acceptodds
Under review as a conference paper at ICLR 2027

PebbleVideo: On-Device Video Generation via Extreme Model Compression

Abstract

Diffusion Transformers (DiTs) have substantially advanced video generation quality, yet their quadratic bidirectional attention and billion-scale parameter incur significant memory and storage overhead, hindering deployment on resource-constrained devices. Existing approaches improve efficiency through attention redesign or streaming generation, but their restricted global spatiotemporal interactions limit modeling capacity. In contrast, we explore an alternative direction, model compression, and introduce **PebbleVideo**, which aggressively reduces the parameter footprint through structured compression. The core of PebbleVideo is to evaluate compression sensitivity in functional space, preserving generation-critical functions rather than relying on indirect proxies. For the dominant DiT backbone, we measure structural importance through the perturbation induced on the diffusion vector field, directly exposing each module's influence on generation dynamics. For heterogeneous condition encoders, cross-scale representation consistency determines whether compression proceeds through compact-model substitution or in-place structural pruning. Combined with DMD-based distillation, PebbleVideo achieves 512320 video generation at 28 FPS with only 7.45 GB peak memory. Further incorporating model quantization, we deploy the quantized PebbleVideo on a Samsung mobile NPU, achieving a VBench score of 87.46.

Primary area:
optimization
Keywords:
Model compression, Video generation, Device deployment

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_JdHSOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents a new efficient video generation system called PebbleVideo which incorporates DiT pruning, encoder distillation, step distillation, and quantization for on-device deployment. Its main contribution is a novel criterion that ranks DiT blocks by comparing the depth-wise sequence of denoising predictions produced by the original and block-skipped networks, using dynamic time warping for alignment. The selected blocks are structurally pruned using Optimal Brain Surgeon (OBS) and recovered through teacher-student distillation. The pipeline also compresses the text and image encoders, reduces sampling from 50 to 4 steps vid DMD distillation, and applies mixed-precision quantization. Experiments on ReCamMaster report a reduction from 8.13B to 1.57B total parameters (of the whole system) and deployment with 1.8 GiB peak memory on a mobile NPU.

Critical strengths
  1. A novel block/structured pruning scheme: Rather than relying only on local representation similarity, it measures how skipping a block changes the depth-wise sequence of denoising predictions. The method seems to be superior to previously proposed heuristics.

  2. The work presents a complete mobile deployment pipeline: It combines several compression techniques and demonstrates that the resulting model can run on a mobile NPU using only 1.8 GiB of peak memory.

Critical weaknesses
  1. The system is not competitive with prior mobile video-generation work, and its novelty is quite limited. For example, Neodragon generates videos at 640 x 1024 x 49 resolution in approximately 7 seconds on-device, whereas the proposed system requires 146.8 seconds to generate 512 x 320 x 49, making it roughly 20 times slower despite operating at a substantially lower spatial resolution. Moreover, PebbleVideo takes an image and an existing video as inputs, making its V2V setting substantially more constrained than Neodragon’s T2V setting. Most components of the proposed pipeline are established techniques already used in prior work. The pruning scheme is the main new idea, but it is evaluated on only one V2V backbone (ReCamMaster) leaving its generality and usefulness for other models unclear.

  2. The evaluation relies heavily on small differences in a custom VBench score. The paper averages only six selected dimensions rather than reporting the official aggregate score. Moreover, VBench metrics can be noisy and do not necessarily provide a reliable measure of perceptual quality, particularly when interpreting small score differences. Although the paper includes a human evaluation of the final system, it does not establish whether the small improvements attributed to the proposed pruning method are perceptually meaningful.

Questions for authors
  1. Why was ReCamMaster chosen as the only backbone, given that it is specialized for camera-controlled V2V generation?

  2. Can the authors evaluate TAFD and show its superiority over other heuristics on another DiT backbone, preferably a general-purpose T2V model?

  3. Can the authors provide a controlled comparison with the latest mobile video generation systems (e.g., SnapGen-V, Neodragon, S2DiT, MobileWan) using the same task, hardware, resolution, and frame count? What are the advantages of the proposed system over previous work?

6 comments
AuthorsOpenReview ↗

Response to Reviewer JdHS (1/3)

Many thanks for the reviewer's valuable comments, which help us a lot to improve our work. We address your concerns as follows.


W1.1: The system is not competitive with prior mobile video-generation work, For example, Neodragon generates videos at 640 x 1024 x 49 resolution in approximately 7 seconds on-device, whereas the proposed system requires 146.8 seconds to generate 512 x 320 x 49, making it roughly 20 times slower despite operating at a substantially lower spatial resolution.

Thank you for your valuable comment.

