Training-Free Efficient Video Generation via Dynamic Token Carving

Dear reviewer cBBb,

Thank you for your valuable feedback and constructive comments. We are grateful for your positive assessment of our work, particularly your recognition of its clarity and comprehensive experiments. We will address your comments below and incorporate the necessary revisions in the updated paper.

Q1 Latent Alignment May Introduce Artifacts

We acknowledge this limitation in our discussion (L302-305) and provide a detailed discussion in Appendix C.1. Indeed, latent alignment requires certain trade-offs as a training-free method, since pure pixel-level decode-encode operations consume approximately 40 seconds, which is unacceptable for a video generation acceleration framework. Nevertheless, we provide several mitigation strategies and discussions:

1. High noise levels minimize alignment impact from resize+renoise operations. During early denoising timesteps when noise levels are substantial (883/1000 of the actual denoising trajectory, in G2mf Q2), the resize and renoise operations have minimal impact on latent alignment since the semantic content is not yet well-formed.

2. Noise shift compensation. Discussed in Appendix L339-342, we introduce adaptive noise shift adjustments to maintain consistency across resolution transitions, helping preserve the underlying latent structure.

3. Enhanced semantic conditioning. By incorporating richer semantic conditions during the alignment process (i.e., using enhanced prompts, with results in Appendix Table 5 as validation), we can better preserve content coherence across different resolutions.

4. Faster VAE integration. Future work can leverage faster VAE variants to reduce the computational overhead of pixel-level operations while maintaining alignment quality.

Q2 SFC is Static

Your observation is accurate. SFC construction is completed before denoising and only depends on resolution and dimension ordering (e.g., hwt or thw, representing different orderings). As acknowledged in our limitation section (L305-307), this approach has room for improvement. We explain the rationale behind Jenga's SFC choice and challenges with adaptive partitioning:

1. SFC block partition exhibits inherent locality properties. Attention computation for high-resolution data predominantly exhibits local characteristics (e.g., STA, SeedVR). SFC naturally possesses local-neighborhood properties, where neighbors on the 1D curve correspond to neighbors in 3D space. We validated in Table 3(b) that SFC improves both latency and performance while reducing line-wise drifting artifacts.

2. SFC demonstrates remarkable parameter insensitivity. Our Generalized Hilbert curves maintain locality while being insensitive to parameters. Hand-crafted block partition strategies (local windows in STA & CLEAR, Swin windows in SeedVR) restrict models to specific resolutions, with dimensional padding severely impacting performance. SFC's 1D nature makes it insensitive to resolution and block_size parameters. The table below shows additional padding and computational overhead for 720×1280×129f videos with block_size=128:

"Optimized Local" represents the optimal GPU-friendly block partitioning, still significantly larger than Jenga's 3D SFC.

3. Adaptive Partition: Cannot be defined at arbitrary timesteps. Text-to-video generation starts from pure noise, making it challenging for adaptive token partitioning to extract meaningful semantic information. Determining adaptive partitions through latent semantic correspondence in early denoising timesteps is problematic.

4. Benefits of static approaches: Frequent reordering degrades generation speed. We experimented with changing SFC dimension scanning directions during interleaved forward attention passes to enhance block interactions (similar to Swin-window shift). This approach yielded no significant quality improvements while introducing substantial processing overhead (~20s per video). Token reordering operations cause memory discontinuity and defragmentation overhead. Adaptive block partition strategies would face similar challenges, with additional overhead for constructing the adaptive strategy itself--an overhead that static SFC avoids entirely through pre-computation.

Q3 Adaptability of the Attention Carving

Following our discussion on block partitioning in Q2, we acknowledge that Attention Carving employs heuristic strategies. In practice, during Attention Carving, within each attention head, the block selection for each query block is adaptive and content-dependent. The importance mask $\mathbf{B}_\text{top}$ is dynamically computed based on the current latent features and query-key relationships, making the carving process responsive to the evolving content during generation. While the SFC partitioning strategy is static, the actual attention patterns and block selections are adaptive to the semantic content at each timestep and layer, ensuring that the most relevant spatial-temporal relationships are preserved during the carving process.

As a pure training-free method, the current version avoids the need for offline tuning procedures. Jenga's significantly smaller search space makes it easily adaptable to new models without extensive hyperparameter tuning. It would be a future direction to design a metric to determine adaptive sparsity.

Q4 Title: "Token Carving", It Seems Not Related With ProRes?

