We sincerely thank you for your positive evaluation. Your comments are deeply encouraging with valuable insights. If you have any further suggestions or would like to discuss our work in more detail, we would be delighted to engage in a conversation. Your feedback is invaluable to us. Detailed responses are provided in the following rebuttal.

Weakness 1: Discussion and comparison with more recent SOTA KD methods.

Thank you for highlighting these related works, which are all feature-based distillation methods. Below, we will discuss the difference between our MiPKD and these methods in terms of feature alignment, and provide the experimental comparison on SR task.

1. ViTKD aligns the student’s features with the teacher’s by using the FitNet loss for shallow layers and a generative loss for deeper layers. Its generative KD loss can be seen as a special case of our Feature Prior Mixer in MiPKD, without the auto-encoder structure or stochastic feature mixing.

2.  PEFD employs a projector ensemble to prevent overfitting to the teacher’s feature space, which learns the task-relevant discriminative feature. In contrast, MiPKD employs the feature mixture in the latent space by adding the part feature information as the prior, which ensures better feature learning for student.

3. KD-SRRL computes a Softmax Regression loss by feeding the student’s features to the teacher’s classification head for distillation In contrast, MiPKD leverages a block prior mixer to stochastically passing the enhanced feature maps for the alignment on the final outputs, which effectively reduces the capacity disparity between the teacher and student.

4. DMAE applies a feature-based KD loss to align the intermediate features between the pre-trained MAE teacher model and the student based on the masked image, and employs the decoder after the last student’s layer to reconstruct the original image. Similar to the FitNet approach, it forces direct feature alignment on the intermediate features of teacher and student. In contrast, MiPKD uses an auto-encoder like structure to enable feature prior mixture in the latent space. Moreover, MiPKD employs the random mask on the feature space, instead of the masking on the input images in DMAE. A more detailed comparison between the MiPKD and DMAE is below

   • The DMAE aligns intermediate features between teacher and student using a simple L1 loss, with lightweight MLPs employed to match their dimensions. In contrast, MiPKD adopts a Feature Prior Mixer to encode teacher and student feature maps into a unified latent space for mixing and reconstruction, further stabilized by an auxiliary auto-encoder loss $\mathcal{L}^{ae}\_k$.
   • While DMAE does not explicitly distill at the block level, MiPKD introduces a Block Prior Mixer, which dynamically switches between teacher and student blocks for coarse-grained distillation.
   • In terms of knowledge transfer, DMAE relies on masked image reconstruction to facilitate feature learning, whereas MiPKD mixes feature maps and blocks at multiple granularities, aligning models’ logits, features, and reconstruction outputs to address capacity disparity.
   • The DMAE is tightly integrated into the MAE framework, requiring the teacher to be pre-trained with masked image modeling. In contrast, MiPKD is universally compatible with various SR architectures and can be extended to diverse computer vision tasks.

Furthermore, we apply these methods to the SR tasks for the experimental comparison. As shown in the below table, our MipKD significantly outperforms all above feature-based distillation methods. For example, we outperform DMAE by 0.19db PSNR for x4 SR, when distilling RCAN model.

Weakness 2: Discussion with DMAE and RDEN.

The difference between our method and DMAE is discussed in the answer of Weakness 1. For RDEN, it introduces an efficient SR network design by leveraging re-parameterization and a progressive training strategy. Especially, the FitNet loss is used for distillation in the second training stage. As a general distillation framework, MiPKD can serve as a replacement for RDEN’s original distillation method, to potentially further enhance  student performance. In the revision, we have cited this paper and discussed it in the related work.

Weakness 3: Exploration on the optimal distillation position.

Thank you for raising this important point regarding the positioning of distillation components. We agree that understanding the optimal placement of the feature and block prior mixers can provide valuable insights into the flexibility and effectiveness of MiPKD. As suggested, we conduct additional ablation experiments to evaluate the impact of applying the FPM and BPM at different network positions, which is shown in the below table.

The results indicate that applying MiPKD uniformly across all layers yields the best performance compared to the specific single layer. To explain, the uniform strategy captures hierarchical feature interactions between the teacher and student models.

We sincerely thank you for your positive evaluation. Your comments are deeply encouraging with valuable insights. If you have any further suggestions or would like to discuss our work in more detail, we would be delighted to engage in a conversation. Your feedback is invaluable to us. Detailed responses are provided in the following rebuttal.

Weakness 1: Inconsistent notations and undefined terms

We sincerely thank the reviewer for pointing out these issues, which are essential for improving the clarity and consistency of our manuscript. In future version , we will carefully revise the notations, and add explicit description and reference for them in the relevant sections to avoid ambiguity.

Weakness 2: Experiments on SwinIR model

Thank you for pointing out the limited evaluation on SwinIR. We acknowledge that the coverage for this model is not as extensive as for EDSR and RCAN. The SwinIR model, being based on a Transformer architecture, poses unique challenges for knowledge distillation, as most existing SR KD methods are primarily designed for CNN-based networks. In the experiments, we applied two feature-based KD methods (namely, FitNet and FAKD) on SwinIR, and found that both of them lead to worse student model comparing to training from scratch, while the MiPKD improves SwinIR student model significantly.

Question 1: Rationale for randomly sampling the forward propagation path.

The block prior mixer dynamically assembles network blocks from the teacher and student models. Randomly sampling the forward propagation path ensures a diverse set of combinations, encouraging the student model to generalize better by learning from various teacher-student configurations. Such a stochastic approach mitigates overfitting and reduces the capacity disparity between teacher and student models. By selectively using teacher or student blocks during propagation, the mixer provides flexibility and enables the student to progressively inherit and adapt the teacher’s capabilities.