Neodragon's video generation pipeline consists of three stages: (1) SSD generates the first frame; (2) T2V diffusion generates a video at 512 × 320 × 49 resolution; and (3) QuickSRNet performs 2× super-resolution to produce a video at 640 × 1024 × 49 resolution. Neodragon reports an end-to-end latency of 6.7 seconds for the complete pipeline.

Since inference latency can vary significantly across hardware platforms and deployment pipelines, we deploy Neodragon on the same Snapdragon 8 Elite Gen 5 platform using the same deployment procedure as our method for a fair comparison. We evaluate Neodragon with its first-frame generation and video diffusion stages, excluding super-resolution, at 512 × 320 × 49 resolution. The results are shown below:

ModelMaximum Visual Token Sequence LengthLatency (s)
Neodragon-T2V4,48093.02
PebbleVideo-V2V8,960146.80
PebbleVideo-V2V5,12064.24

As shown, our model performs V2V generation, whereas Neodragon performs T2V generation. Our task requires processing twice as many visual tokens, resulting in longer inference latency. However, at comparable visual token sequence lengths, our model achieves lower latency than Neodragon, indicating a lower per-token computational cost enabled by aggressive model compression.

This evaluation on the same hardware platform demonstrates that the higher latency of our model primarily stems from the greater computational demands of the V2V task rather than limitations of our method's design.

W1.2 & Q1: PebbleVideo takes an image and an existing video as inputs, making its V2V setting substantially more constrained than Neodragon’s T2V setting. Why was ReCamMaster chosen as the only backbone, given that it is specialized for camera-controlled V2V generation?

We sincerely thank you for your valuable comment.

We would like to emphasize that the core motivation of this work is to explore whether computationally intensive video generation models can be deployed on resource-constrained mobile devices solely through aggressive model compression. To this end, our objective is to develop a unified framework based on functional-space compression sensitivity assessment to select appropriate compression strategies for different components of video generation models.

Based on this motivation and objective, we explain our choice of V2V as the baseline in the Appendix C.1, Model Selection of the original paper. Specifically, at the same generation resolution and frame count, V2V processes nearly twice as many visual tokens as T2V and I2V, substantially increasing computational and memory costs. Thus, V2V imposes greater resource demands on mobile devices, better aligning with our motivation of enabling on-device deployment through aggressive model compression.

Moreover, V2V involves text, image, and video encoders, providing more diverse heterogeneous conditioning components than T2V and I2V. This allows us to systematically investigate the compression characteristics of different encoders within a single model and establish a unified functional-space compression sensitivity assessment framework, better aligning with our research objective.

AuthorsOpenReview ↗

Response to Reviewer JdHS (2/3)

W1.3: Its novelty is quite limited. Most components of the proposed pipeline are established techniques already used in prior work.

Thank you for your valuable comment.

Many existing studies on mobile video generation adopt a combination of model compression, few-step distillation, and hardware quantization for deployment. Our work likewise employs established DMD distillation and quantization techniques. However, we would like to emphasize that our core contribution is not a simple combination of existing techniques, but a unified framework based on functional-space compression sensitivity assessment to guide structural compression of different video generation components.

Specifically, existing compression methods typically rely on indirect proxies, such as feature similarity or changes in generation objectives, to assess component importance. These proxies may not accurately reflect the impact of structural removal on the model's core generative function. To address this limitation, we propose two key designs based on the functions of different video generation components.

  • For the DiT backbone, we map block outputs at different depths into a unified denoising function space and directly assess the vector-field deviation caused by structural removal by aligning the functional trajectories of the original and pruned models, rather than relying on indirect metrics such as local feature similarity or generation loss. Building on this, we further develop a progressive pruning and recovery strategy guided by TAFD sensitivity assessment.

  • For heterogeneous conditioning encoders, we employ cross-scale representation consistency to evaluate how well different compression strategies preserve conditioning functionality, thereby selecting either compact model replacement or in-place structural pruning. This enables encoder compression strategies to be guided uniformly by functional preservation rather than predefined empirical choices.

W1.4 & Q2: The pruning scheme is the main new idea, but it is evaluated on only one V2V backbone (ReCamMaster) leaving its generality and usefulness for other models unclear. Can the authors evaluate TAFD and show its superiority over other heuristics on another DiT backbone, preferably a general-purpose T2V model?

Thank you for your valuable and constructive suggestion.

Following your suggestion, we further apply the TAFD pruning method to Wan2.1-1.3B-T2V to validate its generalization across different DiT backbones and generation tasks.

