uses model-generated depth predictions as autoregressive history, which is consistent with the underlying idea of Self Forcing [1], which addresses the train–inference gap through self-generated history. However, the paper lacks a discussion of their relationship and differences.

  2. Multi-scale Depth Memory: This module closely resembles the history compression mechanism in FramePack [2], even adopting the same three-level compression configurations. However, FramePack is neither cited nor discussed in the manuscript.

  3. Processing video in chunks with bidirectional attention within each chunk and causal attention across chunks is also a commonly adopted design in autoregressive video works.

While I acknowledge the practical value of this work, I believe that adequately discussing related work and clearly distinguishing technical contributions from existing methods are particularly important for a research conference like ICLR.

[1] Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion, NeurIPS 2025

[2] Frame Context Packing and Drift Prevention in Next-Frame-Prediction Video Diffusion Models, NeurIPS 2025

Questions for authors
  1. How do the core components of this paper relate to and differ from the prior methods mentioned in the Weaknesses?

I also have the following specific questions:

  1. In Sec. 3.4, historical depth latents are compressed at different spatiotemporal scales and concatenated with the current RGB tokens before being fed into the DiT. However, it is unclear how RoPE positional encoding is handled for tokens.

  2. Fig. 5 presents qualitative comparisons with and without the two-stage training strategy. Since the second stage primarily aims to improve robustness to accumulated prediction errors during autoregressive inference, it would be helpful to clarify whether the visualized depth predictions come from the initial chunk or subsequent chunks.

  3. In Fig. 9, Only Stage 2 eventually outperforms Only Stage 1, which is somewhat surprising and deserves further analysis. Additionally, the training procedure for Only Stage 2 needs clarification, particularly how it is implemented without initialization from Stage 1.

6 comments

Overall:

We sincerely thank the reviewer for recognizing our systematic analysis, strong depth estimation performance, temporal consistency, and inference efficiency. Below, we address each weakness (W) and question (Q) point by point. We will incorporate the clarifications and citations into the revised manuscript.

W1 & Q1: Differences & Relationships from Self Forcing

The two methods address fundamentally distinct tasks: Dandelion focuses on accurate and efficient geometry estimation, whereas Self Forcing focuses on generating robust videos. The former has corresponding GT depth available for constructing historical geometric context, whereas the latter lacks a unique GT, as generation requires diversity. This difference leads to the following distinctions in geometric anchoring.

Stage 1 uses GT depth as depth history to establish a reliable geometric anchor. The model learns to use this anchor as a geometric reference to guide subsequent predictions. Stage 2 inherits the geometric anchoring prior learned in Stage 1 and continually updates the geometric reference with new depth predictions. In contrast, Self Forcing uses self-generated RGB history to guide video generation and optimizes distribution matching, without a corresponding geometric anchor. Moreover, we admit that using self-generated histories reduces the training–inference distribution gap, which is shared by both methods.

W2 & Q1: Differences from FramePack

The tasks differ fundamentally. Dandelion addresses a geometric prediction task, recovering accurate depth from observed RGB videos. FramePack [1] addresses a video generation task, synthesizing diverse RGB content. This distinction in task objectives fundamentally changes the purpose and motivation of multi-scale history compression.

Our three-level compression is motivated by the observed temporal propagation of depth history. Fig. 6(b) shows three levels of detail relevance across temporal distances. Recent history exhibits strong detail relevance, whereas distant history retains persistent scale similarity. These observations motivate preserving recent geometric details while encoding intermediate and distant history at progressively coarser resolutions to retain global geometric guidance. In contrast, FramePack studies context packing based on temporal proximity for video generation, unable to investigate these depth-specific characteristics, such as detail relevance and scale similarity.

[1] Frame Context Packing and Drift Prevention in Next-Frame-Prediction Video Diffusion Models, NeurIPS 2025

W3 & Q1: Differences from Chunk Design

Notably, to our knowledge, prior streaming video depth estimators operate frame by frame and have not explored the chunk-based mechanism. Our experiments in Sec. 4.2 show that frame-wise processing is suboptimal, motivating us to introduce streaming chunks for video depth estimation.

Then, the information processed within each chunk differs: Dandelion uses predicted depth history and current RGB observations, treating past geometry as a reference and limited future RGB observations as supplementary evidence. This chunk-based mechanism allows causal attention across chunks to convey past depth information to the current chunk, where bidirectional attention integrates this geometric reference with current RGB observations to predict the current depth. In contrast, autoregressive video diffusion generators condition on generated RGB content from past chunks and denoise target latents initialized from noise in the current chunk, using within-chunk interaction to synthesize new RGB content.