ProRes fundamentally reduces token interactions as well. By reducing resolution (and thus the input token count), ProRes effectively reduces token interactions. Both ProRes and Attention Carving share the same core motivation: reducing computational complexity through strategic token interaction reduction. ProRes achieves this through progressive resolution scaling, while Attention Carving accomplishes this through selective attention masking. The "carving" metaphor aptly describes both processes--ProRes "carves" away lower-resolution details in the early stage, while Attention Carving "carves" away less important attention connections. Together, they form a unified framework for efficient video generation that operates across both input token dimensions and attention pattern dimensions.

Q5 Two Methods are Independent (similar to iQ7G Q1.3)

While ProRes and Attention Carving operate independently, the core innovation of Jenga lies in recognizing that attention sparsity and progressive resolution are fundamentally complementary techniques that address different aspects of the computational bottleneck in video diffusion models:

1. Temporal Complementarity: Based on diffusion denoising theory, the generation process naturally progresses from low to high-frequency content. ProRes leverages this by using lower resolutions during early timesteps when content structure is being established, while Attention Carving becomes most effective during later timesteps when high-resolution details are refined but full attention is no longer necessary.

2. Spatial Complementarity: ProRes reduces the total number of tokens (addressing the N in O(N²) complexity), while Attention Carving focuses on reducing the density of token interactions in high resolution (addressing the practical computation within the O(N²) framework). This dual approach creates a multiplicative effect on acceleration.

3. Content-Generation Constraints and Resolution Trade-offs: During content generation phases, AttenCarve cannot employ extremely high attention sparsity levels without compromising content quality, thus limiting its acceleration potential. However, content generation fundamentally does not require high-resolution inputs. ProRes strategically exploits this by reducing token interactions during content generation through lower resolutions, while preserving the ability to apply higher sparsity levels during high-resolution detail refinement when content structure is already established. This creates a sparsity-resolution scheduling: conservative sparsity at low resolution (content generation) and aggressive sparsity at high resolution (detail refinement), maximizing acceleration while maintaining generation quality.

4. Orthogonality: Unlike competing acceleration methods that often interfere with each other, ProRes and Attention Carving operate on orthogonal dimensions--resolution scaling vs. attention sparsity--allowing them to be combined without negative interactions. This explains why our combined approach achieves 7.22× speedup compared to individual contributions of 3.28× (ProRes) and 2.17× (Attention Carving).

Q6 Ablation on Mask Strategies

In Table 3(b) and Figure 7, we provided an ablation study on SFC selection and mask selection. We acknowledge that more comprehensive ablation studies (i.e., incrementally adding strategies) would better validate our method's effectiveness. Here we present a more detailed ablation study to provide better validation of our method:

This comprehensive ablation demonstrates that: (1) Each mask component contributes meaningfully to both quality and efficiency, (2) The importance mask provides the most significant impact on both metrics, (3) The combination of all three masks achieves the optimal balance between generation quality and computational efficiency.

For additional details regarding mask-related implementations, please refer to our response in G2mf Q1.1.

Dear reviewer iQ7G,

Thank you for your valuable feedback and constructive comments. We are grateful for your positive assessment of our work, particularly your recognition of Jenga's generalizability and comprehensive experiments. We will address your comments below and incorporate the necessary revisions in the updated paper.

Q1 The Novelty & Motivation of Jenga

Q1.1 ProRes: Low-to-High Resolution

We agree that multi-stage resolution progression is not entirely novel in isolation. However, several challenges arise in training-free pipelines, including the abnormal FOV problem discussed in L60. Jenga is the first training-free open-sourced progressive generation pipeline for Video Diffusion Transformers. As stated in G2mf Q5, most importantly, ProRes solves the content degradation when adopting high sparsity attention across all timesteps, maintaining both efficiency and quality throughout the generation process.

Regarding the inconsistent expression, early steps and generate low resolution contents represent distinct concepts. In high-resolution generation, early steps produce global content while maintaining normal field of view (FOV). Similarly, in low-resolution generation, early steps also generate global content encompassing overall layout and structures. However, low-resolution content inherently introduces a "zoom-in" FOV degradation, where the generated layout and structures appear zoomed-in, presenting only partial views rather than the intended global perspective. We will clarify this distinction in our revision.