We follow the pruning sensitivity ablation settings in Fig. 5 of the main paper, keeping the pruning configuration and recovery training procedure identical while varying only the sensitivity assessment criterion for a fair comparison. For the pruned models, we follow the official VBench evaluation protocol and report the official aggregate score (Total Score), Quality Score, and Semantic Score. The results are shown below:

MethodTotal Score ↑Quality Score ↑Semantic Score ↑
Wan2.1-1.3B-T2V (Original)82.4783.3379.01
Input–Output Deviation80.9983.0572.77
DiT Output Deviation80.6782.8272.09
TAFD81.5583.3074.56

As shown, TAFD outperforms previous feature-deviation-based sensitivity assessment methods across all three VBench metrics. These results demonstrate that TAFD is effective not only for V2V but also for general-purpose T2V generation, more accurately identifying pruning-insensitive structures and better preserving generation quality under the same compression settings.

AuthorsOpenReview ↗

Response to Reviewer JdHS (3/3)

W2: The evaluation relies heavily on small differences in a custom VBench score. The paper averages only six selected dimensions rather than reporting the official aggregate score. Moreover, VBench metrics can be noisy and do not necessarily provide a reliable measure of perceptual quality, particularly when interpreting small score differences. Although the paper includes a human evaluation of the final system, it does not establish whether the small improvements attributed to the proposed pruning method are perceptually meaningful.

Thank you for your valuable comments.

VBench does not support an official aggregate score for V2V tasks. Therefore, the original V2V model, ReCamMaster, does not report an official aggregate score either. Instead, it evaluates generation quality using six VBench metrics covering video consistency, visual aesthetics, and other aspects. To ensure consistency with the evaluation protocol of the original model, we adopt the same six metrics and additionally report their arithmetic mean to facilitate a more intuitive and concise comparison of overall performance across methods.

Experimental results show that our compressed model achieves an average score only 0.49 points lower than the original model, demonstrating the effectiveness of our method in preserving generation quality.

We agree that existing automated video evaluation metrics can be noisy and may not accurately reflect changes in perceptual quality. Nevertheless, VBench remains a widely adopted and well-recognized benchmark in video generation, providing a valuable quantitative reference for generation quality. For this reason, we also conducted a human preference evaluation. The results show that our method achieves a preference rate of 76.5% against on-device Sora and 38.6% against the original ReCamMaster. These results further validate the perceptible visual quality of our compressed model from a human perceptual perspective.

Q3: Can the authors provide a controlled comparison with the latest mobile video generation systems (e.g., SnapGen-V, Neodragon, S2DiT, MobileWan) using the same task, hardware, resolution, and frame count? What are the advantages of the proposed system over previous work?

Thank you for your valuable and constructive suggestion.

Following your suggestion, we further apply the complete compression pipeline to Wan2.1-1.3B-T2V, including encoder compression, structured DiT pruning, and few-step distillation. Due to time constraints, we have not yet completed mobile quantization and deployment of this model and therefore cannot provide a mobile quality–latency comparison under strictly matched hardware conditions.

However, as discussed in our response to W1.1, we have reproduced Neodragon on the same hardware platform and found that PebbleVideo achieves lower on-device inference latency at comparable visual token sequence lengths. This provides further experimental evidence of the inference efficiency of our compression method. Based on our completed T2V compression experiments, we further compare model sizes and generation quality, as shown below:

MethodDiT BackboneEncoderTotal Score ↑Quality Score ↑Semantic Score ↑
Wan2.1-1.3B-T2V1.42B5.81B82.4783.3379.01
Neodragon E2E2.97B1.97B81.6183.6873.36
Neodragon T2V Multi-Step2.97B1.97B80.2183.5466.90
Ours1.02B0.65B80.8283.6269.61

As shown, Neodragon achieves a higher Total score than our method through its multi-stage generation pipeline incorporating SSD-based first-frame generation, and we acknowledge this advantage. However, when excluding the multi-stage pipeline and focusing solely on model compression, we compare our method with the other Neodragon variant (reported in NeoDragen paper), which directly compresses the T2V diffusion model without SSD-based first-frame generation. Our compressed model achieves a Total Score of 80.82 with only four denoising steps, outperforming the 80.21 achieved by this Neodragon variant using multi-step generation. Furthermore, our compression pipeline reduces the DiT backbone and encoder parameters to 1.02B and 0.65B, respectively, substantially reducing model storage and weight memory overhead. We sincerely thank you again for your valuable comments.

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

This paper presents PebbleVideo, an end-to-end compression framework for deploying video diffusion Transformers (DiTs) on resource-constrained devices while preserving full bidirectional attention. The framework addresses the substantial parameter, memory, and computational overhead of existing video generation models.The paper's central contribution is a function-space perspective on structured model compression, supported by an end-to-end deployment pipeline and experiments on both server and mobile hardware.

Critical strengths
  1. The paper provides a systematic, theoretically motivated comparison of existing block-pruning criteria and identifies their fundamental limitations in the context of video diffusion transformers. In particular, local feature similarity may fail to capture the downstream effects of removing a block, while changes in aggregate objectives may obscure how pruning errors propagate along the denoising trajectory.