Finally, our core contribution here is the systematic analysis of chunk length for streaming depth estimation. This analysis is independent of studies in video generation and, to our knowledge, is the first systematic study of chunk length for streaming video depth estimation. Our experiments further identify an optimal chunk length tailored to depth estimation.

Q2: RoPE Alignment

We use a shared 3D RoPE coordinate system for historical depth and current RGB tokens. Within each rolling window, consecutive temporal indices are assigned to the history followed by the current chunk. To align RoPE with compressed history tokens, we average-pool its cosine/sine tensors on the base patch grid with factors of (1,1,1), (2,2,2), and (4,4,4) for short-, mid-, and long-range history, respectively, with corresponding boundary padding. The aligned positional encodings are concatenated in the same [long, mid, short, RGB] order as the tokens and applied to queries and keys before attention.

Q3: Chunk Selection

The examples in Fig. 5 are taken from subsequent chunks during autoregressive inference.

Q4: Stage 2 Initialization and Performance

Only Stage 1 uses GT depth history throughout training, whereas the depth history in Only Stage 2 is generated by the current model from the outset. Both variants start from pretrained streaming video generation weights.

We interpret Fig. 9 as follows: GT history provides reliable conditioning for faster initial learning but leaves a mismatch at inference. Only Stage 2 is initialized directly from pretrained streaming video generation weights, without using a Stage 1 checkpoint. It already has some ability to extract information from temporal history. Although it faces more challenging conditioning initially, it is trained under history conditions that better match autoregressive inference, consistent with its eventual advantage over Only Stage 1.

Thank you for the clarification. Throughout my review, I have consistently acknowledged the motivation for adapting a FramePack-like mechanism to the current task. The more important thing is that the similarities between the proposed design and FramePack go far beyond the high-level idea. The implementation details, including the RoPE alignment strategy newly clarified in the rebuttal, are all highly consistent with FramePack.

Were these designs independently developed from scratch, or were they adapted from FramePack for the current task? If the latter, why was FramePack neither cited in the original manuscript nor explicitly acknowledged in the rebuttal, while these technical designs were presented as the authors' own contributions?

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

The paper proposes Dandelion, a streaming framework for long-video depth estimation built on a pretrained video generative model. It combines chunk-wise prediction, a two-stage training strategy using ground-truth and predicted depth histories, and multi-scale history compression to improve efficiency and temporal consistency. Experiments on ScanNet, KITTI, and Bonn demonstrate strong depth estimation performance, with additional evaluation on 1,000-frame videos and a reported throughput of 24 FPS on an RTX 4090.

Critical strengths

The paper addresses a meaningful practical challenge: maintaining accurate and temporally consistent depth predictions while processing long videos efficiently. The overall framework is well motivated, and the reported improvements across multiple benchmarks, together with component ablations and long-video evaluation, provide useful evidence of its effectiveness.

Critical weaknesses
  1. The choice of history length and allocation across scales needs further justification. The effect of multi-scale compression on depth accuracy and temporal consistency also needs further evaluation.
  2. The contribution of Multi-scale History Attention beyond FramePack needs clarification.
Questions for authors
  1. How were the history cache length and the allocation across scales determined? The paper assigns 6, 2, and 1 latent frames to long-, mid-, and short-range history, respectively. Have the authors evaluated alternative history lengths or allocations in terms of accuracy and computational cost?
  2. Figure 10 compares efficiency with and without multi-scale history compression. Could the authors also report depth accuracy and temporal consistency for these two settings, using the same historical frames? This would clarify the trade-off between prediction quality and efficiency.
  3. The paper presents Multi-scale History Attention as a main contribution. However, FramePack[1] has already explored compressing historical frames using patch embeddings at different scales based on temporal proximity. Could the authors clarify the specific differences between the proposed design and FramePack? [1] Frame Context Packing and Drift Prevention in Next-Frame-Prediction Video Diffusion Models.
2 comments

Overall:

We sincerely thank the reviewer for recognizing the practical importance of consistent and efficient long-video depth estimation and the effectiveness of our framework. Below, we address each weakness (W) and question (Q) point by point. We will incorporate the additional experiments and clarifications into the revised manuscript.

W1 & Q1: History Length and Allocation

We conducted ablations on history length and compared multi-scale allocations. The results are shown in the table below. All models are evaluated on the ScanNet dataset for a fair comparison.

