Plug-and-Play Context Feature Reuse for Efficient Masked Generation

We thank the reviewer for the thoughtful and constructive feedback. We appreciate your recognition of the reasonableness of our method and the quality of our presentations. To address your concerns, we have conducted additional experiments and analyses. These include comprehensive ablations over key hyperparameters, an evaluation on a 512×512 text-to-image (T2I) task, and a comparison and discussion of learnable guidance methods such as [1]. These additions further justify the robustness and general applicability of ReCAP across various generation settings.

Reporting performance variations based on different hyperparameter settings would enhance the paper.

It would be valuable to analyze whether reducing the total sampling timestep is more beneficial than setting l to a value greater than 1.

We thank the reviewer for their valuable feedback. While the choice of Full-FE ($T$) and Local-FE ($T'$) steps is central to our method, we find that ReCAP’s performance is robust across a wide range of these hyperparameters. Even with more efficient caching (e.g., $l=3$), ReCAP outperforms naive step reduction.

Specifically, $u$ denotes the point when Local-FE begins to be inserted, and $l$ controls the number of Local-FE steps inserted per grouped decoding step. These implicitly determine $T$ and $T'$, with $T = u + \frac{S - u}{l+1}$ and $T' = S - T$ ($S$ is the total sampling step). In the main paper, we set $u = \frac{T+T'}{2}$ and $l = 1$ for simplicity. Below, we expand our analysis by varying both $u$ and $l$ （Inference time is measured using an NVIDIA RTX 4090 GPU with a batch size of 32）:

MaskGIT-r (Baseline):

MaskGIT-r+ReCAP:

Conclusion1: ReCAP consistently improves quality-efficiency trade-offs over the baseline across different $u$, demonstrating robustness. Smaller $u$ enables better efficiency but may lead to greater quality degradation, as early context features are relatively unstable. As $u$ increases, cached features become more stable and improvements become more significant.

1. $l_{u:u+T'}=1$

Conclusion2: Increasing $l$ further enhances efficiency while maintaining quality. Even with aggressive caching (e.g., $l=3$), ReCAP outperforms naive step reduction. However, larger $l$ increases sensitivity to $u$, requiring careful scheduling to avoid excessive cache degradation.

2. $l_{u:u+\frac{T'}{2}}=2$

3. $l_{u:u+\frac{T'}{3}}=3$

ReCAP requires Full-FE for more than half of the total steps.

We thank the reviewer for the insightful observation. While our default setting in the main paper uses more Full-FE than Local-FE steps, this design already leads to noticeable improvements, demonstrating the practicality of ReCAP even under a balanced schedule. In the following, we further explore more adaptive and efficient caching strategies and demonstrate that the number of Full-FE steps can be reduced to less than half of the total steps, while still achieving strong generation quality and even better quality-efficiency trade-offs.

Specifically, motivated by our earlier ablations that features in later decoding stages are more stable and cacheable, we implement a dynamic cache schedule in MAGE+ReCAP for unconditional generation. We set $u = 0$ and adaptively increase $l$ from 1 to 2 when the number of remaining masked tokens falls below 128 (i.e., half of the 256-token sequence). Below we compare baseline, static ReCAP ($l=1$), and hybrid ReCAP ($l\in\{1,2\}$):

These results confirm that:

• ReCAP with adaptive scheduling achieves comparable or better FID with lower latency.
• For example, at 80 steps total, hybrid ReCAP reaches 7.38 FID in 0.31s, outperforming the base model (8.04 FID,0.31s) and the static ReCAP (7.38 FID, 0.34s).
• As steps increase, the efficiency benefit scales, showing potential for even larger gains with longer sequences.

This demonstrates that ReCAP’s efficiency gain is not inherently bounded by needing many Full-FE steps — instead, careful scheduling (e.g., hybrid $l$) enables fewer Full-FE steps with strong performance, highlighting ReCAP’s flexibility and future potential.

A larger resolution (512) or Text-to-Image (T2I) model (e.g., [2]) will make the paper more complete.

We thank the reviewer for the constructive suggestion. We have incorporated additional experiments by applying ReCAP to the Harmon-0.5B T2I model from [2] at 512×512 resolution on the MJHQ-30k benchmark.