2.The paper presents a well-motivated and practically grounded framework, supported by extensive experiments and ablation studies that validate its design choices and demonstrate its practical effectiveness.

Critical weaknesses

1.The paper could benefit from a brief discussion of the computational cost of the compression pipeline. While the primary goal is to reduce on-device inference cost, reporting the overall compression time or training overhead would provide a more complete picture of the practical trade-off between the one-time optimization cost and the resulting inference efficiency.

2.Limited evidence on cross-distribution generalization. The default TAFD calibration uses a single prompt–video pair, and the evaluation primarily relies on VBench metrics. Although the authors show that the pruning order remains consistent with more calibration pairs, this does not necessarily establish that the compressed model preserves generation quality across diverse prompt and motion distributions. Evaluating the model on an independent benchmark or a more diverse set of prompts would strengthen the evidence for its generalization capability.

Questions for authors

1.How robust is TAFD across models, tasks, and sampling configurations? Since the main experiments are built on ReCamMaster, have the authors evaluated whether the block rankings and pruning benefits generalize to other video DiTs, conditioning modalities, or denoising-step configurations?

2.Could the authors clarify the datasets used for TAFD calibration, vector-field recovery, and DMD distillation, and whether they overlap with the VBench evaluation distribution? How sensitive is the final generation quality to the diversity of calibration prompts and motion patterns?

3.Table 2 compares PebbleVideo’s four-step DMD generator with baselines that mostly use default multi-step sampling. Because sampling steps are not matched, the reported 34× speedup mixes structural compression with step distillation. How much each factor contributes? The paper lacks a controlled comparison under the same sampling steps.

8 comments
AuthorsOpenReview ↗

Response to Reviewer Hewt (1/2)

Many thanks for the reviewer's valuable comments, which help us a lot to improve our work. We address your concerns as follows.


W1: The paper could benefit from a brief discussion of the computational cost of the compression pipeline. While the primary goal is to reduce on-device inference cost, reporting the overall compression time or training overhead would provide a more complete picture of the practical trade-off between the one-time optimization cost and the resulting inference efficiency.

Thank you for your constructive and valuable comment.

The computational cost of our compression pipeline consists of four stages: text encoder compression, image encoder compression, DiT backbone compression, and few-step distillation. All experiments were conducted on 16 NVIDIA A800 GPUs. Following your suggestion, we report the computational cost of each stage in the table below.

StageText Encoder CompressionImage Encoder CompressionDiT Backbone CompressionFew-Step Distillation
GPU Hours (A800)16016670114

W2 & Q2: Limited evidence on cross-distribution generalization. Could the authors clarify the datasets used for TAFD calibration, vector-field recovery, and DMD distillation, and whether they overlap with the VBench evaluation distribution? Evaluating the model on an independent benchmark or a more diverse set of prompts would strengthen the evidence for its generalization capability. How sensitive is the final generation quality to the diversity of calibration prompts and motion patterns?

Thank you for your valuable and constructive comments.

The Prompt–Video Pairs used for TAFD calibration were randomly sampled from the official Wan2.1-I2V video set provided by VBench. Both Vector-Field Recovery and DMD Distillation use the MultiCamVideo dataset, the original training dataset of ReCamMaster, which contains over 120K videos. These training data configurations are detailed in Tables 7 and 8 of Appendix C.4.

TAFD aims to measure the relative functional contribution of each block to the pretrained model's denoising trajectory. Since this relative importance is primarily determined by the model's learned denoising mechanisms, it exhibits a certain degree of stability across different input samples. Table 12 in Appendix J confirms this: even when expanding the calibration set to 10 Prompt–Video Pairs, TAFD produces exactly the same block pruning order.

In contrast, Vector-Field Recovery and DMD Distillation aim to restore and optimize the generation capability of the compressed model, and their training involves learning the data distribution. We train on MultiCamVideo and achieve generation quality close to the original model on VBench, which comes from a different data source. This demonstrates the cross-distribution generalization capability of the compressed model.

We also performed TAFD calibration using one Prompt–Video Pair from the MultiCamVideo dataset. Due to time constraints, we evaluated the pruning selections for the first six blocks and found them to be identical to those obtained using VBench Prompt–Video Pairs.

Following your suggestion, we further provide evaluation results on FiVE-Bench, which includes both real and synthetic videos covering diverse motion patterns involving humans, animals, vehicles, and other categories. As shown in the table below, our model maintains generation quality close to that of the original model Recammaster.

