MagCache: Fast Video Generation with Magnitude-Aware Cache

Thanks for the valuable feedback. MagCache is a unified, plug-and-play acceleration method compatible with a wide range of video generation models, including Wan 2.1, Open-Sora, HunyuanVideo, CogVideoX, few-step distilled models (FusionX), and low-precision models. It supports ComfyUI interface for ease of use.

---

Q1: Compatibility with other acceleration methods: Can MagCache be integrated with other diffusion model acceleration techniques, such as model distillation and low-precision arithmetic?

A1: Thanks for the insightful question. MagCache is compatible with other acceleration techniques, including model distillation and low-precision arithmetic.

Although distilled models tend to exhibit larger variations between denoising steps, MagCache remains effective even for few-step distilled models by caching and reusing intermediate computations with minimal differences. As shown in Table R4.1, we evaluated MagCache on FusionX (a 10-step distilled version of Wan2.1 14B) and achieved a 1.66× speedup without significant degradation in visual quality. Since distillation has reduced redundancy between steps, the speedup ratio of MagCache (1.66×) on the distilled model (FusionX) is lower than that on the non-distilled version (Wan 2.1). Notably, simply reducing the inference steps to T=6 yields much worse performance than using MagCache with T=10, even at the same speedup of 1.66×.

We also evaluated MagCache under low-precision settings. As shown in Table R4.2, MagCache is compatible with 4-bit quantized models, maintaining acceleration benefits with only a minor drop in speedup due to the reduced numerical precision. Both the “fast” and “slow” variants maintain strong visual quality when used with low-precision models.

---

Table R4.1: Evaluation of MagCache on Few-Step Distilled Model (Wan 2.1 14B FusionX, 33 frames, 480P)

---

Table R4.2: Evaluation of MagCache on Low-precision Arithmetic Model (Wan 2.1 14B, 33 frames, 480P)

Dear Reviewer YTog,

Thank you sincerely for your valuable time to review our paper and providing such insightful feedback. We deeply appreciate the thought and effort you have invested in this process. We wanted to update you that the reviewer AVyF and 9jpp , who initially assigned scores of 3, have not yet joined any discussion. Given this, we would be extremely grateful if you might be willing to reconsider our work with a higher score.

Any slight adjustment to your score, or your continued support when engaging in discussions with the Area Chair, would mean a great deal to us as we strive to address feedback and strengthen our submission.

Thank you again for your support and consideration.

Sincerely,

Authors of Paper 16663

Q1: Insufficient motivation and generality. While MagCache aims to accelerate inference by exploiting motion-aware magnitude patterns, the empirical motivation is not fully convincing. The current analysis on two video models and three prompts is limited in scope.

A1: Thanks for the insightful feedback. Our core motivation is to accelerate video generation by exploiting redundancy in residuals across adjacent steps. Emperically, the residual difference is small in the first 80% steps across diverse models and prompts. Notably, MagCache is step-aware, not motion-aware. Both motion and semantics evolve minimally between early adjacent steps. Besides, we extend experiments to include more models (Table R3.1) and a broader set of prompts (Table R3.2 & R3.3) to demonstrate the effectiveness and generality of MagCache.

Q1.1: Model diversity is limited to support the plug-and-play claim.

A1.1: To strengthen the generality, we integrate MagCache to HunyuanVideo, CogVideoX, and Flux (Table R3.1), observing competitive or superior performance across these models. We also qualitatively validated MagCache on additional models (FramePack, VACE, Flux Kontext, Chroma, and OmniGen2), and will release supporting code and demos. In summary, MagCache is a unified, plug-and-play acceleration method applicable to a wide range of diffusion models.

Table R3.1 Generalization of MagCache across Diverse Diffusion Models.

---

Q1.2: Figure 1 uses only three prompts, insufficient to substantiate the prompt-invariance.

A1.2: To assess prompt invariance, we analyzed magnitude ratios over 944 VBench prompts, grouped into five 20% diffusion intervals. As shown in Table R3.2, the magnitude ratios exhibit low standard deviation and a consistent decreasing trend, with a sharp drop in the final 20%. Cosine distances follow a similar pattern, aligning with Figure 1.

Further, Table R3.3 demonstrates that the choice of calibration prompt has negligible effect. Whether using a random prompt, the full prompt set, or an outlier, performance remains stable. In summary, magnitude ratios are consistent across prompts, and MagCache is robust to prompt choice.

Table R3.2 Statistics of Magnitude Ratios and Cosine Distance across 944 Vbench prompts. 0.9965 (± 0.0004) denotes mean ± standard deviation

Table R3.3 Effect of Calibration Prompt.

---

Q2: Lack of theoretical insight into Equation (5).