We first reproduce the baseline results using the official implementation (inference time measured on NVIDIA A800 with batch size = 50 and the latency of the denoising MLP module is excluded):

Baseline (Harmon-0.5B T2I, 512×512):

Harmon-0.5B+ReCAP:

ReCAP consistently improves the quality-efficiency trade-off of the base T2I model:

• Compared to the 64-step baseline (FID 6.461, 0.657s), ReCAP (32 + 32) achieves similar FID (6.467) with nearly 2× faster inference (0.347s).
• With the same number of Full-FE steps (T = 48), ReCAP (48 + 16) yields better FID (6.463 vs. 6.481) and is faster (0.453s vs. 0.490s) than the 48-step baseline. This efficiency gain is likely due to the Harmon model incorporating a causal LLM: although the number of Full-FE steps is the same, the input token lengths to the LLM differ, leading to reduced computation in ReCAP.
• Moreover, in the first three rows of the second table, larger T' under the same T consistently leads to better FID with only marginal latency increase, showcasing the strength of aggressive caching (larger $l$).

These results further validate the generality of ReCAP on high-resolution and multimodal T2I models.

I wonder if ReCAP can be integrated with the approach in [1], or if not, the authors should provide a comparison between these methods.

We thank the reviewer for highlighting recent advances in learnable guidance methods such as [1]. Importantly, the two methods are orthogonal in both design and objective: [1] improves quality by enforcing semantic consistency through a learned smoothing encoder. In contrast, ReCAP reuses cached context features to construct low-cost decoding steps, efficiently approximating the effect of multiple full-feature evaluations while maintaining consistency.

Moreover, ReCAP is fully compatible with the approach in [1]. Combining the two is straightforward—ReCAP’s caching mechanism can be applied to both the base model and the TOAST module in [1] to further improve the quality-efficiency trade-off.

We also note a key distinction: [1] requires additional training on top of the base model to learn the guidance network, whereas ReCAP is entirely training-free and can be directly applied to any pretrained MGM. We will include a discussion of [1] in the next version of the paper.

[1] Hur, J., Lee, D., Han, G., Choi, J., Jeon, Y., & Kim, J. (2024). Unlocking the Capabilities of Masked Generative Models for Image Synthesis via Self-Guidance. Advances in Neural Information Processing Systems, 37, 130977-130999.

[2] Wu, S., Zhang, W., Xu, L., Jin, S., Wu, Z., Tao, Q., ... & Loy, C. C. (2025). Harmonizing visual representations for unified multimodal understanding and generation. arXiv preprint arXiv:2503.21979.

Thank you for your positive feedback. We appreciate your recognition of the paper’s clarity, the effectiveness of the proposed approach, and the breadth of experimental validation across multiple models.

We sincerely appreciate the reviewer’s insightful and constructive feedback. Thank you for recognizing the strengths of our work. We hope the following responses will fully address your concerns and further clarify our contributions. Specifically, we have conducted comprehensive ablations under various settings, expanded our evaluation to include a text-to-image (T2I) generation task, and discussed related work such as TeaCache and Flux. These additions further justify the robustness, practical applicability, and future potential of ReCAP.

The results in Table 1 show a notable performance drop for MaskGIT when reducing the number of steps to 16, with the FID score worsening from 4.46 to 5.02.

We thank the reviewer for the careful observation. We would like to clarify that the result reported in Table 1 of the main paper with (FID 5.02, 0.032s) at 16 steps corresponds to ReCAP with $u = 0$ and still largely outperforms the base model (FID 5.49 with 0.042s with 12 steps) in quality-speed trade-offs. Specifically, the parameter $u$ denotes the point at which Local-FE begins to be inserted, offering a flexible trade-off between quality and efficiency. In the following, we provide ReCAP performance under varying $u$ and demonstrate that the effectiveness of ReCAP is robust to various configurations.

MaskGIT-r (Baseline):

MaskGIT-r+ReCAP:

As shown in the table, ReCAP consistently improves quality-efficiency trade-offs over the baseline across different $u$, demonstrating robustness. Smaller $u$ enables better efficiency but may lead to greater quality degradation, as early context features are relatively unstable. As $u$ increases, cached features become more stable and improvements become more significant.

It would significantly strengthen the work to demonstrate its effectiveness in the T2I setting.

We thank the reviewer for the constructive suggestion. We have incorporated additional experiments by applying ReCAP to the Harmon-0.5B T2I model from [1] at 512×512 resolution on the MJHQ-30k benchmark.

We first reproduce the baseline results using the official implementation (inference time measured on NVIDIA A800 with batch size = 50). To ensure fair comparison, we exclude the latency of the denoising MLP module:

Baseline (Harmon-0.5B T2I, 512×512):

Harmon-0.5B+ReCAP (T = #Full FE, T' = #Local FE):

ReCAP consistently improves the quality-efficiency trade-off of the base T2I model:

• Compared to the 64-step baseline (FID 6.461, 0.657s), ReCAP (32 + 32) achieves similar FID (6.467) with nearly 2× faster inference (0.347s).
• With the same number of Full-FE steps (T = 48), ReCAP (48 + 16) yields better FID (6.463 vs. 6.481) and is faster (0.453s vs. 0.490s) than the 48-step baseline. This efficiency gain is likely due to the Harmon model incorporating a causal LLM: although the number of Full-FE steps is the same, the input token lengths to the LLM differ, leading to reduced computation in ReCAP.
• Moreover, in the first three rows, larger T' under the same T consistently leads to better FID with only marginal latency increase, showcasing the strength of aggressive caching (larger $l$).