Q1.2 AttenCarve: Sparse Attention is Not New

We acknowledge that utilizing sparsity for attention acceleration is not fundamentally novel. Our work does not claim to be "the first sparse attention design for video generation." Instead, we focus on how to effectively leverage sparse attention through principled design choices and achieve higher sparsity levels while maintaining quality.

Current widely-used attention approaches suffer from limitations: they are either static (lacking content adaptivity, cBBb Q2) or achieve only low sparsity levels ($\sim$0.5). Our method introduces SFC-based block partitioning and cutoff probability mechanisms to address resolution-dependent challenges and achieve substantially higher sparsity ($\sim$0.75-0.80) without quality degradation. The technical contribution lies in the systematic combination of these components to create a practical, training-free acceleration framework that works across different resolutions and model architectures.

Q1.3 Complementary Nature of ProRes and Attention Carving

The core innovation of Jenga lies in recognizing that attention sparsity and progressive resolution are fundamentally complementary techniques that address different aspects of the computational bottleneck in video diffusion models:

1. Temporal Complementarity: Based on diffusion denoising theory, the generation process naturally progresses from low to high-frequency content. ProRes leverages this by using lower resolutions during early timesteps when content structure is being established, while Attention Carving becomes most effective during later timesteps when high-resolution details are refined but full attention is no longer necessary.

2. Spatial Complementarity: ProRes reduces the total number of tokens (addressing the N in O(N²) complexity), while Attention Carving focuses on reducing the density of token interactions in high resolution (addressing the practical computation within the O(N²) framework). This dual approach creates a multiplicative effect on acceleration.

3. Content-Generation Constraints and Resolution Trade-offs: During content generation phases, AttenCarve cannot employ extremely high attention sparsity levels without compromising content quality, thus limiting its acceleration potential. However, content generation fundamentally does not require high-resolution inputs. ProRes strategically exploits this by reducing token interactions during content generation through lower resolutions, while preserving the ability to apply higher sparsity levels during high-resolution detail refinement when content structure is already established. This creates a sparsity-resolution scheduling: conservative sparsity at low resolution (content generation) and aggressive sparsity at high resolution (detail refinement), maximizing acceleration while maintaining generation quality.

4. Orthogonality: Unlike competing acceleration methods that often interfere with each other, ProRes and Attention Carving operate on orthogonal dimensions--resolution scaling vs. attention sparsity--allowing them to be combined without negative interactions. This explains why our combined approach achieves 7.22× speedup compared to individual contributions of 3.28× (ProRes) and 2.17× (Attention Carving).

Q2 Structural Fidelity Metrics

Thank you for raising this important point about evaluation metrics. We acknowledge that structural fidelity metrics like FVD (Fréchet Video Distance) provide additional validation of our method's effectiveness. While our primary evaluation focuses on VBench for comprehensive video quality assessment, we recognize the value of complementary metrics that capture different aspects of video quality. We also provide detailed VBench-Q quality assessments in Appendix Table 6, including video consistency, visual quality, smoothness, dynamic degree, etc.

We conducted additional experiments with FVD to provide a more complete evaluation. As FVD evaluates the distance between training data distribution and inference data distribution, as a training-free method, we adopt 1. the official VBench test result (with different sampling seeds) and 2. the high-quality video dataset Inter4K as real data distributions to evaluate FVD.

The tables in Q3 further compare methods using FVD metrics. The results show that Jenga preserves video generation quality comparable to the baseline, achieving an FVD score of 141 versus 144 in terms of structural fidelity. The FVD results align with the VBench trends observed across different methods.

Q3 TeaCache & Jenga: Independent Results (with FVD metrics)

We appreciate your interest in understanding the relationship between our work and complementary acceleration techniques like TeaCache. The experiments with TeaCache demonstrate that Jenga achieves substantial acceleration independently and can be combined with other orthogonal methods, including feature reuse and few-step models (G2mf Q3).

Here we also report updated performance on the Wan2.1 model. Due to time constraints, we only tested AttenCarve + Skip on the Wan series in the submission. We have updated the Jenga-Turbo setting, which further boosts efficiency.

As discussed in Q2 and G2mf Q3, our results demonstrate that Jenga's acceleration is fundamentally independent of feature caching techniques like TeaCache. We believe that further improvement on the basis of a mature acceleration method is both challenging and necessary. One reason for the less significant speedup ratio is that the native input resolution (480P videos) is relatively small, resulting in a smaller proportion of attention computation. However, the reduction in the timestep dimension is linear. While TeaCache focuses on reusing computed features across timesteps, Jenga reduces computation through progressive resolution and selective attention, while also boosting generation quality. This orthogonality means the methods can be combined for potentially even greater acceleration benefits without quality degradation.