HistoryAbsRel ⬇️δ1\delta_1 ⬆️TAE ⬇️Total Tokens
[6,2,1]6.297.90.4512460
[4,4,1]6.897.30.5212680
[4,2,3]6.697.50.5114780
[2,2,1]7.396.90.5712380
[14,2,1]6.597.70.4812620

The history configuration [6,2,1] achieves the best depth estimation accuracy and temporal consistency among the evaluated configurations, with only marginally more tokens than the shortest history and fewer tokens than the other alternatives. These results demonstrate the appropriateness of our history length and cross-scale allocation.

W2 & Q3: Differences from FramePack

First, the tasks differ fundamentally. Dandelion addresses a geometric prediction task, recovering accurate and consistent depth from observed RGB videos. FramePack [1] addresses a video generation task, synthesizing coherent and diverse RGB content.

Second, FramePack compresses history to retain longer context within a bounded token budget. Our motivation comes from depth geometry: as discussed in Sec. 3.4, long-range depth history provides important global geometric guidance, while short-range history preserves fine-grained near details. Multi-scale History Attention integrates distant global information with recent geometric details to improve the effectiveness of depth history for accurate streaming depth estimation.

Third, in Dandelion, the multi-scale 3D patch embeddings extract depth-distribution information from predicted depth history. They learn geometry-relevant history representations through depth supervision. In FramePack, the corresponding embeddings extract contextual information from RGB history. They learn visual representations through context optimization. Although the multi-scale 3D patch-embedding architecture is shared, the learned representations are fundamentally distinct. Our contribution concerns this geometry-motivated adaptation and its exploration for streaming depth estimation. We will explicitly cite FramePack and clarify this architectural relationship in the revision.

[1] Frame Context Packing and Drift Prevention in Next-Frame-Prediction Video Diffusion Models, NeurIPS 2025

Q2: Ablation of Multi-scale History Compression

We evaluated both settings on the ScanNet dataset using the same historical frames under matched training and evaluation conditions, as shown in the table below.

HistoryAbsRel ⬇️δ1\delta_1 ⬆️TAE ⬇️FPS
w/ multi-scale6.297.90.4524.2
w/o multi-scale6.397.60.5112.7

Multi-scale compression achieves 1.91 times throughput and reduces TAE by 11.8%, while maintaining comparable depth accuracy. This demonstrates a favorable balance between depth accuracy, temporal consistency, and efficiency, supporting our choice for streaming inference. We will include this comparison in the revised manuscript.

Thanks to the authors' responses, my concerns have been essentially resolved. After carefully reading the other reviewers' comments and the authors' responses, and considering everything together, I have decided to maintain my positive assessment.

Reviewer_x73mOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Dandelion estimates relative depth from video in successive chunks. It uses a pretrained video diffusion model and conditions each chunk on earlier depth predictions. The history cache is updated as new predictions arrive; recent depth latents retain more spatial detail, while older latents are compressed more heavily. Training first uses noisy ground-truth depth history and then the model's own predictions. The paper evaluates depth accuracy and temporal consistency on indoor and outdoor videos, including 1,000-frame sequences.

Critical strengths
  1. Strong Results and Useful Ablations: Dandelion improves depth accuracy over DVD and GemDepth across the evaluated video benchmarks. On the 1,000-frame Bonn sequences, AbsRel improves from DVD's 7.8 to 5.5. Removing depth history in Table 3 and comparing the training stages show that both choices help. Training with the model's own predictions is a reasonable way to reduce the difference between training and inference, following the same general motivation as Self Forcing, although the task, loss, and stored history differ.

  2. Presentation: I found the overall method easy to follow. Figure 4 explains how chunk size affects accuracy and latency. The temporal depth profiles are also useful for showing errors that are difficult to see in individual frames.

Critical weaknesses
  1. Limited Ability to Handle Large Scale Changes: My main concern is the depth range that the normalized predictions can represent. If the normalization is fixed by the initial frames, later scenes may exceed the representable range or have insufficient depth precision under that normalization. For example, moving from a small room to an outdoor scene can substantially increase the required depth range. Updating the history cache does not by itself solve this problem. Adapting the normalization also requires keeping the cached depth and new predictions in compatible coordinates. The paper needs to explain how it handles this change while preserving consistency with earlier outputs. Without such a mechanism, the method is limited to depth ranges that its initial normalization can represent accurately.

  2. Source of the Depth Improvements: Table 3 shows that history helps, but it does not separate better local geometry from reduced scale drift. Figure 6 measures ground-truth depth-distribution similarity, rather than drift in predicted depth. The single-image gains over DVD also show that some improvements do not require temporal history. I would like a more direct explanation of what the distant history improves and when it is useful.