Modelbackground_consistencyimaging_qualityTemporal Flickeringaesthetic_qualitymotion_smoothnesssubject_consistencyVBench Score
ReCamMaster95.0966.9996.1951.8298.5993.1983.64
Ours94.7666.5896.7550.9798.6092.6583.38

[1] FiVE-Bench: A Fine-grained Video Editing Benchmark for Evaluating Emerging Diffusion and Rectified Flow Models. ICCV 2025.

AuthorsOpenReview ↗

Response to Reviewer Hewt (2/2)

Q1: How robust is TAFD across models, tasks, and sampling configurations? Since the main experiments are built on ReCamMaster, have the authors evaluated whether the block rankings and pruning benefits generalize to other video DiTs, conditioning modalities, or denoising-step configurations?

Thank you for your valuable and constructive suggestion.

Following your suggestion, we further apply TAFD to Wan2.1-1.3B-T2V and perform structured pruning with 25 denoising steps to evaluate its generalization across different models, tasks, and sampling configurations.

Given the high computational cost of post-pruning recovery training and the limited time available, we primarily evaluate the effectiveness of TAFD by comparing the generation quality of the pruned model with that of the original model.

As shown in the table below, the TAFD-pruned model achieves a six-metric VBench mean (VBench Score) only 0.44 points lower than the original model. This indicates that TAFD maintains good pruning stability across different video DiT architectures, conditioning modalities, and denoising-step configurations.

Wan2.1-T2V (step=25)background_consistencyimaging_qualityTemporal Flickeringaesthetic_qualitymotion_smoothnesssubject_consistencyVBench Score
Original95.9556.1997.7252.7998.4193.9782.50
TAFD95.0255.2997.1252.0299.3193.5982.06

Q3: Table 2 compares PebbleVideo’s four-step DMD generator with baselines that mostly use default multi-step sampling. Because sampling steps are not matched, the reported 34× speedup mixes structural compression with step distillation. How much each factor contributes? The paper lacks a controlled comparison under the same sampling steps.

Thank you for your valuable comment.

Following existing mobile video generation works, Table 2 reports the overall efficiency gains of our compression pipeline, incorporating both structural compression and few-step distillation. We provide a detailed breakdown of their individual contributions in Table 9 of Appendix E.

Specifically, structural compression alone reduces the model size from 8.13B to 1.57B parameters and peak memory usage from 16.93 GiB to 7.46 GiB, while achieving a 1.56× speedup. Further incorporating few-step distillation increases the overall speedup to 34.33×. For your convenience, we reproduce Table 9 from Appendix E below. We sincerely thank you again for your valuable comments.

MethodParams (B)Peak Memory (GiB)Latency (s)
Baseline8.1316.9360.08
+ Encoder compression2.348.8359.00
+ DiT pruning1.577.4638.05
+ DMD distillation1.577.451.75

The FiVE-Bench results show a 0.26 deficit for the proposed method, and the response describes this as maintaining generation quality close to that of the original model ReCamMaster. In the paper, a 0.43 difference is presented as an advantage over the baseline, and a 0.28 difference separates TAFD from the competing criterion. Judged by the same standard, the 0.26 deficit here is equally an indication that the method has a weakness. So I keep my current score.

Reviewer_5wRXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PebbleVideo, an end-to-end compression pipeline for deploying a pretrained video diffusion model on mobile hardware. Its main proposed method, TAFD, scores DiT blocks by comparing projected depth-wise denoising trajectories with and without a block, aligning the trajectories before measuring their deviation. The selected blocks undergo progressive attention-head and FFN-channel pruning, OBS compensation, and vector-field recovery. The authors use CKA to choose between compact-model substitution for the text encoder and in-place pruning for the image encoder, followed by four-step DMD distillation and mixed-precision quantization. The reported pipeline shrinks from 8.13B to 1.57B parameters and generates a 49-frame video in 1.75 seconds on an RTX 5090.

Critical strengths
  1. The system-level implementation is substantial. The paper addresses the condition encoders, DiT backbone, sampling steps, and quantized deployment together. Its staged results help explain which components reduce memory and which reduce latency.

  2. TAFD is an identifiable methodological contribution. It attempts to assess how skipping a block affects downstream denoising predictions, while accounting for the depth shift caused by the skip.

Critical weaknesses
  1. The overall pipeline overlaps substantially with prior work, while TAFD’s independent contribution remains insufficiently isolated. There were a lot of prior works already cover closely related combinations of mobile video generation, sensitivity-guided pruning, recovery, and few-step distillation. TAFD is therefore the central novelty claim. In Figure 5, it improves the reported score from 87.12 for point-wise trajectory deviation to 87.40, but this experiment prunes only eight blocks. The final comparison also changes the pruning and recovery procedure, making it difficult to attribute its gain specifically to TAFD.