Future Integration: We plan to explore combinations with TeaCache and other orthogonal efficient generation methods in future work, such as token purning, SageAttention (quantized attention), etc.

[1] Stergiou, Alexandros and Poppe, Ronald, AdaPool: Exponential Adaptive Pooling for Information-Retaining Downsampling, 2021

Q2. In Table 2, the acceleration effect of the last row is obviously not as good as TeaCache.

The differences between methods across the two tables have been mentioned in the original Q3.

One reason for the less significant speedup ratio is that the native input resolution (480P videos) is relatively small, resulting in a smaller proportion of attention computation.

HunyuanVideo's input is 115K tokens total, while Wan2.1-1.3B's input is 31K tokens total. This causes the attention computation time to account for a much smaller proportion of the overall computation (77.8% for 115K and 47.1% for 31K), which is precisely the core acceleration principle of Jenga. Additionally, due to different attention sparsity levels between the two models, we set different cutoff probabilities $p$, which also affect the acceleration ratio of corresponding operators.

Meanwhile, TeaCache's setting proportions differ between the two models. In HunyuanVideo, TeaCache runs for 23/50 inference steps, while in the Wan2.1 series, TeaCache only runs for 15/50 steps. Below is a simple revisit of the tables in the rebuttal:

We respectfully request that you acknowledge and understand our research efforts. Our method focuses on "token carving," which reduces token interactions, so it is fundamentally different from feature-reuse at the methodological level. Even with very few token inputs or low attention sparsity, TeaCache may be faster than our proposed method. We did not simply comparing AttenCarve + ProRes with TeaCache, as they essentially address problems from two orthogonal dimensions. While our method demonstrates greater utility for high-resolution and large-scale models, we position it not as a direct competitor to TeaCache, but rather as a complementary approach addressing different optimization dimensions.

Most importantly, we can achieve additional acceleration on top of methods like TeaCache without performance degradation. However, TeaCache cannot be used together with few-step distilled models, such as AccVideo (5NFE, Jenga achieved a 3.16× speed-up, discussed in G2mf Q3).

Therefore, we are not concerned that using AttenCarve + ProRes achieves lower acceleration ratios than TeaCache in certain specific scenarios, because the combined results of the Jenga and TeaCache are better than using TeaCache alone in terms of both efficiency and quality.

We will include the above discussion in our revision and emphasize the importance and orthogonality of the timestep skip. Thank you very much for your insights and response!

Dear reviewer iQ7G

Thank you for your timely comments! We will address each question below:

Q1. In the first table, FVD (the last row) surpasses the original performance, but in the second table, it does not. Why?

Your observation is very astute! FVD metrics are related to the video generation distribution itself, particularly the potential changes in generated content domain when introducing ProRes. As we can see, after introducing ProRes, the FVD-OriginModel results become "worse", but the FVD-Inter4K results become "slightly better". This indicates that while our method may have some distribution differences from the original model (even though these differences, as explained in A1, fall within normal fluctuations), these differences may be related to the model architecture itself.

1. Since HunyuanVideo uses the mmDiT architecture while Wan uses the Self-Cross Attention architecture, the attention score amplifier we designed to solve FOV degradation in mmDiT cannot be fully utilized in Self-Cross Attention. Even if we simply increase the cross-attention weights, we cannot directly modify the self-attention results through this weight. Therefore, there may be some differences in results, but we believe these differences are acceptable. We must also acknowledge that some designs are indeed more suitable for the mmDiT architecture.

2. This difference does not affect the distribution gap between our method and high-quality real-world videos (FVD-Inter4K), ensuring that the generated videos maintain high quality. As mentioned in our response, VBench includes multiple metrics related to video quality. We provide detailed VBench-Q quality assessments in Appendix Table 6, including video consistency, visual quality, smoothness, dynamic degree, etc.

3. FVD metric reliability: We contend that FVD metrics inherently allow for a certain degree of fluctuation. When we conducted inference using a different seed, the results improved substantially. We believe that metrics within reasonable ranges do not necessarily indicate degradation. Moreover, removing TeaCache from Jenga-Turbo consistently outperforms TeaCache-Only across all quality metrics.