These results further validate the generality of ReCAP on high-resolution and multimodal T2I models.

[1] Wu, S., Zhang, W., Xu, L., Jin, S., Wu, Z., Tao, Q., ... & Loy, C. C. (2025). Harmonizing visual representations for unified multimodal understanding and generation. arXiv preprint arXiv:2503.21979.

The core idea bears similarities to TeaCache

We thank the reviewer for pointing out this connection. TeaCache is a representative work on caching for continuous diffusion models, which also leverages temporal similarities between high-level features for acceleration. Our work targets token-based masked generative models (MGMs) / masked diffusion models (MDMs), where the decoding dynamics and caching strategies differ. We will include a discussion of TeaCache in the next version of the paper.

ReCAP is not very competitive compared to one-step generators like Flux. Furthermore, MGMs currently lag behind diffusion models, which limits the practical applicability of this approach.

We appreciate the reviewer bringing up Flux as example of one-step generators that offer fast image generation. However, Flux is fundamentally based on rectified-flow transformer architectures with distillation-based guidance, which represents a distinct paradigm from token-based masked generative models (MGMs). Flux requires specialized architecture design and supervised distillation training. In contrast, ReCAP is a training-free, plug-and-play module specifically designed for token-based MGMs, where generation naturally involves multiple iterative decoding steps.

Moreover, we respectfully note that the generation quality of MGMs has been rapidly improving. Recent models such as MAR have significantly closed the gap with state-of-the-art continuous diffusion models employing advanced sampling strategies. In addition, MGMs enjoy efficiency advantages over traditional diffusion models. More importantly, MGMs, as sequence modeling frameworks, have the potential to unify language and image generation under a single architecture. This makes them particularly well-suited for multimodal large language models (MLLMs). Notably, more recent work such as Mmada [2] demonstrates that MGMs can serve as a strong backbone for large-scale multimodal generative models, further highlighting the practical applicability and future potential of ReCAP.

[2] Yang, L., Tian, Y., Li, B., Zhang, X., Shen, K., Tong, Y., & Wang, M. (2025). Mmada: Multimodal large diffusion language models. arXiv preprint arXiv:2505.15809.

Additionally, the acceleration on MAR-H appears insufficient and may not justify the trade-off in performance.

We thank the reviewer for the comment. We would like to highlight that ReCAP achieves a 2.4× speedup over the state-of-the-art FID of MAR-H baseline, with minimal performance drop. Moreover, as shown in Figure 5 of the main paper, MAR-H + ReCAP matches the performance of REPA [3], even though REPA is further equipped with advanced guidance interval sampling strategies. Notably, REPA is a state-of-the-art flow-matching and represents a very strong baseline, as it incorporates self-supervised features from powerful pretrained vision foundation models, resulting in rich semantic priors and high-quality generation. In contrast, ReCAP is entirely training-free and does not rely on any external guidance—yet still achieves comparable performance. This highlights the effectiveness and applicability of ReCAP.

[3] Yu S, Kwak S, Jang H, et al. Representation alignment for generation: Training diffusion transformers is easier than you think[J]. arXiv preprint arXiv:2410.06940, 2024.

We thank the reviewer for the thoughtful and constructive feedback, and for recognizing the motivation, design, and empirical validation of ReCAP to be sound and broadly applicable. To further address the reviewer’s concerns regarding the sensitivity and generality of ReCAP’s hyperparameters, we have conducted additional ablation studies and implemented a more dynamic and efficient caching schedule, as well as a new evaluation on a text-to-image (T2I) generation task. These results offer deeper insight into the behavior of ReCAP and demonstrate its robustness across a range of settings.

How does the efficiency-quality trade-off evolve under different allocations of Full-FE (T) and Local-FE (T′) steps? In particular, what happens when ReCAP operates under more aggressive schedules?

We thank the reviewer for their valuable feedback. While the choice of Full-FE ($T$) and Local-FE ($T'$) steps is central to our method, we find that ReCAP’s performance is robust across a wide range of these hyperparameters. Even with more aggressive schedules (e.g., $l=3$), ReCAP outperforms naive step reduction. Below, we conduct a comprehensive evaluation across different allocations of $T$ and $T'$, examining their impact on quality-efficiency trade-offs.