  2. The mobile efficiency and quality claims need more careful qualification. The 28 FPS figure is measured on an RTX 5090; Table 3 reports 146.8 seconds for complete generation on the mobile NPU. These results establish that mobile execution is feasible, but limit stronger conclusions about practical mobile speed and near-lossless quality.

Questions for authors
  1. How different are the blocks selected by TAFD from those selected by the Neodragon and FastLightGen criteria in the final compression setting?

  2. Could the authors report ranking correlations and compare these criteria while holding the pruning ratios, OBS compensation, and recovery procedure fixed?

  3. TAFD constructs its trajectory across network depth at a fixed diffusion timestep. How strongly does its score correlate with the actual vector-field error and video quality after the proposed within-block pruning and recovery?

  4. Could the authors provide mobile quality–latency results under matched input modalities, resolution, and frame count, and clearly separate their six-metric mean from the official VBench aggregate?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 5wRX (1/3)

Many thanks for the reviewer's valuable comments, which help us a lot to improve our work. We address your concerns as follows.


W1.1: The overall pipeline overlaps substantially with prior work, while TAFD’s independent contribution remains insufficiently isolated. There were a lot of prior works already cover closely related combinations of mobile video generation, sensitivity-guided pruning, recovery, and few-step distillation.

Thank you for your valuable feedback.

Model compression (including sensitivity-guided pruning and recovery) and few-step distillation are widely adopted as complementary techniques in existing mobile video generation methods, targeting redundancy in model size and inference steps, respectively. In this regard, PebbleVideo shares a similar overall optimization pipeline with prior works. However, we would like to emphasize that the core contribution of PebbleVideo lies not in combining these techniques, but in addressing structural redundancy from the perspective of component functional spaces, enabling more accurate compression sensitivity assessment and structural selection.

W1.2: TAFD improves the reported score from 87.12 for point-wise trajectory deviation to 87.40, but this experiment prunes only eight blocks. The final comparison also changes the pruning and recovery procedure, making it difficult to attribute its gain specifically to TAFD.

Thank you for your valuable and constructive comments.

Following your suggestion, we have conducted additional ablation experiments on DiT sensitivity metrics at the target pruning scale (20 blocks). The results are presented in the table below. As shown, the advantages of TAFD become more pronounced under this more challenging pruning setting.

MethodBackbone ParamsVBench ScoreBackground ConsistencyImaging QualityTemporal FlickeringAesthetic QualityMotion SmoothnessSubject Consistency
Point-wise0.92 B86.1396.1371.0696.2460.0098.1995.16
TAFD0.92 B87.1796.8172.9096.6661.6598.4696.55

We would like to clarify that the ablation study in Fig. 5 of the main paper follows a strictly controlled experimental setup, where the pruning scale and recovery training strategy are kept identical, with only the sensitivity evaluation metric varied. Specifically, we compare TAFD against existing metrics based on (1) block-level input–output discrepancies, (2) final DiT output discrepancies, (3) endpoint deviations of function-space trajectories, and (4) accumulated deviations across trajectory points. By isolating the sensitivity metric from other factors, these experiments directly demonstrate the pruning performance gains attributable to the more accurate sensitivity assessment provided by TAFD.

Considering the computational cost of full-scale pruning experiments, Fig. 5 adopts a setting in which 8 blocks are pruned. Following your suggestion, we will further complete the ablation experiments at the full target pruning scale in future work. Thank you again for your constructive suggestion.

AuthorsOpenReview ↗

Response to Reviewer 5wRX (2/3)

W2: The mobile efficiency and quality claims need more careful qualification. The 28 FPS figure is measured on an RTX 5090; Table 3 reports 146.8 seconds for complete generation on the mobile NPU. These results establish that mobile execution is feasible, but limit stronger conclusions about practical mobile speed and near-lossless quality.

Thank you for your valuable comments.

Regarding generation efficiency, we have explicitly stated in the original Introduction that the 28 FPS generation speed was measured on an RTX 5090, and reported the complete mobile generation time of 146.8 seconds in Table 3 of the efficiency evaluation section.

The high on-device inference latency is primarily attributable to our choice of baseline model. As described in Appendix C.1, Model Selection, our motivation for choosing a V2V model was to validate our central hypothesis: aggressive model compression can enable the deployment of computationally intensive video generation models on resource-constrained mobile devices. To this end, we deliberately selected a V2V model with higher computational demands than T2V models. V2V processes approximately twice as many tokens as T2V, and the quadratic complexity of attention further increases its computational cost. Therefore, we acknowledge that our current mobile implementation does not achieve real-time generation. However, our primary objective is to demonstrate the feasibility of deploying computationally intensive video generation models on mobile devices while achieving real-time generation on high-performance GPUs.