A2: Thanks for the valuable comment. Equation (5) states that when the cosine distance between adjacent residuals is small, the residual difference $||r\_t - r\_{t-1}||$ can be approximated by the magnitude difference $\left|||r\_t|| - ||r\_{t-1}||\right|$:

$$||r_t - r_{t-1}|| \approx \left|||r_t|| - ||r_{t-1}||\right|. \tag{5}$$

This holds when $r\_t$ and $r\_{t-1}$ are nearly colinear (i.e., $1 - \cos(r\_t, r\_{t-1}) \approx 0$). Empirically, this occurs in the first 80% of steps. Theoretically, we can analysize the cosine law:

$$||r_t - r_{t-1}||^2 = (||r_t|| - ||r_{t-1}||)^2 + 2||r_t||||r_{t-1}||(1 - \cos(r_t, r_{t-1})).$$

As $1 - \cos(r\_t, r\_{t-1}) \to 0$, the second term vanishes, yielding Eq. (5). Let $r(t) = v\_\theta(x\_t, t) - x\_t$. By the chain rule:

$$\frac{dr}{dt} = (\partial_x v_\theta - I) x'(t) + \partial_t v_\theta.$$

Assuming $||x'\_t|| \le M$, spatial lipschitz continuity $||\partial\_x v\_\theta|| \le L\_x$, and temporal lipschitz continuity $||\partial\_t v\_\theta|| \le L\_t$, we get

$$\left|\left|\frac{dr}{dt}\right|\right| \le (L_x + 1)M + L_t = B,$$

where $B$ aggregates the ODE’s stiffness ($M$) and model smoothness ($L_x, L_t$), which is usually within a small constant range. Integrating over step size $h$(small in early steps), we can get $||r_t - r_{t-1}|| \le Bh.$ Assuming $||r\_t||, ||r\_{t-1}|| \ge c > 0$, we get:

$$1 - \cos(r_t, r_{t-1}) = \frac{||r_t - r_{t-1}||^2 -(||r_t|| - ||r_{t-1}||)^2}{2||r_t||||r_{t-1}||}\le \frac{||r_t - r_{t-1}||^2}{2||r_t||||r_{t-1}||} \le \frac{B^2 h^2}{2c^2} = O(h^2).$$

Hence, $r\_t$ and $r\_{t-1}$ are nearly aligned when $h$ is small. In non-uniform schedules (e.g., shift=8), early steps use small $h$ (e.g., 0.001), making $1 - \cos(r\_t, r\_{t-1})$ negligible, validating Equation (5).

---

Q3: Parameter sensitivity and lack of consistency.

A3: The varying $\delta$ and $K$ settings across models were chosen to ensure fair comparisons at similar speedups. In practice, MagCache performs robustly with default parameters across different models, requiring only minor $\delta$ adjustments. We offer two predefined modes:

• Slow mode: $K = 2$, $\delta = 0.06$ (default)
• Fast mode: $K = 4$, $\delta = 0.06$ (default)

Here, $K$ controls speedup mode, and $\delta$ adjusts the quality–speed trade-off within the selected mode. In Table R1.1, varying $\delta$ offers fine control with just 1–2 adjustments (doubling or halving). We recommend the slow mode ($K=2, \delta=0.06$), which works well for various models. Our ComfyUI interface allows users to easily adjust these parameters in a few seconds.

Table R3.4 Parameter Configurations in Slow and Fast inference mode.

---

Q4: Figure 2 is difficult to interpret and lacks clarity in both visual design and explanation.

A4: We have revised Figure 2 for improved clarity. It now better illustrates how $K$ and $\delta$ jointly govern the adaptive caching and reusing, with annotations and a step-wise flow to highlight the caching logic. Due to the rebuttal policy, we will update the figure in the revised paper.

---

Q5: How does the type of video content (e.g., human, natural scene, dynamic textures) influence the the effectiveness of MagCache? How sensitive is MagCache to different prompt types?

A5: We group the 944 Vbench prompts into 5 types. Table R3.5 shows that prompt types have negligible effect on MagCache’s performance. Stable magnitude ratios across prompts (Table R3.2) and calibration robustness (Table R3.3) also support this conclusion.

Table R3.5 Robustness of MagCache across Different Prompt Types.

Dear Reviewer 9jpp,

Thank you for your valuable time in reviewing our manuscript. As the discussion phase is ending soon, we would greatly appreciate it if you can consider our response and supplementary experiments in your final assessment. We demonstrated the robustness and generality of MagCache across more models and prompts (Table R3.1-R3.3, R3.5), and included a theoretical analysis of Equation (5) in A2. Additionally, we have improved the clarity of Figure 2, and shown that the parameters of MagCache are consistent, robust, and easy to set (Table R3.4). We hope that our response and the supplementary experiments have satisfactorily addressed your concerns.