Specifically, $u$ denotes the point when Local-FE begins to be inserted, and $l$ controls the number of Local-FE steps inserted per grouped decoding step. These implicitly determine $T$ and $T'$, with $T = u + \frac{S - u}{l+1}$ and $T' = S - T$ ($S$ is the total sampling step). The following tables report the performance with varying $u$ and $l$ (All reported inference time is measured using a single NVIDIA RTX 4090 GPU with a batch size of 32):

MaskGIT-r (Baseline):

MaskGIT-r+ReCAP:

Conclusion1: ReCAP consistently improves quality-efficiency trade-offs over the baseline across different $u$, demonstrating robustness. Smaller $u$ enables better efficiency but may lead to greater quality degradation, as early context features are relatively unstable. As $u$ increases, cached features become more stable and improvements become more significant.

1. $l_{u:u+T'}=1$

Conclusion2: Increasing $l$ further enhances efficiency while maintaining quality. Even with aggressive caching (e.g., $l=3$), ReCAP outperforms naive step reduction. However, larger $l$ increases sensitivity to $u$, requiring careful scheduling to avoid excessive cache degradation.

2. $l_{u:u+\frac{T'}{2}}=2$

3. $l_{u:u+\frac{T'}{3}}=3$

While I understand that the authors chose simple, generalizable settings across models, is there substantial headroom left if hyperparameters are finely tuned for each architecture? Also, is the behavior of performance vs. hyperparameter choice consistent across different models?

We appreciate the reviewer’s insightful question. The overall behavior of performance vs. hyperparameter choice is largely consistent across models. For instance, a larger $u$—which delays the use of Local-FE—typically leads to more stable caching and better quality. Based on this observation, we adopt simple and generalizable hyperparameter settings (e.g., $u = \frac{T + T'}{2}$, $l = 1$) in our main results and demonstrate the broad applicability and robustness of ReCAP across different MGMs.

Nevertheless, the sensitivity to these hyperparameters does vary by model. Some models, such as MAGE, are more robust to early caching with smaller $u$ without significant quality degradation. Others, such as MaskGIT and MAR, benefit more from conservative schedules with larger $u$. Therefore, further performance gains can be achieved by fine-tuning these hyperparameters for each architecture. For instance, in the main paper, we adopt a more aggressive configuration ($u = 0$, $l = 1$) for MAGE to maximize efficiency.

To explore this further, we include two additional experiments:

(i) Dynamic caching strategy for MAGE+ReCAP: We implement a adaptive cache schedule with $u = 0$, and adaptively increase $l$ from 1 to 2 when the number of remaining masked tokens falls below 128 (i.e., half of the 256-token sequence). This is motivated by the observation that features in later decoding stages are more stable and thus more cacheable.

These results show that ReCAP with adaptive scheduling achieves comparable or better FID than fixed schedules with better efficiency. For example, at 80 steps total, hybrid ReCAP reaches 7.38 FID in 0.31s, outperforming the base model (8.04 FID,0.31s) and the static ReCAP (7.38 FID, 0.34s), highlighting ReCAP’s future potential with more dynamic and principled schedules.

(ii) ReCAP for a T2I model (Harmon-0.5B [1]) under different hyperparameters: We evaluate ReCAP in a 512×512 text-to-image setting using Harmon-0.5B on MJHQ30K.

Baseline (Harmon-0.5B T2I, 512×512):

Harmon-0.5B+ReCAP (T = #Full FE, T' = #Local FE):

These results indicate that a larger $l$ (e.g., $l = 2$) is particularly beneficial in this T2I setting, offering stronger acceleration without degrading FID. For example, ReCAP at (16, 2) achieves FID 6.467 in 0.347s, matching the 64-step baseline with ~2× lower latency.

Together, these results demonstrate that while a general hyperparameter configuration works well across models, ReCAP allows for further tuning and adaptation, and holds great promise when combined with more dynamic caching strategies tailored to specific architectures or tasks.

[1] Wu, S., Zhang, W., Xu, L., Jin, S., Wu, Z., Tao, Q., ... & Loy, C. C. (2025). Harmonizing visual representations for unified multimodal understanding and generation. arXiv preprint arXiv:2503.21979.

If there is a reason why, between two not Local-FE steps, the previously unmasked tokens KV are not added to the cache but appear to be recomputed in Figure 3.

In our design, each Full-FE aims to recompute precise and up-to-date features for all tokens. Because MGMs rely on bidirectional attention, the features of previously unmasked tokens can change across steps as more context is revealed.