Regarding generation quality, the mobile generation visualizations of Fig. 4 and the reported VBench score of 83.47 demonstrate that the mobile model maintains good generation quality.

Q1: How different are the blocks selected by TAFD from those selected by the Neodragon and FastLightGen criteria in the final compression setting.

Thank you for your valuable suggestion.

Neodragon and FastLightGen assess block sensitivity using input–output discrepancies and trajectory endpoint deviations, respectively. In contrast, TAFD evaluates each block's impact on the generation process through function-space trajectory deviations.

Following your suggestion, we report the block pruning sequences under the final compression setting, ordered by pruning priority:

  • TAFD: 23, 22, 13, 11, 12, 15, 14, 10, 20, 21, 24, 16, 6, 8, 9, 7, 5, 4, 17, 3
  • Neodragon: 17, 24, 23, 22, 14, 3, 15, 4, 6, 11, 10, 9, 16, 13
  • FastLightGen: 5, 14, 11, 23, 22, 13, 15, 24, 12, 20, 16, 21, 10, 25

Due to differences in per-block pruning costs, TAFD prunes 20 blocks, while Neodragon and FastLightGen each prune 14 blocks under comparable overall pruning budgets.

The three methods exhibit different sensitivity rankings, particularly among the highest-priority blocks. Specifically, TAFD prioritizes Block 23, whereas Neodragon and FastLightGen prioritize Blocks 17 and 5, respectively. Notably, these two blocks rank only 19th and 17th under TAFD, highlighting the differences in how these criteria assess block importance.

In Appendix Fig. 7, we visualize the generation trajectory deviations after pruning the top-ranked block selected by TAFD, Neodragon, and FastLightGen. The results show that TAFD better preserves consistency with the original model's generation trajectory than Neodragon and FastLightGen, demonstrating its effectiveness in identifying low-sensitivity blocks.

AuthorsOpenReview ↗

Response to Reviewer 5wRX (3/3)

Q2: Could the authors report ranking correlations and compare these criteria while holding the pruning ratios, OBS compensation, and recovery procedure fixed?

Thank you for your valuable suggestion.

The sensitivity metric ablation study in Fig. 5 already follows a strictly controlled experimental setup, where the pruning ratios, OBS compensation, and recovery procedure are kept identical, with only the block sensitivity metric varied. This allows us to directly isolate the impact of different sensitivity criteria on pruning performance.

The resulting block pruning sequences, ordered by pruning priority, are as follows:

  • Input-output deviation: 17, 24, 23, 11, 10, 22, 14, 3
  • DiT output deviation: 14, 23, 22, 24, 21, 19, 18, 11
  • Trajectory endpoint deviation: 14, 23, 24, 25, 22, 18, 19, 11
  • Trajectory point deviation sum: 5, 14, 23, 11, 22, 13, 24, 15
  • TAFD: 23, 22, 13, 11, 12, 15, 14, 10

As can be seen, these criteria exhibit weak ranking correlations, highlighting the importance of accurate sensitivity assessment for pruning performance.

Q3: TAFD constructs its trajectory across network depth at a fixed diffusion timestep. How strongly does its score correlate with the actual vector-field error and video quality after the proposed within-block pruning and recovery?

Thank you for your insightful question.

Although each depth-wise trajectory is constructed at a fixed diffusion timestep, TAFD aggregates sensitivity scores across all denoising timesteps, rather than relying on a single timestep.

A direct correlation with post-recovery vector-field error or video quality is difficult to establish, as progressive pruning and recovery continuously alter block sensitivities, while the final performance reflects cumulative interactions among pruned blocks rather than the contribution of any individual block. However, as shown in Eq. (7) of paper, under Euler sampling, the deviation between the final projected states satisfies: ∥g~tL−gtL∥2=∣Δt∣∥v~t−vt∥2\|\widetilde{g}_t^L-g_t^L\|_2=|\Delta t|\|\widetilde{v}_t-v_t\|_2. Since the DTW alignment cost in Eq. (14) incorporates this terminal deviation, TAFD reflects vector-field perturbations while additionally capturing discrepancies across intermediate states. Consequently, selecting blocks with lower TAFD scores helps reduce pruning-induced functional perturbations, facilitating subsequent recovery. As demonstrated by our sensitivity metric ablation study, TAFD achieves better generation quality compared with alternative metrics.

Q4: Could the authors provide mobile quality–latency results under matched input modalities, resolution, and frame count, and clearly separate their six-metric mean from the official VBench aggregate?

Thank you for your valuable suggestion.

Following your suggestion, we have applied the complete compression pipeline to Wan2.1-1.3B-T2V, including encoder compression, structured DiT pruning, and few-step distillation. Due to time constraints, we have not yet completed mobile quantization and deployment of this model and therefore cannot provide actual mobile latency results. However, based on our completed T2V compression experiments, we report the evaluation results of the current compressed model on a A100 GPU.