Thank you for your valuable time!

Sincerely,

Authors of Paper 16663

Dear Reviewer 9jpp,​

I hope this message finds you well. Please accept my sincere apologies for following up again; we understand your time is extremely valuable. However, I’m writing with urgency as tomorrow marks the final day of our paper’s rebuttal review period, and we haven’t yet received your feedback on our rebuttal.​

We’ve worked diligently to address all concerns from your initial review, hoping our rebuttal clarifies our work. With only one day left until the deadline, we’d be deeply grateful if you could spare a few moments to review it at your convenience. Your insights are critical to our paper’s evaluation at this stage, and we hope our rebuttal has adequately addressed the points you raised before the deadline closes.​

Thank you again for your time and consideration. We apologize for any inconvenience and look forward to your input.​

Sincerely,

Authors of Paper 16663

Q1: Comparison with two recent state-of-the-art diffusion acceleration methods: TaylorSeer[1] and DuCa[2]

A1: Thanks for the valuable suggestion. We have added detailed comparisons with TaylorSeer [1] and DuCa [2] across three representative models: Open-Sora, Wan2.1 1.3B, and HunyuanVideo in Table R2.1. A comparison will be included in the revised paper.

Compared with DuCa on Open-Sora and HunyuanVideo, MagCache achieves better visual quality at similar speedups, while requiring only 0.1 GB of extra memory—significantly less than DuCa’s 23.9 GB (Open-Sora) and 14.7 GB (HunyuanVideo). Against TaylorSeer, MagCache delivers better visual quality on both Wan2.1 1.3B and HunyuanVideo at similar speedups, with much lower memory overhead (0.1GB-0.5 GB vs. 40–46 GB).

It is worth noting that TaylorSeer and DuCa cache outputs of all transformer layers, while MagCache only caches a single residual output, greatly reducing memory usage. In summary, MagCache consistently outperforms TaylorSeer and DuCa in both visual quality and memory efficiency.

Table R2.1: Comparison with Other Cache-based Methods across Various Models.

---

Q2: The ~2× speedup is solid but not as high as some recent approaches (e.g., TaylorSeer ~5×).

A2: Thanks for the comment. As shown in Table R2.1, at similar speedup levels (~5×), MagCache can achieve a superior performance than TaylorSeer on HunyuanVideo with significantly less extra memory cost. MagCache with ~2x speedup targets a balanced trade-off between speed and quality, allowing users to enjoy acceleration without sacrificing lots of quality. If users desire higher speedups, MagCache can be configured with larger $K$ and $\delta$ values to further accelerate inference.

---

Q3: The experiments are conducted on two large text-to-video diffusion models (Open-Sora and Wan), but broader validation is somewhat limited. The method’s generality would be better established by testing on more diverse setups (e.g., CogVideoX).

A3: Thanks for valuable suggestions. In the Table 5 of supplementary material, we have included a comparison on HunyuanVideo (540P) and Flux. To strengthen the generality of our approach, we futher expand our evaluation to include HunyuanVideo (480P & 540P), CogVideoX, and Flux. As shown in Table R2.2, MagCache consistently delivers strong performance across all various models and resolutions, demonstrating its robustness and generality.

Moreover, we have also qualitatively validated MagCache on additional models such as FramePack, VACE, Flux Kontext, Chroma, and OmniGen2, and will release demos and code in our open-source repository.

Table R2.2 Generalization of MagCache across Diverse Models and Resolutions.

Dear Reviewer AVyF,

Thank you for your valuable time in reviewing our manuscript. As the discussion phase is ending soon, we would greatly appreciate it if you could consider our response and supplementary experiments in your final assessment. MagCache outperforms both TaylorSeer and DuCa in quality, speed, and memory efficiency (Table R2.1), and we validate its effectiveness on more video/image generation models, including CogVideoX, HunyuanVideo, and Flux (Table R2.2). We hope that our response and the supplementary experiments have satisfactorily addressed your concerns.

Thank you for your valuable time!

Sincerely,

Authors of Paper 16663

Dear Reviewer AVyF,​

I hope this message finds you well. Please accept my sincere apologies for following up again; we understand your time is extremely valuable. However, I’m writing with urgency as tomorrow marks the final day of our paper’s rebuttal review period, and we haven’t yet received your feedback on our rebuttal.​

We’ve worked diligently to address all concerns from your initial review, hoping our rebuttal clarifies our work. With only one day left until the deadline, we’d be deeply grateful if you could spare a few moments to review it at your convenience. Your insights are critical to our paper’s evaluation at this stage, and we hope our rebuttal has adequately addressed the points you raised before the deadline closes.​

Thank you again for your time and consideration. We apologize for any inconvenience and look forward to your input.​

Sincerely,

Authors of Paper 16663

Q1: The performance relies on the tuning of two key hyperparameters $\delta$ and $K$ for each model.