  3. Real-Time Claim and Latency: The reported 24 FPS measures throughput, but does not establish low latency for live input. Figure 10 and Appendix A.2 report around 1.5 s to the first prediction with preloaded inputs and 2.6 s when video acquisition is included. The model uses bidirectional attention within chunks of nine latents (approximately 36 frames), so earlier frames in each chunk must wait for later frames, even after startup. Steady-state capture-to-output latency is not reported. Although the appendix acknowledges buffering, the title and abstract should qualify the real-time claim. The paper needs to distinguish throughput from capture-to-output latency and state which applications can tolerate this delay.

The reported accuracy is strong, but maintaining a stable scale is only part of the problem. For general long-video depth estimation, the method also needs to accommodate changes in the depth range without losing depth precision or consistency with earlier predictions. This is my main reservation, and the authors' explanation will be important to my recommendation. The real-time claim also needs to state the buffering delay clearly.

Questions for authors
  1. For the full model and the variant without long-range history in Table 3, how does predicted depth scale change over time? An analysis of existing predictions would help separate scale stability from other accuracy gains.

  2. At the reported operating point, what are the average and worst-case capture-to-output latencies after startup, including chunk buffering? Please state the input frame rate and resolution and distinguish these measurements from throughput and first-prediction latency.

3 comments

Overall

We sincerely thank the reviewer for recognizing our strong depth estimation results, useful ablation studies, and clear presentation. Below, we address each weakness (W) and question (Q) point by point. We will incorporate the additional experiments and clarifications into the revised manuscript.


W1: Handling Depth Range Changes

The depth range and normalization parameters are not determined by the initial frames. Following VDA and DVD, Dandelion performs prediction in disparity space, whose underlying representation is s=1/ds=1/d, where dd denotes depth. As in prior work, we left the depth–disparity conversion implicit and will explicitly describe it in the revision. The VAE decoder output u∈[−1,1]u\in[-1,1] is mapped to [0,1][0,1] through 0.5u+0.50.5u+0.5, and disparity is subsequently converted to predicted results by taking its reciprocal. Typically, the smallest predicted disparity in the initial frames is around 0.1 rather than zero. In this representation, distant geometry corresponds to small disparity values close to zero. When the camera transitions to a scene containing more distant geometry, those larger depths can be represented by even smaller disparity values, closer to zero, without expanding the numerical output range or adapting the normalization. Since zero disparity corresponds to infinite depth, this parameterization can, in principle, represent arbitrarily large depths. Thus, our pipeline neither fixes a normalization range nor imposes a fixed maximum metric depth.

W2: Sources of Depth Improvements

Direct explanation

As discussed in Appendix A.2, the temporal profiles in Fig. 11 have provided qualitative evidence: removing recent depth history Hs\mathcal{H}^{s} disrupts near geometry, whereas removing distant depth history Hl\mathcal{H}^{l} introduces broad scale drift, highlighted by the colored boxes in Fig. 11.

When Distant History Is Useful

Distant history can help when foreground content changes while background geometry remains stable, such as a nearby person briefly occluding a static background. Earlier cached predictions can provide a coarse background reference, helping reduce scene-wide depth shifts caused by transient foreground changes.

W3 & Q2: Throughput, Latency, and the Scope of the Real-Time Claim

Capture-to-output latency

Steady-state capture-to-output latency is 2.9 s at 24 FPS and a resolution of 640×480640\times480, including chunk buffering and subsequent processing.

Real-time qualification

Following FlashDepth, our "real-time" claim refers to the model's ability to process incoming video while concurrently producing depth estimates with low latency. From the model's perspective, video acquisition is not the depth prediction process itself. Accordingly, our first-prediction latency is only 1.5 s. We will qualify this terminology in the revised version.

Latency and throughput definitions

Throughput measures the number of frames processed per second. Capture-to-output latency (COL) measures the interval from each frame' s acquisition to its corresponding depth output, including waiting for the chunk to be collected and subsequent processing. First-prediction latency measures the time until the first visible depth prediction.

Application

A typical application is live streaming, where delays of a few seconds are common. Even in the worst observed case, our latency remains below 3 s.

Additional COL

The mean and observed maximum steady-state COL after startup are reported below.

AverageMaximum
2.842.95

Input frame rate and resolution

The first chunk is captured at 30 FPS, while subsequent chunks are captured at 24 FPS during steady-state operation. The input resolution remains 640×480640\times480 throughout.

Q1: Depth Scale Claim

Like prior methods such as Video Depth Anything, Dandelion predicts relative depth without a separate "depth scale" prediction. Our discussion of "scale stability" refers to depth estimation performance over ultra-long autoregressive sequences, with "scale" denoting the temporal scale (e.g., 1, 10, 100, or 1,000 frames). Specifically, the 1,000-frame experiment evaluates whether the depth prediction is accurate over extended inference, using AbsRel and δ1\delta_1. We will make this terminology and evaluation scope explicit in the revised version.