Under matched input modalities (T2V), resolution (512x320), and frame count (49), we compare our compressed model with the original Wan2.1-1.3B-T2V and report the official VBench aggregate scores. The results are presented below, the compressed model also preserves generation quality. Due to time constraints, we reduced the training budget for each component during compression. Matching the training budget of the original work could potentially yield better results. Following your suggestion, we will complete its on-device deployment and evaluation in future work. We sincerely thank you again for your valuable comments.

ModelLatency E2E ↓official VBench scores ↑
Wan2.1-1.3B-T2V27.5 s82.47
Open-Sora V1.2-79.76
Ours1.3 s80.82

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary of Submission 40891

Dear Area Chair,

Thank you for your time and consideration. We sincerely thank all reviewers for their constructive suggestions and valuable insights. Below is a summary of our rebuttal and discussion.

First, we are encouraged by the reviewers' recognition of PebbleVideo's motivation, technical contributions, and practical value. Reviewer Hewt highlighted our "systematic, theoretically motivated comparison of existing block-pruning criteria" and recognized our framework as "well-motivated and practically grounded", supported by extensive experiments. Reviewer 5wRX acknowledged the substantial system-level implementation and identified TAFD as a methodological contribution. Reviewer JdHS also recognized our novel structured pruning scheme and complete mobile deployment pipeline, demonstrating practical deployment on resource-constrained devices.

Second, the reviewers provided valuable suggestions concerning TAFD's novelty and effectiveness, generalization, computational cost, and on-device competitiveness. In response, we conducted extensive additional experiments, clarified our contributions, and revised the manuscript accordingly.

  • We clarified PebbleVideo's core novelty, emphasizing functional-space compression sensitivity assessment rather than simply combining existing compression and distillation techniques (5wRX W1.1; JdHS W1.3).
  • We extended the controlled TAFD ablations to 20 pruned blocks, achieving a 1.04-point improvement over the best competing sensitivity criterion. We also provided block-selection rankings and trajectory visualizations to further validate TAFD's effectiveness (5wRX W1.2, Q1–Q3; Hewt follow-up).
  • We validated TAFD's cross-model generalization on Wan2.1-1.3B-T2V under 25- and 50-step sampling configurations. Under controlled pruning settings, TAFD consistently outperforms alternative sensitivity criteria, demonstrating its effectiveness beyond the original V2V backbone (Hewt Q1; JdHS W1.4, Q2).
  • We extended cross-distribution evaluation to FiVE-Bench, where PebbleVideo achieves 83.38 compared with 83.64 for the original model, preserving most generation quality (Hewt W2, Q2).
  • We conducted additional on-device comparisons with NeoDragon using identical hardware and deployment procedures. We clarified the computational differences between V2V and T2V and showed that PebbleVideo achieves lower latency at comparable visual token lengths (64.24 s vs. 93.02 s) (JdHS W1.1, Q3; 5wRX Q4).
  • We clarified the evaluation protocols and quality comparisons, justified the use of V2V evaluation metrics, and provided additional evidence through T2V experiments and human preference evaluations. (JdHS W2; 5wRX Q4).

Following the reviewers' suggestions, we incorporated these additional results into the revised manuscript, including Sec. 5.3 (computational cost and generalization), Appendix C.5 (datasets and data usage), and Appendix O (on-device inference efficiency comparison).

During the follow-up discussion, Reviewer JdHS raised further questions about system-level competitiveness and perceptual validation. We provided additional clarification on the different task requirements, fair deployment comparisons, and evaluation protocols. We hope these responses help address the reviewer's remaining concerns.

Discussion Outcomes. We are particularly encouraged by the reviewers' positive follow-up feedback:

  • Reviewer 5wRX stated that our responses "mostly addressed my concerns" and described PebbleVideo as "very solid" and making "a contribution to this community." The reviewer maintained the positive rating and explicitly expressed willingness to recommend acceptance during the AC–Reviewer discussion.
  • Reviewer Hewt, after carefully examining the revised Sec. 5.3 and relevant appendices, confirmed that the authors had "addressed most of my concerns" and explicitly expressed willingness to raise the score.
  • Reviewer JdHS acknowledged that the additional T2V experiments strengthened the evidence for TAFD, while maintaining concerns about overall competitiveness and validation. We provided further point-by-point responses and hope they clarify our contributions and experimental evidence.

We believe our additional experiments, detailed clarifications, and manuscript revisions have substantially strengthened the paper and addressed the major technical concerns raised during the review. We sincerely appreciate the reviewers' constructive feedback and consideration.

Finally, we again thank all reviewers for their time and effort, and especially thank the Area Chairs for their 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.