A1: Thanks for the question. The varying $\delta$ and $K$ settings across models were chosen to ensure fair comparisons at similar speedups. In practice, MagCache performs robustly with default parameters across different models, requiring only minor $\delta$ adjustments. We offer two predefined modes:

• Slow mode: $K = 2$, $\delta = 0.06$ (default)
• Fast mode: $K = 4$, $\delta = 0.06$ (default)

Here, $K$ controls speedup mode, and $\delta$ adjusts the quality–speed trade-off within the selected mode. In Table R1.1, varying $\delta$ offers fine control with just 1–2 adjustments. We recommend the default slow mode for general use, and our ComfyUI interface makes parameter tuning fast and user-friendly.

Table R1.1 Parameter Configurations in Slow and Fast inference mode.

---

Q2: MagCache’s may not be useful for few-step diffusion models.

A2: Despite larger residual variations in few-step models, MagCache still provides acceleration by reusing residuals with small differences. On FusionX (a 10-step distilled Wan2.1 14B), MagCache achieves a 1.66× speedup with minimal quality loss(Table R1.2). In contrast, directly reducing the steps to 6 significantly degrades quality, highlighting the advantage of MagCache.

Table R1.2: Evaluation of MagCache on Few-Step Distilled Model (Wan 2.1 14B FusionX, 33 frames, 480P)

---

Q3: Comparison with other caching-based acceleration methods, such as AdaCache and FasterCache.

A3: Thanks for the valuable suggestion. We compare MagCache with AdaCache, FasterCache, DuCa, and TaylorSeer across Open-Sora, Wan2.1, and Hunyuan in Table R1.3. On Open-Sora, MagCache matches AdaCache’s speedup but outperforms it in quality with significantly less extra memory (0.1G vs 11.3G). It also surpasses FasterCache in performance, speed, and memory efficiency (0.1 G vs. 11.3 G) on Open-Sora. This is because MagCache caches only a single residual instead of layer-wise outputs like AdaCache and FasterCache. Additionally, MagCache consistently achieves superior performance on both the Wan and Hunyuan models. Across all models, MagCache offers a strong balance of speed, quality, and memory usage.

Table R1.3: Comparison with Other Cache-based Methods across Various Models.

---

Q4: Sensitivity to the calibration sample. Does prompt selection impact the quality? Would an outlier or particularly complex prompt result in a less representative curve?

A4: MagCache is robust to the selection of the calibration sample, and generation quality remains consistent even with outliers. We first evaluate 944 prompts from VBench and observe very low variance in magnitude ratios and cosine distances(Table R1.4). The magnitude ratios exhibit a stable, monotonically decreasing trend, with sharper drops in the final 20% steps. The cosine distance trends are also stable across prompts.

We also test MagCache with three calibration prompt: a random prompt, all 944 prompts, and an outlier prompt (farthest from the average curve). All three yield similar speedup and visual quality (Table R1.5), confirming MagCache's robustness.

Table R1.4 Statistics of Magnitude Ratios and Cosine Distance across 944 Vbench prompts. 0.9965 (± 0.0004) denotes mean ± standard deviation

Table R1.5 Influence of Calibration Prompt.

---

Q5: Are the hyperparameters dependent on the denoising scheduler? Should hyperparameters be recalibrated when altering denoising scheduler or number of denoising steps?

A5: MagCache’s hyperparameters are robust across different schedulers and numbers of steps. As shown in Table R1.6 and Table R1.7, magnitude ratios calibrated with one scheduler or step setting can generalize well to others without recalibration. For instance, magnitude ratios calibrated under UniPC can be applied to inference using DPM++, and vice versa, with minimal impact on quality or speedup. When steps change, we use nearest-neighbor interpolation to align the ratio curve. Note that using the same inference scheduler or number of steps tends to yield consistent results.

Table R1.6 Robustness of Magnitude Ratios to Different Schedulers.

Table R1.7 Robustness of Magnitude Ratios to Different Numbers of Steps.

Dear Reviewer vpsp,

Thank you sincerely for your valuable time to review our paper and providing such insightful feedback. We deeply appreciate the thought and effort you have invested in this process. We wanted to update you that the reviewer AVyF and 9jpp , who initially assigned scores of 3, have not yet joined any discussion. Given this, we would be extremely grateful if you might be willing to reconsider our work with a higher score.

Any slight adjustment to your score, or your continued support when engaging in discussions with the Area Chair, would mean a great deal to us as we strive to address feedback and strengthen our submission.

Thank you again for your support and consideration.

Sincerely,

Authors of Paper 16663