Dear reviewer N3G7,

Thank you for your valuable feedback and constructive comments. We are grateful for your clear positive assessment of our work, particularly your recognition of its dynamic token carving strategy and comprehensive experiments. We will address your comments below and incorporate the necessary revisions in the updated paper.

Q1 Stage Ablation for Progressive Resolution

Thank you for this insightful question. This is indeed an important ablation study we conducted but did not systematically report. We provide supplementary experiments in the table below:

• The experiment uses Jenga-Turbo's basic setting, containing one low-res > high-res stage transition.

We have included a detailed analysis in G2mf Q2 as follows:

The experimental results reveal key insights about trajectory-guided resolution scheduling. Due to the timestep shift mechanism in DiT, more computational steps are allocated to the early stages of the trajectory. Even in the final 10% of timesteps, the actual trajectory still contains approximately 43.7% of the total steps.

We observe a dramatic quality degradation when low-res steps exceed 50% of the total timesteps, while introducing latent misalignment artifacts, confirming that multi-stage generation requires sufficient high-resolution denoising to maintain video quality. Meanwhile, when stage 1 (low-res) steps are too few, quality also decreases. This aligns with our motivation (L48-49) that low-res content generation requires a higher attention density to establish proper content structure. This finding validates why ProRes and AttenCarve work synergistically--they are not merely independent methods but complementary components that address different phases of the generation process.

The 50% low-res step configuration achieves the best balance between quality (83.07%) and efficiency (225s), demonstrating that trajectory curvature analysis effectively guides optimal resolution scheduling decisions.

Meanwhile, ensuring sufficient high-resolution steps is indeed a critical factor for maintaining generation quality in Jenga. This is partially a consequence of adhering to our training-free methodology. Some works like FlashVideo propose that "high resolution generation is relatively easy to optimize into few steps." Training a few-step detail generation model is certainly possible; however, such approaches require substantial data and computational resources, and may be superseded by stronger foundation generation models. Our training-free approach provides a more sustainable and generalizable solution that can readily adapt to newer and more powerful foundation video generation models without requiring retraining.

Q2 The Impact of Cutoff Probability

This is a valuable question. The cutoff probability threshold $p$ is one of the key hyperparameters that controls the sparsity-quality trade-off in our attention carving mechanism.

We conducted systematic experiments to understand its impact in HunyuanVideo, with a Jenga-Turbo setting.

• The latency results may have some fluctuations (not exceeding 5s), but the overall trend is correct.

From these results, we derive the following analysis: Higher $p$ values make the model results approach Full Attention. We computed the standard deviation (STD) of selected block quantities across different heads and the average difference between AttenCarve and full attention. Due to varying feature extraction capabilities of different heads (visualized in Appendix C.2), the STD increases with higher $p$ values. We also observe that the local nature of attention is revealed: Effective Sparsity is significantly greater than the theoretical average selection value (57.5% vs. 10%).

We believe that both low and high cutoff probability values are reasonable in this context: With high $p=0.9$, covering more probability regions ensures a certain degree of sparsity while allowing each head to retain most of the original features. Although $p=0.9$ does not achieve optimal performance in HunyuanVideo, it closely approaches the full attention baseline (82.57%, same as ProRes + timeskip in Table 1) and achieves the best results in Wan2.1.

With lower $p=0.3$, AttenCarve selects the most important blocks, with the removal of long-tailed features in attention balancing the block selections among heads (with a small STD) and boosting performance. We guarantee higher sparsity while partially preserving the most important components in certain global heads.

Intermediate $p$ values (0.5, 0.7) introduce relatively unimportant and partial features in attention heads (evidenced by the dramatic STD increase across heads without fully preserving global head representations), thereby reducing sparsity without improving results. We will incorporate this discussion into the revision, and thank you for prompting us to explore these ablation phenomena more deeply.

Q3 Module-Wise Speed-Up Ablations

We provided individual results for AttenCarve and ProRes in Table 1. AttenCarve and ProRes achieve speedups of 2.17× and 1.51× respectively (3.28× when combined with timestep skip). For a module-wise evaluation, we provide detailed parameter configurations in Table 4 in the Appendix. Below is an excerpt from Table 1, with additional module combinations. We use a two-stage basic setting (Jenga-Turbo), where attention sparsity is also applied in a two-stage manner in AttenCarve experiments.