Question 2: Specifications of the encoder and decoder networks.

In the Feature Prior Mixer module, separate encoders are designed for the teacher and student models to ensure effective mapping to the latent space, accommodating the distinct feature distributions of each. The decoders transform the fused feature maps back into enhanced representations for subsequent distillation. The encoder and decoder architectures are deliberately lightweight, focusing on dimensional consistency and efficient projection into the latent space rather than performing deep feature extraction. Typically, three convolution layers are deep enough. Table 8 provides comparisons between MLP and CNN-based encoder/decoder networks.

To address your question on the size of encoder and decoder, we include an ablation study that evaluates the impact of increasing the number of convolutional layers in them.

Question 1: Concern on the training stability and hyper-parameter optimization

In our experiments, we followed the standard hyper-parameter setting (learning rate, lr scheduler, and optimizer), as summarized in section 4.1. For the loss weights $\lambda\_{rec}, \lambda\_{kd}, \lambda\_{feat}, \lambda\_{block}$, we have evaluated their effect in Tab. 11. We further add the ablation about the distillation position of MiPKD on RCAN as shown in the table below. Applying MiPKD uniformly at early, intermediate and deep layers achieves the best performance. For simplicity, we apply the setting of all positions for other models, which also achieves the high distillation performance. The training process of MiPKD is stable to distill SR models. For example, the training curve of X4 RCAN network is added in Fig.4 of the supplementary material to demonstrate the training stability.

Weakness 4: Experiments on more recent SOTA networks

As suggested, we apply the proposed MiPKD to distill the DRCT[2] model for x4 SR, as shown in the below table. The proposed MiPKD outperforms the existing KD methods, including Logits-KD and FitNet. For example, MiPKD outperforms the logits KD by 0.13 dB for x4 SR on Urban100.

[2] Hsu, C., Lee, C., & Chou, Y. (2024). DRCT: Saving Image Super-Resolution away from Information Bottleneck. 2024 CVPR

Weakness 3, Question 2: Extension to other vision tasks

Both Feature Prior Mixer and Block Prior Mixer strategies are flexible for different network architectures, which has strong potential for extension to other vision tasks. As such, we add the experiment by applying MiPKD to the ImageNet-1k classification task as an auxiliary and standard task, as shown in Table below. By distilling the ResNet18 model using  the ResNet34, MiPKD still outperforms feature distillation method SRRL[1] by 0.15 Top-1 accuracy. The multi-granularity prior mixing helps the student model capture crucial high-level semantic information, which potentially verify the broader applicability of the proposed MiPKD.

[1] Yang, J., Martínez, B., Bulat, A., & Tzimiropoulos, G. (2021). Knowledge distillation via softmax regression representation learning. ICLR

Weakness 2: More discussions on the mask generation strategies.

MiPKD employs a random 3D mask as the default masking strategy in the Feature Prior Mixer. This approach selectively combines feature maps from the teacher and student models in the latent space, introducing the controlled stochasticity and diversity into the mixing process. It encourages the student model to learn essential teacher features across varied spatial configurations.

In Tab. 10, we compared the random 3d mask generation strategy with two alternatives: grid masking and similarity-based masking (by Cosine and CKA similarities). Grid masking applies the predefined grid pattern to select the masking position, while similarity-based masking guides the mixing based on feature similarity. The ablation study indicates that the random 3D mask achieves the best performance, compared to other strategies.

We sincerely thank you for your positive evaluation. Your comments are deeply encouraging with valuable insights. If you have any further suggestions or would like to discuss our work in more detail, we would be delighted to engage in a conversation. Your feedback is invaluable to us. Detailed responses are provided in the following rebuttal.

Weakness 1: Why random masks are effective in distilling SR networks?

The masking strategy in MiPKD plays a crucial role by using a stochastic mixing of feature maps from the teacher and student models within a unified latent space, and a stochastic routing between teacher and student blocks to assemble a dynamic combination of blocks through the SR output alignment. In this way, MiPKD effectively reduces the capacity disparity between the teacher and student.

Specifically, a random 3D masks ($I$) is used in feature prior mixer (FPM). The encoder-decoder structure in the FPM reconstructs the masked teacher features, facilitating the student's approximation of the teacher's feature distribution. It encourages the student model to align its intermediate representations with the teacher’s by adding the part feature information as the prior, which ensures better feature learning for student.

Block Prior Mixer with Mask $R\_k$​: Mask $R\_k$​ introduces stochastic routing between teacher and student blocks, allowing the student to learn from the teacher’s pathways without replicating its structure. This block-level fusion helps the student model adapt to complex spatial and texture variations, gradually learning the teacher’s feature transformation style to enhance the feature presentation ability of the student.

To further support this explanation, feature map analyses are added to the supplementary material as Fig.3 shown, demonstrating how the masking strategy enhances the student model’s representation learning and overall performance. The random masking approach in MiPKD encourages the student model to focus selectively on critical and diverse features from the teacher's output rather than attempting a full replication. This combination of random masking allows MiPKD to effectively balance global context with detailed textures, making it particularly well-suited for high-detail tasks, such as super-resolution.

Dear Reviewer pGYs,

Thank you for your valuable feedback and constructive comments. We greatly appreciate the time and attention you have dedicated to this process.

Sincerely Yours,

The Authors