AuthorsOpenReview ↗

Authors' Kind Reminder to Reviewer x73m

Dear Reviewer x73m,

Thank you for your time and effort for reviewing our submission. As the discussion deadline (November 18th) approaches, we would greatly appreciate it if you could review our responses and provide any further clarifications at your earliest convenience. We highly value any opportunities to address any potential remaining issues. Your feedback is extremely valuable!

Thanks again for your time and effort!

Best,

Authors of Submission 3867

Thank you. I've checked the rebuttals and would keep my positive recommendation. Please incorporate these new analysis and clarification in the final version.

Comments on the paper

1 comment

Dear Area Chair,

We fully understand the significant workload you are facing during this busy decision period. We are truly grateful for the time and dedication you have invested in maintaining the high standards of the review process.

To assist in your final evaluation, we have prepared a concise summary below.


Rebuttal & Discussion Summary

Reviewer x73m (Score: 3 Weak acceptance)

Strengths:

• Highlighted the improved depth estimation accuracy across the evaluated benchmarks and the ablation studies showing the benefits of depth history and two-stage training. Found the method easy to follow and highlighted the chunk-length analysis and temporal depth profiles as particularly useful.

Concerns:

• Handling large changes in scene depth range, understanding how predicted depth scale evolves with and without distant history, and clarifying the depth-scale and real-time claims, including reporting latency with buffering.

Resolution:

• We explained how disparity accommodates larger depths without rescaling cached predictions, providing ablations to clarify the roles of recent and distant history, and clarified that our depth-scale evaluation measures accuracy over longer sequences. We also measured the capture-to-output latency and distinguished it from throughput and clarified the scope of our real-time claim.

Outcome Summary: After we clarified our claims and provided additional results, the reviewer offered no further comments. We are confident that this productive discussion will lead to a positive adjustment of their score.

Reviewer rZvb (Score: 3 Weak acceptance)

Strengths:

• Recognized the practical importance of accurate, temporally consistent, and efficient depth estimation for long videos and found the framework well motivated. Highlighted the cross-benchmark improvements, component ablations, and long-video evaluations as evidence of the framework's effectiveness.

Concerns:

• Justifying the history cache length and allocation across scales through accuracy and cost comparisons, evaluating how multi-scale compression affects depth accuracy, temporal consistency, and efficiency, and clarifying the contribution of Multi-scale History Attention.

Resolution:

• We added ablations supporting the chosen history length and allocation and showed that, with the same history frames, multi-scale compression improves throughput and temporal consistency while preserving depth accuracy. We also clarified how Multi-scale History Attention combines distant geometry with recent detail.

Outcome Summary: After the discussion, the reviewer explicitly stated, “My concerns have been essentially resolved.” After considering the other reviews and our responses, they decided to maintain their positive assessment.

Reviewer pYai (Score: 2)

Strength:

• Praised the systematic analysis of key challenges in streaming video depth estimation and the clear presentation of the methodology. Highlighted the strong depth accuracy and temporal consistency at real-time inference speeds, underscoring its practical value.

Concern:

• Clarifying Dandelion’s contributions in the context of prior work; explaining how RoPE positions are assigned to compressed depth and RGB tokens; and clarifying the training evidence in Figs. 5 and 9, including which chunks are shown, how Stage 2 alone is initialized and trained, and how to interpret the corresponding results.

Resolution:

• We clarified the relationship by explaining the geometric rationale for two-stage training and multi-scale depth memory. We highlighted our chunk-length analysis for streaming video depth estimation, explained RoPE alignment, and confirmed that Fig. 5 shows later autoregressive chunks. We also clarified the shared initialization and the role of training history in interpreting Fig. 9.

Outcome Summary: After we provided 2 rounds of detailed explanations and illustrated workflows, the reviewer explicitly acknowledged the novelty of our work and made no further comments. We are confident that this productive discussion will lead to an increase in their final score.


We are grateful to the reviewers for the care they put into evaluating our work, and we have made every effort to give each concern the same care in return. Over the course of the rebuttal, we conducted experiments with alternative history configurations, evaluated compression under matched conditions, measured latency, and provided illustrated explanations of our design choices. We believe these efforts address every concern raised across the three reviews. No reviewer raised any further questions after final responses.

We will carefully revise our paper according to reviewer's concerns in final version.

Thank you once again for your service to the community.

With sincere gratitude, The authors of Submission 3867

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.