The results demonstrate that the acceleration ratio can essentially be viewed as the product of individual method speedups. The synergistic combination of these approaches enables Jenga to achieve high acceleration ratios while maintaining video generation quality. The organic integration of ProRes and AttenCarve working together produces even greater performance gains than the sum of their individual contributions would suggest.

Dear reviewer N3G7,

Thank you for your timely comments! We will address your question below:

Q1. STD of selected block quantities

We apologize for not providing a detailed implementation of the STD calculation. We used 500 prompts and recorded the attention head block selection values for each attention forward pass, then averaged these values. The specific formula is as follows:

For the importance mask $\mathbf{B}^{h}_{\text{top}} \in \mathbb{R}^{M \times M}$ (where $h$ denotes the attention head and $M$ is the number of blocks, $M=900$ for one 720P HunyuanVideo inference), we compute the standard deviation across attention heads:

$\text{STD} = \sqrt{\frac{1}{H} \sum_{h=1}^{H} ||\mathbf{B}^{h}_{\text{top}} - \mu||_F^2}$

where $\mu = \frac{1}{H} \sum_{h=1}^{H} \mathbf{B}^{h}_{\text{top}}$ is the mean block selection mask across all heads, and $||\cdot||_F$ denotes the Frobenius norm.

The final reported STD is averaged across all timesteps $T$, layers $L$, and the number of sampled prompts throughout the generation process.

This metric measures the diversity and consistency of different attention heads in their block selection strategies. Thus, it serves as a reference for block selection balance.

Dear reviewer G2mf,

Thank you for your feedback and comments. We appreciate the thoroughness with which you have reviewed our manuscript, as evidenced by your insightful and valuable comments. We are grateful for your positive assessment of our work, particularly your recognition of its novelty and results. We will address your comments below and incorporate the necessary revisions in the updated paper.

Q1 Clarification of Block-Level Attention

We address your concerns by breaking them down into the following specific points:

Q1.1 How frequently the block-wise attention mask is updated?

The table below provides detailed information about the number of updates and computation timing for each mask type in an $L$-layer DiT model with $T$ actual denoising timesteps across $S$ resolution stages in ProRes.

Attention Carving directly replaces the normal Full Attention. This means the head-block-wise attention mask $\mathbf{B}$ is updated in per-layer attention. If we perform inference on a 50-step model with 60 layers, it will update 50×60=3000 times. This design choice is motivated by the fact that attention patterns may vary significantly across different heads, layers and timesteps, as visualized in Appendix Fig. 13. We have experimented with reusing masks across different timesteps, but this approach failed to generate satisfying video results (VBench < 80%). We will include this explanation in the revision for better clarification. Additionally, while the block-selection operation (determining the importance mask $\mathbf{B}_\text{top}$) is performed layer-wise, we demonstrate in Appendix C.4 and Fig. 15 that the computational overhead of block selection is minimal (less than 3% of the total sparse attention computation).

Regarding the Condition Mask and Adjacency Mask, as stated in L162-163,

In Jenga, Condition Mask and Adjacency Mask are pre-computed, and only determined by the resolution and the partition function $\mathcal{G}$.

These two masks are static and precomputed as the resolution-aware SFC ($\mathcal{G}$) is predetermined. The only updates occur during stage switches when the latent resolution changes.

Q1.2 Which Mask Contributes Most to Performance?

In Table 3(b) and Figure 7, we provided an ablation study on SFC selection and mask selection, demonstrating the effectiveness of our mask selection scheme. Here we present a more comprehensive ablation study to provide better validation of our method. The "No Mask" baseline is the result with ProRes and timestep skip.

The above results demonstrate that: (1) Each mask component contributes meaningfully to quality or efficiency, (2) The importance mask provides the most significant impact on both metrics, (3) The combination of all three masks achieves the optimal balance between visual quality and computational efficiency. The "w/o Importance" case shows the most dramatic speedup but significant quality degradation, confirming that content-aware adaptive block selection is crucial for maintaining generation quality.

Q2 Connection Between Trajectory Curvature and Resolution Scheduling

The denoising trajectory serves as a guidance mechanism for optimal resolution scheduling. We conducted an ablation study on resolution scheduling to better understand the relationship between trajectory curvature and optimal transition timing:

• The experiment uses Jenga-Turbo's basic setting, containing one low-res > high-res stage transition.

The experimental results reveal key insights about trajectory-guided resolution scheduling. Due to the timestep shift mechanism in DiT, more computational steps are allocated to the early stages of the trajectory. Even in the final 10% of timesteps, the actual trajectory still contains approximately 43.7% of the total steps.

We observe a dramatic quality degradation when low-res steps exceed 50% of the total timesteps, while introducing latent misalignment artifacts, confirming that multi-stage generation requires sufficient high-resolution denoising to maintain video quality. Meanwhile, when stage 1 (low-res) steps are too few, quality also decreases. This aligns with our motivation (L48-49) that low-res content generation requires a higher attention density to establish proper content structure. This finding validates why ProRes and AttenCarve work synergistically--they are not merely independent methods but complementary components that address different phases of the generation process.

The 50% low-res step configuration achieves the best balance between quality (83.07%) and efficiency (225s), demonstrating that trajectory curvature analysis effectively guides optimal resolution scheduling decisions.

Q3 Discussion on Consistency and Low-Step Models

We may not have clearly explained that AccVideo is actually a low-step model (5 NFE) derived from HunyuanVideo. In Table 2(b), we adopt Jenga on AccVideo with single and multi-GPU integration, achieving a 2.12× acceleration with only AttenCarve with comparable VBench and CLIPScore performance. Our method emphasizes the orthogonality with step-reduction (distilled model like AccVideo) and feature reuse (TeaCache) without affecting generation quality.

Meanwhile, based on our discussion in Q2 regarding ProRes and trajectory analysis, as resolution switching fundamentally depends only on the corresponding trajectory timestep, we believe few-step diffusion models can directly apply ProRes for further acceleration. The table above demonstrates that using a two stage ProRes can maintain quality while providing 3.16× speedup even in few-step scenarios. This reinforces the orthogonal nature of our approach--it can complement existing few-step models without interference.

Q4 Adaptivity of Sparsity Masks

As discussed in Q1, the importance mask is updated adaptively in each attention forward pass, with adaptation for each attention head in our implementation. This means the block selection operation is online and adaptive.

Although SFC is pre-determined by resolution, it has several benefits (a) adaptive resolution capability and (b) inherent local nature (which motivates our "Jenga" analogy). For a more detailed discussion on static or adaptive block partition, please refer to response in cBBb Q2.

Although $k$ and threshold $p$ are pre-defined, the block selection for each query block remains adaptive across different attention heads based on the current content. As Jenga achieved a higher Dynamic Degree score in VBench in Appendix Table 6, we may not directly treat the high sparsity to the degradation in generating highly dynamic contents.

As a pure training-free method, current version avoids the need for offline per-head optimal search procedures (a fitted static sparsification, but not adaptive with input feature) that methods like STA[1] require. Jenga's significantly smaller search space makes it easily adaptable to new models without extensive hyperparameter tuning. It would be a future direction to design a metric to determine adaptive or learnable sparsity.

Q5 Novelty is Incremental?

We acknowledge that there are previous works in sparse attention and multi-resolution sampling and discussed them in L92-93, L98-101. However, our contribution represents an advance beyond existing approaches:

Fundamental Motivation Difference: Jenga's core motivation differs fundamentally from previous work. We treat resolution scaling and attention carving as a unified framework, both targeting token interaction reduction. The central challenge is balancing large attention sparsity with content generation quality. Our unified approach addresses this systematically.

Comparison with SparseAttn: SparseAttn achieves only moderate sparsity (0.5) with extra tuning, compared to Jenga's high sparsity (0.75). This substantial improvement in sparsity-quality trade-off represents a meaningful technical contribution. AttenCarve's higher sparsity is achieved by the introduction of ProRes, which pre-generates (noisy) content in a low-res space.

Comparison with Bottleneck Sampling: The BottleNeck design requires additional computation in the first stage and achieves only sub-optimal speedup ratios, even with some timestep reduction. Our goal is to achieve extreme speedup ratios through systematic optimization.

Technical Novelty: Our method introduces several technical innovations: (1) 3D SFC-based block partitioning and adaptive block selection with hybrid masks, (2) Text-attention amplifier that addresses FOV problems in training-free low-res generation. Most importantly, ProRes solves the content degradation when adopting high sparsity attention across all timesteps, maintaining both efficiency and quality throughout the generation process. While technical directions may share similarities, we believe Jenga represents a technically solid contribution that combines strong practical utility in video generation.

[1] P. Zhang .et.al., Fast Video Generation with Sliding Tile Attention, 2025